diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index aa9cc6304f..28bd28b84e 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -327,7 +327,7 @@ wait_for_server_ready() {
 }
 
 # Run benchmark serving with standardized parameters
-# All parameters are required except --endpoint, --use-chat-template, --dsv4, and --trust-remote-code
+# All parameters are required except the optional parameters listed below.
 # Parameters:
 #   --model: Model name
 #   --port: Server port
@@ -346,9 +346,14 @@ wait_for_server_ready() {
 #           template. Implies --use-chat-template.
 #   --trust-remote-code: Optional flag to trust remote code from HuggingFace
 #   --server-pid: Optional server process ID to monitor during benchmark
+#   --tokenizer: Optional tokenizer name/path override
+#   --tokenizer-mode: Optional tokenizer implementation mode
+#   --host: Optional server host (default: 0.0.0.0)
+#   --request-rate: Optional request rate (default: inf)
+#   --num-warmups: Optional warmup request count (default: 2 * concurrency)
 run_benchmark_serving() {
     # In eval-only mode, skip the throughput benchmark entirely.
-    if [ "${EVAL_ONLY}" = "true" ]; then
+    if [ "${EVAL_ONLY:-false}" = "true" ]; then
         echo "EVAL_ONLY mode: skipping throughput benchmark"
         return 0
     fi
@@ -372,6 +377,9 @@ run_benchmark_serving() {
     local server_pid=""
     local tokenizer=""
     local tokenizer_mode=""
+    local host="0.0.0.0"
+    local request_rate="inf"
+    local num_warmups=""
 
     while [[ $# -gt 0 ]]; do
         case $1 in
@@ -448,6 +456,18 @@ run_benchmark_serving() {
                 tokenizer_mode="$2"
                 shift 2
                 ;;
+            --host)
+                host="$2"
+                shift 2
+                ;;
+            --request-rate)
+                request_rate="$2"
+                shift 2
+                ;;
+            --num-warmups)
+                num_warmups="$2"
+                shift 2
+                ;;
             *)
                 echo "Unknown parameter: $1"
                 return 1
@@ -500,6 +520,9 @@ run_benchmark_serving() {
     if [[ -z "$workspace_dir" ]]; then
         workspace_dir=$(pwd)
     fi
+    if [[ -z "$num_warmups" ]]; then
+        num_warmups="$((2 * max_concurrency))"
+    fi
 
     # Profiling support: when PROFILE=1, ensure profiler dir exists, add --profile flag,
     # and cap num_prompts to keep traces small.
@@ -518,18 +541,18 @@ run_benchmark_serving() {
         python3 "$workspace_dir/utils/bench_serving/benchmark_serving.py"
         --model "$model"
         --backend "$backend"
-        --base-url "http://0.0.0.0:$port"
+        --base-url "http://${host}:$port"
         --dataset-name random
         --random-input-len "$input_len"
         --random-output-len "$output_len"
         --random-range-ratio "$random_range_ratio"
         --num-prompts "$num_prompts"
         --max-concurrency "$max_concurrency"
-        --request-rate inf
+        --request-rate "$request_rate"
         --ignore-eos
         "${profile_flag[@]}"
         --save-result
-        --num-warmups "$((2 * max_concurrency))" \
+        --num-warmups "$num_warmups" \
         --percentile-metrics 'ttft,tpot,itl,e2el'
         --result-dir "$result_dir"
         --result-filename "$result_filename.json"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx10dep4_gen1dep16_batch64_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx10dep4_gen1dep16_batch64_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..c8dd651d11
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx10dep4_gen1dep16_batch64_eplb384_mtp3.yaml
@@ -0,0 +1,154 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx10dep4_gen1dep16_batch64_eplb384_mtp3"
+
+# ctx: 10 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true, max_batch=64
+# concurrency: 1229
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 10
+  prefill_workers: 10
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 64
+      max_num_tokens: 256
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep16_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 16
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep8_batch512_eplb384_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep8_batch512_eplb384_mtp1.yaml
new file mode 100644
index 0000000000..afde0cab83
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep8_batch512_eplb384_mtp1.yaml
@@ -0,0 +1,210 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx12dep4_gen1dep8_batch512_eplb384_mtp1"
+
+# ctx: 12 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP8/EP8, enable_attention_dp=true, max_batch=512
+# concurrency: 4301
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 12
+  prefill_workers: 12
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 2
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 1
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+          - 136
+          - 144
+          - 152
+          - 160
+          - 168
+          - 176
+          - 184
+          - 192
+          - 200
+          - 208
+          - 216
+          - 224
+          - 232
+          - 240
+          - 248
+          - 256
+          - 264
+          - 272
+          - 280
+          - 288
+          - 296
+          - 304
+          - 312
+          - 320
+          - 328
+          - 336
+          - 344
+          - 352
+          - 360
+          - 368
+          - 376
+          - 384
+          - 392
+          - 400
+          - 408
+          - 416
+          - 424
+          - 432
+          - 440
+          - 448
+          - 456
+          - 464
+          - 472
+          - 480
+          - 488
+          - 496
+          - 504
+          - 512
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        tokens_per_block: 128
+      max_batch_size: 512
+      max_num_tokens: 1024
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep8_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 1
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "4301"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml
new file mode 100644
index 0000000000..1da0c8b563
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml
@@ -0,0 +1,139 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch1_eplb0_mtp3"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, max_batch=1
+# concurrency: 8
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 1
+      max_num_tokens: 4
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "8"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
new file mode 100644
index 0000000000..15083f7578
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
@@ -0,0 +1,139 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch1_eplb0_mtp3"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=1
+# concurrency: 10
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 1
+      max_num_tokens: 4
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "10"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml
new file mode 100644
index 0000000000..04ae7b8563
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml
@@ -0,0 +1,137 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch2_eplb0_mtp3"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=2
+# concurrency: 15
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "15"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml
new file mode 100644
index 0000000000..e63fdfd01d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml
@@ -0,0 +1,139 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch4_eplb0_mtp3"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=4
+# concurrency: 30
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 4
+      max_num_tokens: 16
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "30"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch2_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch2_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..562b0bad61
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch2_eplb384_mtp3.yaml
@@ -0,0 +1,146 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx2dep4_gen1dep32_batch2_eplb384_mtp3"
+
+# ctx: 2 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=2
+# concurrency: 84
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "84"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx3dep4_gen1dep32_batch4_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx3dep4_gen1dep32_batch4_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..2705d4b167
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx3dep4_gen1dep32_batch4_eplb384_mtp3.yaml
@@ -0,0 +1,146 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx3dep4_gen1dep32_batch4_eplb384_mtp3"
+
+# ctx: 3 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=4
+# concurrency: 180
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 3
+  prefill_workers: 3
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 4
+      max_num_tokens: 16
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "180"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..6daa450c67
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb384_mtp3.yaml
@@ -0,0 +1,147 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx4dep4_gen1dep32_batch8_eplb384_mtp3"
+
+# ctx: 4 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=8
+# concurrency: 333
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 4
+  prefill_workers: 4
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 8
+      max_num_tokens: 32
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "333"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..b3806e169f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb384_mtp3.yaml
@@ -0,0 +1,150 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx6dep4_gen1dep16_batch32_eplb384_mtp3"
+
+# ctx: 6 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true, max_batch=32
+# concurrency: 666
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 6
+  prefill_workers: 6
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 32
+      max_num_tokens: 128
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep16_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 16
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "666"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep8_batch128_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep8_batch128_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..f6b144e299
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep8_batch128_eplb384_mtp3.yaml
@@ -0,0 +1,162 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx7dep4_gen1dep8_batch128_eplb384_mtp3"
+
+# ctx: 7 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP8/EP8, enable_attention_dp=true, max_batch=128
+# concurrency: 1229
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 7
+  prefill_workers: 7
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 2
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        tokens_per_block: 128
+      max_batch_size: 128
+      max_num_tokens: 512
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep8_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep32_batch16_eplb384_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep32_batch16_eplb384_mtp3.yaml
new file mode 100644
index 0000000000..6cab42a71e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep32_batch16_eplb384_mtp3.yaml
@@ -0,0 +1,148 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx8dep4_gen1dep32_batch16_eplb384_mtp3"
+
+# ctx: 8 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=16
+# concurrency: 615
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 8
+  prefill_workers: 8
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 16
+      max_num_tokens: 64
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "615"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx9dep4_gen1dep8_batch256_eplb384_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx9dep4_gen1dep8_batch256_eplb384_mtp1.yaml
new file mode 100644
index 0000000000..03ad62777e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/MTP/ctx9dep4_gen1dep8_batch256_eplb384_mtp1.yaml
@@ -0,0 +1,178 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx9dep4_gen1dep8_batch256_eplb384_mtp1"
+
+# ctx: 9 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP8/EP8, enable_attention_dp=true, max_batch=256
+# concurrency: 2253
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 9
+  prefill_workers: 9
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 2
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 1
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+          - 136
+          - 144
+          - 152
+          - 160
+          - 168
+          - 176
+          - 184
+          - 192
+          - 200
+          - 208
+          - 216
+          - 224
+          - 232
+          - 240
+          - 248
+          - 256
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        tokens_per_block: 128
+      max_batch_size: 256
+      max_num_tokens: 512
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep8_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 1
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2253"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx10dep4_gen1dep8_batch512_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx10dep4_gen1dep8_batch512_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..a53eacde9a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx10dep4_gen1dep8_batch512_eplb384_mtp0.yaml
@@ -0,0 +1,204 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx10dep4_gen1dep8_batch512_eplb384_mtp0"
+
+# ctx: 10 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP8/EP8, enable_attention_dp=true, max_batch=512
+# concurrency: 4301
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 10
+  prefill_workers: 10
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 2
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+          - 136
+          - 144
+          - 152
+          - 160
+          - 168
+          - 176
+          - 184
+          - 192
+          - 200
+          - 208
+          - 216
+          - 224
+          - 232
+          - 240
+          - 248
+          - 256
+          - 264
+          - 272
+          - 280
+          - 288
+          - 296
+          - 304
+          - 312
+          - 320
+          - 328
+          - 336
+          - 344
+          - 352
+          - 360
+          - 368
+          - 376
+          - 384
+          - 392
+          - 400
+          - 408
+          - 416
+          - 424
+          - 432
+          - 440
+          - 448
+          - 456
+          - 464
+          - 472
+          - 480
+          - 488
+          - 496
+          - 504
+          - 512
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        tokens_per_block: 128
+      max_batch_size: 512
+      max_num_tokens: 512
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep8_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "4301"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen1dep32_batch4_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen1dep32_batch4_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..8e1eb368a8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen1dep32_batch4_eplb384_mtp0.yaml
@@ -0,0 +1,140 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen1dep32_batch4_eplb384_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=4
+# concurrency: 154
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "154"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..73f09b764b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
@@ -0,0 +1,133 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch1_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, max_batch=1
+# concurrency: 4
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 1
+      max_num_tokens: 1
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "4"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..943241e70f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
@@ -0,0 +1,133 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch1_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=1
+# concurrency: 5
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 1
+      max_num_tokens: 1
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "5"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..642b6fd272
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml
@@ -0,0 +1,133 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch2_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=2
+# concurrency: 15
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 2
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "15"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..1f94a546ff
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
@@ -0,0 +1,133 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch4_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=4
+# concurrency: 25
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "25"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..5f4eaffc25
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
@@ -0,0 +1,134 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch8_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, max_batch=8
+# concurrency: 55
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+        enable_padding: true
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+        tokens_per_block: 128
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 4
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "55"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..7fa94e6d23
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb384_mtp0.yaml
@@ -0,0 +1,141 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx2dep4_gen1dep32_batch8_eplb384_mtp0"
+
+# ctx: 2 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=8
+# concurrency: 308
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "308"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..6b36508383
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb384_mtp0.yaml
@@ -0,0 +1,142 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx4dep4_gen1dep32_batch16_eplb384_mtp0"
+
+# ctx: 4 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=16
+# concurrency: 615
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 4
+  prefill_workers: 4
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 16
+      max_num_tokens: 16
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "615"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..0961046024
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb384_mtp0.yaml
@@ -0,0 +1,148 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx5dep4_gen1dep16_batch64_eplb384_mtp0"
+
+# ctx: 5 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true, max_batch=64
+# concurrency: 1229
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 5
+  prefill_workers: 5
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 64
+      max_num_tokens: 64
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep16_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 16
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..6493ba43aa
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb384_mtp0.yaml
@@ -0,0 +1,144 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx6dep4_gen1dep32_batch32_eplb384_mtp0"
+
+# ctx: 6 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true, max_batch=32
+# concurrency: 1127
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 6
+  prefill_workers: 6
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 32
+      max_num_tokens: 32
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep32_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 32
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 32
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1127"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep8_batch256_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep8_batch256_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..507a169c88
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep8_batch256_eplb384_mtp0.yaml
@@ -0,0 +1,172 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx6dep4_gen1dep8_batch256_eplb384_mtp0"
+
+# ctx: 6 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP8/EP8, enable_attention_dp=true, max_batch=256
+# concurrency: 2253
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 6
+  prefill_workers: 6
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 2
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+          - 136
+          - 144
+          - 152
+          - 160
+          - 168
+          - 176
+          - 184
+          - 192
+          - 200
+          - 208
+          - 216
+          - 224
+          - 232
+          - 240
+          - 248
+          - 256
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        tokens_per_block: 128
+      max_batch_size: 256
+      max_num_tokens: 256
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep8_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 8
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2253"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb384_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb384_mtp0.yaml
new file mode 100644
index 0000000000..069a7f11da
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/DeepSeek-V4-Pro/disagg/trtllm_dynamo/gb300_mxfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb384_mtp0.yaml
@@ -0,0 +1,156 @@
+name: "dsv4pro_mxfp4_ISL8K_OSL1K_ctx9dep4_gen1dep16_batch128_eplb384_mtp0"
+
+# ctx: 9 prefill workers, TP4/EP4, EPLB: num_slots=384
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true, max_batch=128
+# concurrency: 2253
+
+model:
+  path: "deepseek-ai/DeepSeek-V4-Pro"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1"
+  precision: "fp4"
+
+extra_mount:
+  - "tmpfs:/dev/shm:size=100%"
+  - "configs/dsv4-moe-load-balancer-configs:/dsv4-eplb-configs"
+resources:
+  gpu_type: "gb300"
+
+  prefill_nodes: 9
+  prefill_workers: 9
+  gpus_per_prefill: 4
+
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: "1"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "-1"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  decode_environment:
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+    PYTHONUNBUFFERED: "1"
+    ENROOT_ALLOW_DEV: "yes"
+    MIMALLOC_PURGE_DELAY: "0"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    PYTHONWARNINGS: "ignore::DeprecationWarning:cutlass.cute.core"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+
+  trtllm_config:
+    prefill:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_attention_dp: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.5
+        tokens_per_block: 128
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8232
+      moe_config:
+        backend: TRTLLM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_ctx_ep4_384.yaml"
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      tensor_parallel_size: 4
+
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 8192
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+        enable_padding: true
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.7
+        tokens_per_block: 128
+      max_batch_size: 128
+      max_num_tokens: 128
+      max_seq_len: 9256
+      moe_config:
+        backend: MEGAMOE_DEEPGEMM
+        load_balancer: "/dsv4-eplb-configs/moe_load_balancer_gen_ep16_slots384.yaml"
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      trust_remote_code: true
+      stream_interval: 100
+      tensor_parallel_size: 16
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2253"
+  req_rate: "inf"
+  num_prompts_mult: 10
+  num_warmup_mult: 20
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+  use_chat_template: true
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+dynamo:
+  request_plane: "tcp"
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..83b6633006
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
@@ -0,0 +1,129 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch1_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 8
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 4 workers x TP8 = 32 GPUs = 8 nodes
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      pipeline_parallel_size: 1
+      allreduce_strategy: MNNVL
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 1
+      max_num_tokens: 1
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "8"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..005388ef8b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch4_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 24
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 4 workers x TP8 = 32 GPUs = 8 nodes
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      pipeline_parallel_size: 1
+      allreduce_strategy: MNNVL
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "24"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..953fb48dff
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
@@ -0,0 +1,132 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch16_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 105
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 16
+      max_num_tokens: 16
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "105"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..90e8e3666c
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
@@ -0,0 +1,130 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch1_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 5
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 1
+      max_num_tokens: 1
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "5"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..81f4019f3d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
@@ -0,0 +1,130 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch4_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 30
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "30"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..71134f59c8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch8_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 60
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "60"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..9c02b8195c
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx2dep4_gen1dep32_batch8_eplb0_mtp0"
+
+# ctx: 2 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 333
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 2 workers x TP4 = 8 GPUs = 2 nodes
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "333"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep16_batch64_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep16_batch64_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..b59412be8b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep16_batch64_eplb0_mtp0.yaml
@@ -0,0 +1,138 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx4dep4_gen1dep16_batch64_eplb0_mtp0"
+
+# ctx: 4 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 1229
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 4 workers x TP4 = 16 GPUs = 4 nodes
+  prefill_nodes: 4
+  prefill_workers: 4
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP16 = 16 GPUs = 4 nodes
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 64
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..0cb843d3a7
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml
@@ -0,0 +1,132 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx4dep4_gen1dep32_batch16_eplb0_mtp0"
+
+# ctx: 4 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 666
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 4 workers x TP4 = 16 GPUs = 4 nodes
+  prefill_nodes: 4
+  prefill_workers: 4
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 16
+      max_num_tokens: 16
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "666"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..4da77069ea
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml
@@ -0,0 +1,134 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx6dep4_gen1dep32_batch32_eplb0_mtp0"
+
+# ctx: 6 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 1229
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 6 workers x TP4 = 24 GPUs = 6 nodes
+  prefill_nodes: 6
+  prefill_workers: 6
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 32
+      max_num_tokens: 32
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx7dep4_gen1dep16_batch128_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx7dep4_gen1dep16_batch128_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..a1ca05cdf5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx7dep4_gen1dep16_batch128_eplb0_mtp0.yaml
@@ -0,0 +1,145 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx7dep4_gen1dep16_batch128_eplb0_mtp0"
+
+# ctx: 7 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true, max_batch=128
+# concurrency: 2253
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 7 workers x TP4 = 28 GPUs = 7 nodes
+  prefill_nodes: 7
+  prefill_workers: 7
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP16 = 16 GPUs = 4 nodes
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 128
+      max_num_tokens: 128
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2253"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch256_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch256_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..672583855b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb200Nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch256_eplb0_mtp0.yaml
@@ -0,0 +1,162 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx9dep4_gen1dep16_batch256_eplb0_mtp0"
+
+# ctx: 9 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 4301
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb200"
+
+  # Prefill: 9 workers x TP4 = 36 GPUs = 9 nodes
+  prefill_nodes: 9
+  prefill_workers: 9
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP16 = 16 GPUs = 4 nodes
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 256
+      max_num_tokens: 256
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+          - 136
+          - 144
+          - 152
+          - 160
+          - 168
+          - 176
+          - 184
+          - 192
+          - 200
+          - 208
+          - 216
+          - 224
+          - 232
+          - 240
+          - 248
+          - 256
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "4301"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..c11f3a36c3
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch1_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 4
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 4 workers x TP8 = 32 GPUs = 8 nodes
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      pipeline_parallel_size: 1
+      allreduce_strategy: MNNVL
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 1
+      max_num_tokens: 1
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "4"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch2_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch2_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..3e55488a52
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch2_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch2_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 12
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 4 workers x TP8 = 32 GPUs = 8 nodes
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      pipeline_parallel_size: 1
+      allreduce_strategy: MNNVL
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 2
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "12"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..29b843dedc
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen4tep8_batch4_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 4 decode workers, TP8/EP8, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 24
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 4 workers x TP8 = 32 GPUs = 8 nodes
+  decode_workers: 4
+  decode_nodes: 8
+  gpus_per_decode: 8
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      pipeline_parallel_size: 1
+      allreduce_strategy: MNNVL
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "24"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..705f29a902
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
@@ -0,0 +1,132 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch16_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 115
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 16
+      max_num_tokens: 16
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "115"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..a6957ba9ff
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
@@ -0,0 +1,130 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch4_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 30
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 4
+      max_num_tokens: 4
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "30"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..eff0524b1b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx1dep4_gen5tep4_batch8_eplb0_mtp0"
+
+# ctx: 1 prefill worker, TP4/EP4
+# gen: 5 decode workers, TP4/EP4, allreduce_strategy=MNNVL, enable_attention_dp=false
+# STP (no speculative decoding)
+# concurrency: 60
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 1 worker x TP4 = 4 GPUs = 1 node
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 4
+
+  # Decode: 5 workers x TP4 = 20 GPUs = 5 nodes
+  decode_workers: 5
+  decode_nodes: 5
+  gpus_per_decode: 4
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.9
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "60"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: false
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..f4a4d26f27
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
@@ -0,0 +1,131 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx2dep4_gen1dep32_batch8_eplb0_mtp0"
+
+# ctx: 2 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 333
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 2 workers x TP4 = 8 GPUs = 2 nodes
+  prefill_nodes: 2
+  prefill_workers: 2
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 8
+      max_num_tokens: 8
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "333"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx3dep4_gen1dep32_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx3dep4_gen1dep32_batch16_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..b74a10e6b2
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx3dep4_gen1dep32_batch16_eplb0_mtp0.yaml
@@ -0,0 +1,132 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx3dep4_gen1dep32_batch16_eplb0_mtp0"
+
+# ctx: 3 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 615
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 3 workers x TP4 = 12 GPUs = 3 nodes
+  prefill_nodes: 3
+  prefill_workers: 3
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 16
+      max_num_tokens: 16
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "615"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep32_batch32_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep32_batch32_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..47a3fabfbf
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep32_batch32_eplb0_mtp0.yaml
@@ -0,0 +1,134 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx5dep4_gen1dep32_batch32_eplb0_mtp0"
+
+# ctx: 5 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 1229
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 5 workers x TP4 = 20 GPUs = 5 nodes
+  prefill_nodes: 5
+  prefill_workers: 5
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 32
+      max_num_tokens: 32
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "1229"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep16_batch128_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep16_batch128_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..bd66cfe13e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep16_batch128_eplb0_mtp0.yaml
@@ -0,0 +1,146 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx6dep4_gen1dep16_batch128_eplb0_mtp0"
+
+# ctx: 6 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP16/EP16, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 2253
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 6 workers x TP4 = 24 GPUs = 6 nodes
+  prefill_nodes: 6
+  prefill_workers: 6
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP16 = 16 GPUs = 4 nodes
+  decode_workers: 1
+  decode_nodes: 4
+  gpus_per_decode: 16
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 128
+      max_num_tokens: 128
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+          - 72
+          - 80
+          - 88
+          - 96
+          - 104
+          - 112
+          - 120
+          - 128
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2253"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx8dep4_gen1dep32_batch64_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx8dep4_gen1dep32_batch64_eplb0_mtp0.yaml
new file mode 100644
index 0000000000..66cdd85f3d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/kimi2.5/trtllm_dynamo/disagg/gb300Nvfp4/ISL8K_OSL1K/STP/ctx8dep4_gen1dep32_batch64_eplb0_mtp0.yaml
@@ -0,0 +1,138 @@
+name: "kimi_k25_nvfp4_ISL8K_OSL1K_ctx8dep4_gen1dep32_batch64_eplb0_mtp0"
+
+# ctx: 8 prefill workers, TP4/EP4
+# gen: 1 decode worker, TP32/EP32, enable_attention_dp=true
+# STP (no speculative decoding)
+# concurrency: 2151
+
+model:
+  path: "nvidia/Kimi-K2.5-NVFP4"
+  container: "nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-dev.1-cuda13"
+  precision: "fp4"
+
+resources:
+  gpu_type: "gb300"
+
+  # Prefill: 8 workers x TP4 = 32 GPUs = 8 nodes
+  prefill_nodes: 8
+  prefill_workers: 8
+  gpus_per_prefill: 4
+
+  # Decode: 1 worker x TP32 = 32 GPUs = 8 nodes
+  decode_workers: 1
+  decode_nodes: 8
+  gpus_per_decode: 32
+
+  gpus_per_node: 4
+
+backend:
+  type: trtllm
+
+  prefill_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  decode_environment:
+    ENROOT_ALLOW_DEV: "yes"
+    NCCL_GRAPH_MIXING_SUPPORT: "0"
+    TLLM_LOG_LEVEL: "INFO"
+    TRTLLM_ENABLE_PDL: "1"
+    TRTLLM_SERVER_DISABLE_GC: "1"
+    TRTLLM_WORKER_DISABLE_GC: "1"
+    MIMALLOC_PURGE_DELAY: "0"
+    UCX_TLS: "cuda_ipc,cuda_copy,sm,self,tcp"
+
+  trtllm_config:
+    prefill:
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 4
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      disable_overlap_scheduler: true
+      trust_remote_code: true
+      max_batch_size: 2
+      max_num_tokens: 16384
+      max_seq_len: 8232
+      print_iter_log: true
+      cuda_graph_config: null
+      moe_config:
+        backend: CUTEDSL
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.4
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+
+    decode:
+      tensor_parallel_size: 32
+      moe_expert_parallel_size: 32
+      pipeline_parallel_size: 1
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: false
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 64
+      max_seq_len: 9256
+      print_iter_log: true
+      stream_interval: 100
+      num_postprocess_workers: 4
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+          - 1
+          - 2
+          - 4
+          - 8
+          - 16
+          - 24
+          - 32
+          - 40
+          - 48
+          - 56
+          - 64
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      kv_cache_config:
+        dtype: fp8
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.75
+      cache_transceiver_config:
+        backend: DEFAULT
+        kv_transfer_timeout_ms: 600000
+        max_tokens_in_buffer: 16384
+      nvfp4_gemm_config:
+        allowed_backends:
+          - cutlass
+          - cublaslt
+          - cutedsl
+          - cuda_core
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
+  isl: 8192
+  osl: 1024
+  concurrencies: "2151"
+  req_rate: "inf"
+
+frontend:
+  type: "dynamo"
+  enable_multiple_frontends: true
+
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+
+dynamo:
+  request_plane: tcp
+  install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml
index f46e40782e..28ec1815b2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml
index 6305f653af..77a8fd7f41 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml
@@ -109,7 +109,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml
index ff0bb705a1..e93800e443 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p4d-dep8-tp8-10-c64.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p4d-dep8-tp8-10-c64.yaml
index 84349c2775..89afabbdff 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p4d-dep8-tp8-10-c64.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p4d-dep8-tp8-10-c64.yaml
@@ -134,7 +134,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "64"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-2p1d-dep8-dep16-8-c1536.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-2p1d-dep8-dep16-8-c1536.yaml
index 478c91b046..6bf374e294 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-2p1d-dep8-dep16-8-c1536.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-2p1d-dep8-dep16-8-c1536.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1536"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-4p1d-dep8-dep16-12-c4096.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-4p1d-dep8-dep16-12-c4096.yaml
index 11434ca534..0ecb4a6d36 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-4p1d-dep8-dep16-12-c4096.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-4p1d-dep8-dep16-12-c4096.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-5p1d-dep8-dep16-14-c8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-5p1d-dep8-dep16-14-c8192.yaml
index 9962318e8a..3a464508af 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-5p1d-dep8-dep16-14-c8192.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-5p1d-dep8-dep16-14-c8192.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-6p1d-dep8-dep12-15-c8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-6p1d-dep8-dep12-15-c8192.yaml
index e88d4b7d53..e3ac60b586 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-6p1d-dep8-dep12-15-c8192.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-6p1d-dep8-dep12-15-c8192.yaml
@@ -148,7 +148,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml
index 0b2423a8ec..e5c875cd4c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p1d-tp8-tp8-mtp.yaml
@@ -113,7 +113,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml
index 79c9a46bda..699f1cba7d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-low-latency-1p6d-dep8-tp8-mtp.yaml
@@ -120,7 +120,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml
index 1bf4f0e859..6edc00c0d1 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-1p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-2p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-2p1d-dep8-dep16-mtp.yaml
index 82519e378c..87584f06d3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-2p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-2p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-3p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-3p1d-dep8-dep16-mtp.yaml
index e69c5e6043..6a7a0a9152 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-3p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-3p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep16-mtp.yaml
index 73bcecaec5..bb8c0ccff6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-4p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep16-mtp.yaml
index 66829c4044..5a05012a25 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-5p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep16-mtp.yaml
index 34b71a9180..270d3f097f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-mid-curve-6p1d-dep8-dep16-mtp.yaml
@@ -132,7 +132,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep32-18-c2500.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep32-18-c2500.yaml
index 528aa5721a..e91fd22e16 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep32-18-c2500.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep32-18-c2500.yaml
@@ -149,7 +149,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2500"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep24-18-c3000.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep24-18-c3000.yaml
index 32a5124c2d..3635f7bacc 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep24-18-c3000.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep24-18-c3000.yaml
@@ -149,7 +149,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "3000"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-14p1d-dep4-dep16-18-c8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-14p1d-dep4-dep16-18-c8192.yaml
index adc6b15507..fa8de70ae4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-14p1d-dep4-dep16-18-c8192.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-14p1d-dep4-dep16-18-c8192.yaml
@@ -149,7 +149,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-15p1d-dep4-dep12-18-c12000.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-15p1d-dep4-dep12-18-c12000.yaml
index 73ad15750d..dfa1791e6c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-15p1d-dep4-dep12-18-c12000.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-15p1d-dep4-dep12-18-c12000.yaml
@@ -149,7 +149,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "12000"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml
index 7cdb779c71..7ea8bb37f7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml
@@ -173,7 +173,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml
index 9382dd6dda..fdb26123c5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml
@@ -156,7 +156,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep40-18-c2048.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep40-18-c2048.yaml
index 269b92e124..6e532ee45f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep40-18-c2048.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep40-18-c2048.yaml
@@ -150,7 +150,8 @@ backend:
 
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-6p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-6p1d-dep4-dep8-mtp.yaml
index 39e11b7199..a91dd3f79e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-6p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-6p1d-dep4-dep8-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-8p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-8p1d-dep4-dep8-mtp.yaml
index a4cf477ed3..b7a2b0e545 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-8p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-high-conc-8p1d-dep4-dep8-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p1d-tp4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p1d-tp4-tp4-mtp.yaml
index be5c4cf389..ca3f698eaa 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p1d-tp4-tp4-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p1d-tp4-tp4-mtp.yaml
@@ -111,7 +111,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p6d-dep4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p6d-dep4-tp4-mtp.yaml
index 5657ad84d6..c0966bf1f5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p6d-dep4-tp4-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-low-latency-1p6d-dep4-tp4-mtp.yaml
@@ -120,7 +120,8 @@ backend:
       context-length: 16384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep16-mtp.yaml
index f4ae3076ce..9972200ddb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep16-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep16-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep8-mtp.yaml
index 4f9f902769..2cba35bb88 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-1p1d-dep4-dep8-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-2p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-2p1d-dep4-dep8-mtp.yaml
index 17018a57ea..d2ca721142 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-2p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-2p1d-dep4-dep8-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-4p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-4p1d-dep4-dep8-mtp.yaml
index 15578537a4..6a05aca028 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-4p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-mid-curve-4p1d-dep4-dep8-mtp.yaml
@@ -131,7 +131,8 @@ backend:
       stream-interval: 60
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   random_range_ratio: 0.8
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
index 11f56972ee..72c77ad02a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
@@ -115,7 +115,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
index acbec6f11a..ce3b1ff38c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
@@ -115,7 +115,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_0.yaml
index 5f67dfb50f..841f27693c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_0.yaml
@@ -120,7 +120,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_1.yaml
index 36ceac6122..129d92aef5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/1k1k/disagg/mtp/1k1k_mtp_maxtpt_1.yaml
@@ -120,7 +120,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_1p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_1p1d.yaml
index 13a1a7d9d0..7272666dde 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_1p1d.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_1p1d.yaml
@@ -135,7 +135,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_2p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_2p1d.yaml
index 37225d44e8..863aa17187 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_2p1d.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_2p1d.yaml
@@ -135,7 +135,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_3p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_3p1d.yaml
index d8e401620a..b193320ddb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_3p1d.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_3p1d.yaml
@@ -135,7 +135,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_4p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_4p1d.yaml
index a3923b02db..c762458310 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_4p1d.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_4p1d.yaml
@@ -135,7 +135,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_5p1d.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_5p1d.yaml
index bcbedcb68b..1a6282da07 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_5p1d.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp2_throughput_5p1d.yaml
@@ -135,7 +135,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_4p1d_c2048.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_4p1d_c2048.yaml
index fc2b0f3366..9532a234a6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_4p1d_c2048.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_4p1d_c2048.yaml
@@ -111,7 +111,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
index 62bbddf84d..eb59efeaad 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
@@ -132,7 +132,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
index 521eddf931..0a6414feb4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
@@ -132,7 +132,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
index 9dfc6331a8..bb6f1c53ed 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
@@ -132,7 +132,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: 300
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_3.yaml
index ba9907840f..62db12c17c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_3.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_4.yaml
index ee2b0a2026..57a6fb5e0e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_hightpt_4.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
index cd01ba6b73..6357ec6378 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_0.yaml
@@ -137,7 +137,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
index 1ec2d053a3..8438a92675 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/1k1k/disagg/mtp/1k1k_mtp_lowlat_1.yaml
@@ -137,7 +137,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_0.yaml
index 6fbd1df10b..0e2b3700ff 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_0.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_1.yaml
index 3c1a641a82..52cdd2943c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_1.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_2.yaml
index 1ee087dfa2..a0cb821f92 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_2.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_3.yaml
index a03cdef22d..fe36683375 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_hightpt_3.yaml
@@ -145,7 +145,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
index baf3aa3609..6678aefaa1 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
@@ -137,7 +137,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
index c386534d01..d0a16ac97f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
@@ -137,7 +137,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
index 561ac280cc..c9061070ff 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1/gb300-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_2.yaml
@@ -137,7 +137,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
index 76edd78e30..eeafc68939 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "2115"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
index 7ff2fb9e63..6f9f339d12 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1156"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
index dde20d03d5..1ac0e78f71 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "556"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
index 1a084f3876..001d4a1d4f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "23"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
index b8cc503e54..6cb8f91f58 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "290"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
index ddc60e9504..a0506f384b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "73"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_maxtpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_maxtpt_0.yaml
index 3f7a83b6c3..de6fece6e0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_maxtpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/1k1k/disagg/stp/1k1k_stp_maxtpt_0.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "3160"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
index eb5b2467e8..00bd6d8a60 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "133"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
index 07b90a7261..e1a3395fc7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "146"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
index f7faa73fe8..5dde716dc3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "103"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
index 242a637dfa..d6247f91b2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "130"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
index 59a3d9fc28..48ceba5c8c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "113"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_5.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_5.yaml
index 595c310d30..4d2bb3393c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_5.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_5.yaml
@@ -136,7 +136,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "23"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
index 8e383ee108..e0a107c957 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "989"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
index da9815cb24..27b7bfc31f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "686"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
index f5b158baee..9cc26f4c79 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1497"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_3.yaml
index 35ae99e403..33de27f3b0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_3.yaml
@@ -141,7 +141,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2674"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/glm5-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/glm5-mtp.yaml
index b4a8864faf..7a8cc0a7aa 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/glm5-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp4/glm5-mtp.yaml
@@ -114,7 +114,8 @@ base:
     max_attempts: 360
     interval_seconds: 10
   benchmark:
-    type: sa-bench
+    type: custom
+    command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
     req_rate: inf
 # ################# 8k1k #################
 zip_override_mtp_8k1k_hightpt:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml
index 854b4023a5..a663b68066 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_hightpt_0.yaml
@@ -124,7 +124,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
index 2ac9997129..8bb57af814 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_0.yaml
@@ -122,7 +122,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
index 1de1730dce..44e86dabb8 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_1.yaml
@@ -122,7 +122,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
index 4efada1878..32506810e3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_2.yaml
@@ -122,7 +122,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
index 23d65f5c5e..70302685f7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_3.yaml
@@ -113,7 +113,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
index 784ced1656..e1540ffaad 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_4.yaml
@@ -113,7 +113,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
index f611c7ad62..acf7121b35 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/1k1k/disagg/stp/1k1k_stp_lowlat_5.yaml
@@ -113,7 +113,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 1024
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml
index 80b88d7964..3ef558666c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_0.yaml
@@ -124,7 +124,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml
index 734aec020f..04072d4bdf 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_1.yaml
@@ -124,7 +124,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml
index feb1c19b73..f715e91a17 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_2.yaml
@@ -124,7 +124,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml
index 29da98c953..a805e75acb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_hightpt_3.yaml
@@ -124,7 +124,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
index 385063085c..a9bbaa821b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
@@ -112,7 +112,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
index 035dda552d..ad7b1b07a9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
@@ -112,7 +112,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
index d2dadc3fa3..d27af50f81 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
@@ -112,7 +112,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   req_rate: inf
   isl: 8192
   osl: 1024
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/glm5-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/glm5-mtp.yaml
index 372c794079..b62dd3c51d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/glm5-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb200-fp8/glm5-mtp.yaml
@@ -103,7 +103,8 @@ base:
     max_attempts: 360
     interval_seconds: 10
   benchmark:
-    type: sa-bench
+    type: custom
+    command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
     req_rate: inf
     # MTP: EAGLE-style spec decoding is trained against chat-formatted
     # inputs — keep the chat template on explicitly rather than relying
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
index 31ac6edf8e..18517d3a04 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
@@ -166,7 +166,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
index 95ffd216b7..ca8a45f14f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_1.yaml
@@ -166,7 +166,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "64"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
index 506dede54f..26d4bf46af 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_2.yaml
@@ -166,7 +166,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "32"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
index 43e96c0a32..40f529081f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_3.yaml
@@ -166,7 +166,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
index 2bcc483e9f..216c140896 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_4.yaml
@@ -166,7 +166,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "12"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-dep4-dep16.yaml
index c110e15993..0fd906e6a6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-dep4-dep16.yaml
@@ -147,7 +147,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-tp4-tp4.yaml
index 143f339d0e..fbdd081d35 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-tp4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/1p1d-tp4-tp4.yaml
@@ -113,7 +113,8 @@ backend:
       disaggregation-mode: "decode"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/2p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/2p1d-dep4-dep16.yaml
index 7b4ae6a03e..85f8a32b1b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/2p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/1k1k/2p1d-dep4-dep16.yaml
@@ -147,7 +147,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml
index d869b247a7..42919159e0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml
@@ -113,7 +113,8 @@ backend:
       disaggregation-mode: "decode"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/4p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/4p1d-dep4-dep16.yaml
index 0eb6c58817..d7ecf653fd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/4p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/4p1d-dep4-dep16.yaml
@@ -150,7 +150,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/8p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/8p1d-dep4-dep16.yaml
index 58f2604a78..54704714d2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/8p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/8p1d-dep4-dep16.yaml
@@ -153,7 +153,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
index df82d47501..f632db58a0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_0.yaml
@@ -134,7 +134,8 @@ backend:
       disaggregation-mode: "decode"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
index 544b98a962..6547ba7caf 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_lowlat_1.yaml
@@ -146,7 +146,8 @@ backend:
       disaggregation-transfer-backend: "mooncake"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_0.yaml
index 486efbdfe9..67923f85fe 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_0.yaml
@@ -162,7 +162,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_1.yaml
index f4a1432310..cf5837a22a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_1.yaml
@@ -162,7 +162,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_2.yaml
index bde189a965..3f4e37e645 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_2.yaml
@@ -162,7 +162,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_3.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_3.yaml
index 0fc8db0c2b..15c1085447 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_3.yaml
@@ -162,7 +162,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_4.yaml
index 2d8bacfc92..75f001c463 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/disagg/mtp/8k1k_mtp_maxtpt_4.yaml
@@ -162,7 +162,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
index 71a0ffb9f3..55c7b4f0cb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml
@@ -1,6 +1,6 @@
 # Qwen3.5-397B-A17B-NVFP4 Disaggregated 1P1D: TP4 Prefill + TP4 Decode
 # Pure tensor parallel, no expert parallel (STP)
-# 8k1k sa-bench concurrency sweep on GB300
+# 8k1k custom benchmark concurrency sweep on GB300
 #
 # Values taken from ni_experiment_config of the
 # sa-qwen-3.5-8k1k-fp4-baseline-low-latency study, row
@@ -167,7 +167,8 @@ backend:
       decode-log-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1x4x8x16x32x64x256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
index 00e576439a..a529d0e6b7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_0.yaml
@@ -169,7 +169,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
index ecab285097..1b7428d593 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_1.yaml
@@ -167,7 +167,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "5120"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
index d35f444694..8cfe8ccff6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_maxtpt_2.yaml
@@ -167,7 +167,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "5120"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-dep4-dep16.yaml
index 39e5315059..f1c82b84fd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-dep4-dep16.yaml
@@ -146,7 +146,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-tp4-tp4.yaml
index 90eaa12aae..ba5fdd2049 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-tp4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/1p1d-tp4-tp4.yaml
@@ -113,7 +113,8 @@ backend:
       disaggregation-mode: "decode"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/2p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/2p1d-dep4-dep16.yaml
index 334636ebb9..abca86c4d4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/2p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/1k1k/2p1d-dep4-dep16.yaml
@@ -146,7 +146,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml
index 0a938fb48c..7775da8e94 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml
@@ -113,7 +113,8 @@ backend:
       disaggregation-mode: "decode"
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/4p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/4p1d-dep4-dep16.yaml
index 7ad092cfd4..0f841bf0c3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/4p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/4p1d-dep4-dep16.yaml
@@ -149,7 +149,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/8p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/8p1d-dep4-dep16.yaml
index 61da0a4ad8..cae55f0a0e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/8p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/8p1d-dep4-dep16.yaml
@@ -152,7 +152,8 @@ backend:
       stream-interval: 50
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: "inf"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch32_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch32_eplb0_mtp3.yaml
index 13da898bc4..bab412692b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch32_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch32_eplb0_mtp3.yaml
@@ -100,7 +100,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '615'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch64_eplb0_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch64_eplb0_mtp1.yaml
index 73b2d10adc..e34c1947fd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch64_eplb0_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep16_batch64_eplb0_mtp1.yaml
@@ -104,7 +104,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '1229'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch16_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch16_eplb0_mtp3.yaml
index 37912a792f..6bbc79017b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch16_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch16_eplb0_mtp3.yaml
@@ -98,7 +98,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '666'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch4_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch4_eplb0_mtp3.yaml
index 9d09244308..316dfa6e85 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch4_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch4_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '180'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch8_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch8_eplb0_mtp3.yaml
index 5af56e0452..cfc206093a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch8_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen1dep32_batch8_eplb0_mtp3.yaml
@@ -97,7 +97,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '308'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch16_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch16_eplb0_mtp3.yaml
index 6a40ed810c..8f93b714ed 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch16_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch16_eplb0_mtp3.yaml
@@ -99,7 +99,8 @@ backend:
         num_nextn_predict_layers: 3
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '92'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml
index e2919b07c9..d24277a4ac 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch1_eplb0_mtp3.yaml
@@ -97,7 +97,8 @@ backend:
         num_nextn_predict_layers: 3
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '8'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch4_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch4_eplb0_mtp3.yaml
index cb043e80a7..0e96e939c6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch4_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch4_eplb0_mtp3.yaml
@@ -97,7 +97,8 @@ backend:
         num_nextn_predict_layers: 3
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '24'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch8_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch8_eplb0_mtp3.yaml
index 87fc3c9899..e215a5d18e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch8_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen4tep8_batch8_eplb0_mtp3.yaml
@@ -98,7 +98,8 @@ backend:
         num_nextn_predict_layers: 3
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '40'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
index 3b76d83207..1f03ab345e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '10'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch128_eplb0_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch128_eplb0_mtp1.yaml
index ae75aea604..ff7eb7316a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch128_eplb0_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch128_eplb0_mtp1.yaml
@@ -112,7 +112,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '2253'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch256_eplb256_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch256_eplb256_mtp1.yaml
index ed5867951f..529ccd19a6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch256_eplb256_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep16_batch256_eplb256_mtp1.yaml
@@ -131,7 +131,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '4301'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep8_batch512_eplb0_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep8_batch512_eplb0_mtp1.yaml
index d21c50f2f9..4a11364b69 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep8_batch512_eplb0_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/MTP/ctx2dep4_gen1dep8_batch512_eplb0_mtp1.yaml
@@ -160,7 +160,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '4301'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch16_eplb0_mtp0.yaml
index d9b2a84884..da3839cf1b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch16_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch16_eplb0_mtp0.yaml
@@ -92,7 +92,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '666'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch32_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch32_eplb0_mtp0.yaml
index 9d0cfb08c3..18107fa606 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch32_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep32_batch32_eplb0_mtp0.yaml
@@ -94,7 +94,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '1229'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep8_batch512_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep8_batch512_eplb0_mtp0.yaml
index f5b2378ceb..9be5d9c83a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep8_batch512_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen1dep8_batch512_eplb0_mtp0.yaml
@@ -154,7 +154,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '4301'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch16_eplb0_mtp0.yaml
index 8a4ab2820d..9b256d1b8b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch16_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch16_eplb0_mtp0.yaml
@@ -93,7 +93,8 @@ backend:
           - cuda_core
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '84'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
index 9a46c38543..c092e01b87 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch1_eplb0_mtp0.yaml
@@ -91,7 +91,8 @@ backend:
           - cuda_core
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '4'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch32_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch32_eplb0_mtp0.yaml
index b5cf5ee77e..67e0f34879 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch32_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch32_eplb0_mtp0.yaml
@@ -95,7 +95,8 @@ backend:
           - cuda_core
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '168'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
index 61473a1f7f..0913d27ded 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch4_eplb0_mtp0.yaml
@@ -91,7 +91,8 @@ backend:
           - cuda_core
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '20'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch64_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch64_eplb0_mtp0.yaml
index f99df622a3..d87df1fdec 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch64_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen4tep8_batch64_eplb0_mtp0.yaml
@@ -99,7 +99,8 @@ backend:
           - cuda_core
       allreduce_strategy: MNNVL
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '284'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
index 7ef8e996f0..0276e0f9ec 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
@@ -90,7 +90,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '5'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
index e3df06c70d..6c33d6aba0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
@@ -90,7 +90,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '25'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch128_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch128_eplb0_mtp0.yaml
index f3166be8c8..c9eaf58bf7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch128_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch128_eplb0_mtp0.yaml
@@ -106,7 +106,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '2151'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch256_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch256_eplb0_mtp0.yaml
index 4c588afcf6..1446d3dadd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch256_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep16_batch256_eplb0_mtp0.yaml
@@ -122,7 +122,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '4301'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep32_batch64_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep32_batch64_eplb0_mtp0.yaml
index 4fa9b93291..ed6bf8693d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep32_batch64_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL1K_OSL1K/STP/ctx2dep4_gen1dep32_batch64_eplb0_mtp0.yaml
@@ -98,7 +98,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: '2151'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep16_batch128_eplb0_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep16_batch128_eplb0_mtp1.yaml
index 1d555a2504..3d4e5229a6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep16_batch128_eplb0_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx12dep4_gen1dep16_batch128_eplb0_mtp1.yaml
@@ -112,7 +112,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '2253'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen1dep32_batch2_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen1dep32_batch2_eplb0_mtp3.yaml
index 9662511d2d..e409cd1d1d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen1dep32_batch2_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen1dep32_batch2_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '90'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
index f82c0c2c6f..637690d5b3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch1_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '5'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml
index 99ff55044f..980fc2206d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch2_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '15'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml
index d72a5d0f2e..0e4b95f827 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx1dep4_gen5tep4_batch4_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '30'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch4_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch4_eplb0_mtp3.yaml
index 5488068262..36a9fc2683 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch4_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx2dep4_gen1dep32_batch4_eplb0_mtp3.yaml
@@ -96,7 +96,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '180'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb0_mtp3.yaml
index 13905b0957..5a481e5d78 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx4dep4_gen1dep32_batch8_eplb0_mtp3.yaml
@@ -97,7 +97,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '333'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb0_mtp3.yaml
index 0d28018a20..67b69cc205 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx6dep4_gen1dep16_batch32_eplb0_mtp3.yaml
@@ -100,7 +100,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '615'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep32_batch16_eplb0_mtp3.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep32_batch16_eplb0_mtp3.yaml
index 3b1c6affcd..144a0ef948 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep32_batch16_eplb0_mtp3.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx7dep4_gen1dep32_batch16_eplb0_mtp3.yaml
@@ -98,7 +98,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 3
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '666'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep16_batch64_eplb0_mtp1.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep16_batch64_eplb0_mtp1.yaml
index 48a2401410..3ca18f15d6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep16_batch64_eplb0_mtp1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/MTP/ctx8dep4_gen1dep16_batch64_eplb0_mtp1.yaml
@@ -104,7 +104,8 @@ backend:
         decoding_type: MTP
         num_nextn_predict_layers: 1
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '1127'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
index e4ec638273..0a5f8ae589 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch16_eplb0_mtp0.yaml
@@ -92,7 +92,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '105'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
index 4ed13a020f..f53536c980 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch1_eplb0_mtp0.yaml
@@ -90,7 +90,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '5'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml
index 8c8c759822..e2c826b125 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch2_eplb0_mtp0.yaml
@@ -90,7 +90,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '10'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
index 5019bb20ed..723876fc52 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch4_eplb0_mtp0.yaml
@@ -90,7 +90,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '25'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
index 4f3971b6c7..1527c57f91 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx1dep4_gen5tep4_batch8_eplb0_mtp0.yaml
@@ -91,7 +91,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '50'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
index 97b70e4833..6ca30a752a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx2dep4_gen1dep32_batch8_eplb0_mtp0.yaml
@@ -91,7 +91,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '308'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml
index 2a48823f9d..0580e5ddd7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx4dep4_gen1dep32_batch16_eplb0_mtp0.yaml
@@ -92,7 +92,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '615'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb0_mtp0.yaml
index a2f97d489e..291f7e00c7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx5dep4_gen1dep16_batch64_eplb0_mtp0.yaml
@@ -98,7 +98,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '1127'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml
index 461d14f5d3..77affe86d6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx6dep4_gen1dep32_batch32_eplb0_mtp0.yaml
@@ -94,7 +94,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '1229'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb0_mtp0.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb0_mtp0.yaml
index ad55ac6056..0dfe7f1190 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb0_mtp0.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5/disagg/trtllm_dynamo/gb200_nvfp4/ISL8K_OSL1K/STP/ctx9dep4_gen1dep16_batch128_eplb0_mtp0.yaml
@@ -106,7 +106,8 @@ backend:
           - cutedsl
           - cuda_core
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: '2151'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p1d-dep4-dep8-c308-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p1d-dep4-dep8-c308-stp.yaml
index 1cb6ab0de4..f22f7a284e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p1d-dep4-dep8-c308-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p1d-dep4-dep8-c308-stp.yaml
@@ -119,7 +119,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c24-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c24-stp.yaml
index 49a71471de..595eefea58 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c24-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c24-stp.yaml
@@ -115,7 +115,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c4-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c4-stp.yaml
index cabec31b18..a9b6614dc8 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c4-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p4d-dep4-tep8-c4-stp.yaml
@@ -115,7 +115,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c115-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c115-stp.yaml
index 298279a010..f548fb4139 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c115-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c115-stp.yaml
@@ -117,7 +117,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c195-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c195-stp.yaml
index f31a7da80e..c21aa53211 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c195-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c195-stp.yaml
@@ -119,7 +119,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c30-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c30-stp.yaml
index 4c4c833192..a802a8c9b4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c30-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c30-stp.yaml
@@ -115,7 +115,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c5-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c5-stp.yaml
index 619c1bb709..18ec20c235 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c5-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c5-stp.yaml
@@ -115,7 +115,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c60-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c60-stp.yaml
index 1a8b4f2988..3d2a791556 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c60-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-1p5d-dep4-tep4-c60-stp.yaml
@@ -116,7 +116,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-2p1d-dep4-dep8-c615-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-2p1d-dep4-dep8-c615-stp.yaml
index abcce625e7..3acf877ea4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-2p1d-dep4-dep8-c615-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-2p1d-dep4-dep8-c615-stp.yaml
@@ -123,7 +123,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-3p1d-dep4-dep8-c1127-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-3p1d-dep4-dep8-c1127-stp.yaml
index 316a88f23a..50be2ab1e3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-3p1d-dep4-dep8-c1127-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-3p1d-dep4-dep8-c1127-stp.yaml
@@ -131,7 +131,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-4p1d-dep4-dep8-c2151-stp.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-4p1d-dep4-dep8-c2151-stp.yaml
index 893a725cae..5999e094cb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-4p1d-dep4-dep8-c2151-stp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4/disagg-B200-4p1d-dep4-dep8-c2151-stp.yaml
@@ -147,7 +147,8 @@ frontend:
   env:
     ETCD_LEASE_TTL: '120'
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-gb200-low-latency-mtp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-gb200-low-latency-mtp2.yaml
index ffc8e4ea1a..388a23f772 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-gb200-low-latency-mtp2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/agg-gb200-low-latency-mtp2.yaml
@@ -68,7 +68,8 @@ backend:
       gpu-memory-utilization: 0.9
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
index e81ef8631c..277452978a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
@@ -100,7 +100,8 @@ benchmark:
   isl: 8192
   osl: 1024
   req_rate: "inf"
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   use_chat_template: true
 identity:
   model:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
index b296440158..368a0c2fa2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
@@ -100,7 +100,8 @@ benchmark:
   isl: 8192
   osl: 1024
   req_rate: "inf"
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   use_chat_template: true
 identity:
   model:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
index b566c05fae..03b72308d3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
@@ -100,7 +100,8 @@ benchmark:
   isl: 8192
   osl: 1024
   req_rate: "inf"
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   use_chat_template: true
 identity:
   model:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
index 43190836a1..74574f3e89 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
@@ -104,7 +104,8 @@ benchmark:
   isl: 8192
   osl: 1024
   req_rate: "inf"
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   use_chat_template: true
 identity:
   model:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
index 283bb839f3..89fa3b12dc 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
@@ -100,7 +100,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml
index 9247a31daf..f2ecf85864 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml
@@ -111,7 +111,8 @@ backend:
       reasoning-parser: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml
index 260e16fc7d..858bb42c20 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml
@@ -108,7 +108,8 @@ backend:
       reasoning-parser: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1x32x64x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml
index 98b701790a..60104c28ab 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml
@@ -111,7 +111,8 @@ backend:
       reasoning-parser: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256x1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe-mtp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe-mtp2.yaml
index 24b134c1b9..37fe90fa0a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe-mtp2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe-mtp2.yaml
@@ -127,7 +127,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe.yaml
index a78b83ce2e..2491911749 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-high-tpt-megamoe.yaml
@@ -136,7 +136,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency-mtp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency-mtp2.yaml
index 7e0d09a0ef..5ad64359f9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency-mtp2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency-mtp2.yaml
@@ -113,7 +113,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: 16x32x64
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency.yaml
index 6c5dec41d5..0939ef1b4c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-latency.yaml
@@ -133,7 +133,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-middle-curve.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-middle-curve.yaml
index 98ec9acdad..b9644cc1e4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-middle-curve.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-low-middle-curve.yaml
@@ -135,7 +135,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256x512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-max-tpt-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-max-tpt-megamoe.yaml
index db3caef440..9abb2dee31 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-max-tpt-megamoe.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-max-tpt-megamoe.yaml
@@ -136,7 +136,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe-mtp2.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe-mtp2.yaml
index 813584eb90..98fc405785 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe-mtp2.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe-mtp2.yaml
@@ -127,7 +127,8 @@ backend:
       tokenizer-mode: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "128x256x512x1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe.yaml
index a1f7cfd5a6..bf571bdce7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb200-mid-curve-megamoe.yaml
@@ -136,7 +136,8 @@ backend:
       enable-sleep-mode: true
       tokenizer-mode: deepseek_v4
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256x512x1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p6d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p6d-dep4-tp4.yaml
index 7a299d81a5..06a1dd130a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p6d-dep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p6d-dep4-tp4.yaml
@@ -107,10 +107,11 @@ backend:
       tokenizer-mode: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "192"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p9d-tep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p9d-tep4-tp4.yaml
index d1e895bbbf..f0bc4bf734 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p9d-tep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p9d-tep4-tp4.yaml
@@ -98,10 +98,11 @@ backend:
       tokenizer-mode: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "18"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep8-24-c4096.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep8-24-c4096.yaml
index 0a3643cf27..0b3ac88329 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep8-24-c4096.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-4p1d-dep4-dep8-24-c4096.yaml
@@ -116,10 +116,11 @@ backend:
       no-enable-flashinfer-autotune: true
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-5p1d-dep4-dep8-28-c4096.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-5p1d-dep4-dep8-28-c4096.yaml
index 698ef94e2b..cd22158c45 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-5p1d-dep4-dep8-28-c4096.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-5p1d-dep4-dep8-28-c4096.yaml
@@ -116,10 +116,11 @@ backend:
       no-enable-flashinfer-autotune: true
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-6p1d-dep4-dep8-32-c4096.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-6p1d-dep4-dep8-32-c4096.yaml
index 85a1d4c6b6..9ff033c6ec 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-6p1d-dep4-dep8-32-c4096.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-6p1d-dep4-dep8-32-c4096.yaml
@@ -116,10 +116,11 @@ backend:
       no-enable-flashinfer-autotune: true
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-7p2d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-7p2d-dep4-dep16.yaml
index 53534f9697..c70b580494 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-7p2d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-gb300-7p2d-dep4-dep16.yaml
@@ -113,10 +113,11 @@ backend:
       tokenizer-mode: deepseek_v4
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "3072"
   req_rate: "inf"
-  tokenizer_mode: "deepseek_v4"
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
   use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml
index b647276179..4ec6ba81d6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-agentic.yaml
@@ -148,7 +148,6 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml
index 9e2b665843..8405286b88 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p2d-tep8-tp8-agentic.yaml
@@ -142,7 +142,6 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml
index 13c8d353e3..9f95983352 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p6d-dep4-tp4-agentic.yaml
@@ -6,7 +6,7 @@ name: "svf-vllm-disagg-gb300-1p6d-dep4-tp4-agentic"
 # the fixed-seq-len 1p6d baseline at the same concurrency point (192).
 #
 # Divergence vs the 8k1k sibling:
-#   - benchmark.type:        sa-bench -> custom (hands off to agentic_srt.sh)
+#   - benchmark.type:        custom (hands off to agentic_srt.sh)
 #   - max-model-len:         removed (let vLLM derive from model config; agentic
 #                              trajectories blow past any small explicit cap)
 #   - no-enable-prefix-caching: dropped (prefix caching MUST be on for
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-4p1d-dep4-dep8-24-c4096-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-4p1d-dep4-dep8-24-c4096-agentic.yaml
index 05b779d541..4fec9c0164 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-4p1d-dep4-dep8-24-c4096-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-4p1d-dep4-dep8-24-c4096-agentic.yaml
@@ -6,7 +6,7 @@ name: "svf-vllm-disagg-gb300-4p1d-dep4-dep8-24-c4096-agentic"
 # node. Sized for concurrency 4096 with deep_gemm_mega_moe on both workers.
 #
 # Divergence vs the 8k1k sibling:
-#   - benchmark.type:        sa-bench -> custom (hands off to agentic_srt.sh)
+#   - benchmark.type:        custom (hands off to agentic_srt.sh)
 #   - max-model-len:         removed (let vLLM derive from model config; agentic
 #                              trajectories blow past any small explicit cap)
 #   - no-enable-prefix-caching: dropped (prefix caching MUST be on for
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml
index 0ee4f3e7bc..2249645140 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep16.yaml
@@ -93,7 +93,8 @@ backend:
       max-cudagraph-capture-size: 384
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4096x6144"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep8.yaml
index 40683d1b8d..d627c790bb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p1d-dep4-dep8.yaml
@@ -93,7 +93,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4096x12288"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p4d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p4d-dep4-tp8.yaml
index bda68b9207..67bed3b38c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p4d-dep4-tp8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb200-1p4d-dep4-tp8.yaml
@@ -89,7 +89,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4x8x32x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep16.yaml
index 7188a5a082..72eff7bf9f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep16.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 256
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "2048x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep24.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep24.yaml
index 79758b546f..d020dd9b01 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep24.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p1d-dep4-dep24.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 128
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p2d-dep4-dep4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p2d-dep4-dep4.yaml
index 15d4169abc..830c70a06c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p2d-dep4-dep4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p2d-dep4-dep4.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "6144"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p7d-tep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p7d-tep4-tp4.yaml
index a0ba50b871..20f2060b14 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p7d-tep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-1p7d-tep4-tp4.yaml
@@ -87,7 +87,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "8x16x32x64x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-2p3d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-2p3d-dep4-dep8.yaml
index cfa6037040..8b415ceb2c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-2p3d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/1k1k/disagg-gb300-2p3d-dep4-dep8.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml
index 429945c05e..5e34443e79 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tep4.yaml
@@ -89,7 +89,8 @@ backend:
       max-cudagraph-capture-size: 16
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tp8.yaml
index 55165aeb3b..c6da14d0c2 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tp8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-1p4d-dep4-tp8.yaml
@@ -89,7 +89,8 @@ backend:
       max-cudagraph-capture-size: 16
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4x8x16x32x256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml
index b4d1e8c82e..f4625a2e93 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-3p1d-dep4-dep16.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 256
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml
index 1f4de8df10..022a7f3635 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-6p1d-dep4-dep16.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 512
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "3072"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-8p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-8p1d-dep4-dep16.yaml
index 45a7214b2d..5e6427b6f9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-8p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb200-8p1d-dep4-dep16.yaml
@@ -93,7 +93,8 @@ backend:
       max-cudagraph-capture-size: 512
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "6144"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p4d-dep4-tp8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p4d-dep4-tp8.yaml
index 0ed5673478..665e0a66ad 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p4d-dep4-tp8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p4d-dep4-tp8.yaml
@@ -89,7 +89,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p8d-dep4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p8d-dep4-tp4.yaml
index 71d897b97d..554b351068 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p8d-dep4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-1p8d-dep4-tp4.yaml
@@ -89,7 +89,8 @@ backend:
       max-cudagraph-capture-size: 1024
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "32x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-2p1d-dep4-dep24.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-2p1d-dep4-dep24.yaml
index b43b9f1367..dbd894d725 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-2p1d-dep4-dep24.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-2p1d-dep4-dep24.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 128
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-3p1d-dep4-dep16.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-3p1d-dep4-dep16.yaml
index 7ee94c9d10..e399ca0fb5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-3p1d-dep4-dep16.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-3p1d-dep4-dep16.yaml
@@ -92,7 +92,8 @@ backend:
       max-cudagraph-capture-size: 128
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-4p1d-dep4-dep8.yaml
index 23fc7f67f0..275b5c3137 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-4p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-4p1d-dep4-dep8.yaml
@@ -93,7 +93,8 @@ backend:
       max-cudagraph-capture-size: 512
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "3072"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-8p1d-dep4-dep24.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-8p1d-dep4-dep24.yaml
index 8b2f2bf80c..6c4a0e3b21 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-8p1d-dep4-dep24.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4/8k1k/disagg-gb300-8p1d-dep4-dep24.yaml
@@ -93,7 +93,8 @@ backend:
       max-cudagraph-capture-size: 128
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "15360"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml
index 710c4d9796..b859b058ef 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-dep8-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1024x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml
index c1b9cf32d6..963f9ba4d9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p1d-dep4-tep8-1k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 8196
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "128x256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml
index 907633ba74..6134dfc637 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-dep8-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1024x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml
index 2e2bbc12a9..edae8be6ac 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p2d-dep4-tep8-1k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 8196
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4x16x64"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml
index 429aa015ea..fc817c5507 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/1k1k/1p4d-dep4-tp4-marlin-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 8196
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "32"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml
index 4129408ebd..10a1c82ce3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml
index 5bde9ac9be..18b63411b6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep4-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml
index 56ed02e5e8..4b9ce2e876 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/1p2d-dep4-tep8-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4x16x32x64x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml
index 22ebfc7836..38e183c3fe 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/2p2d-dep4-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512x1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml
index ba1db206e4..c3aed93778 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/3p2d-dep4-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml
index a54ddcb226..0ab797f970 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8/8k1k/5p2d-dep4-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
index f57d7af090..b7ef8410f1 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p1d-dep2-dep4-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
index b4f457654a..76be752718 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/1p2d-dep2-tep8-1k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 8196
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4x16x64x128x256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
index 6bba9ea864..cad7e88dac 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 8196
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "32"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
index de852e427a..e422be56f0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p3d-dep2-dep4-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml
index 8f7b7b140e..5f12be921a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/2p4d-dep2-dep4-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "8192"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
index f6cf6a59f0..4a00a96e38 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/1k1k/4p2d-dep2-dep8-1k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1024x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
index d990d661b9..3158ca470c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
index d46133924c..130a32897f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p1d-dep2-tep8-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
index e8c606e271..c7a51865cf 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/1p2d-dep2-tep8-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "32x64x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
index 02c3be14ac..bd64ee4d32 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p1d-dep2-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
index 304650d6c1..daaec01a6e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml
index efea8bfacf..3765bec505 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/2p4d-dep2-tep4-8k1k.yaml
@@ -98,7 +98,8 @@ backend:
       max-cudagraph-capture-size: 4096
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
index 97e1ec88c8..5b31a0f2a6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p1d-dep2-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
index 745b2fad44..222b5547e5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/3p2d-dep2-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml
index 9be5cc1772..61fd8a6f6c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8/8k1k/6p1d-dep2-dep8-8k1k.yaml
@@ -100,7 +100,8 @@ backend:
       max-cudagraph-capture-size: 2048
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tep8-1k1k.yaml
index 486af05573..2db08084e4 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tep8-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4x16x64x128x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tp4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tp4-1k1k.yaml
index 532b78a103..c840f29f9e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tp4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p1d-dep2-tp4-1k1k.yaml
@@ -75,7 +75,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1x4x8x16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p2d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p2d-dep2-dep4-1k1k.yaml
index fde8442a18..e3bdf4788a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p2d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/1p2d-dep2-dep4-1k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-dep8-1k1k.yaml
index ed3b5f9950..45bc95a331 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-dep8-1k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "512x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-tep8-1k1k.yaml
index 0784283b91..85d6c26e21 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p1d-dep2-tep8-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "32"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p2d-dep2-tep8-1k1k.yaml
index 59c52da00c..6a2d2601eb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/2p2d-dep2-tep8-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/3p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/3p2d-dep2-tep8-1k1k.yaml
index 7e9f7dec31..3b531b5d1f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/3p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/1k1k/3p2d-dep2-tep8-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp4-8k1k.yaml
index a98e9f9c83..e120bd1106 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp4-8k1k.yaml
@@ -66,7 +66,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "8x16x24x32x48x64"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp8-8k1k.yaml
index 5f699bbcf7..9339941265 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/1p1d-tp1-tp8-8k1k.yaml
@@ -65,7 +65,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1x2x4x8x16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/2p1d-tp1-tp4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/2p1d-tp1-tp4-8k1k.yaml
index 7b527ce70e..8649e04a1a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/2p1d-tp1-tp4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/2p1d-tp1-tp4-8k1k.yaml
@@ -66,7 +66,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/3p1d-tp1-tp4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/3p1d-tp1-tp4-8k1k.yaml
index 7bf977c39e..27fb9a0e1b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/3p1d-tp1-tp4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/3p1d-tp1-tp4-8k1k.yaml
@@ -66,7 +66,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p1d-tp1-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p1d-tp1-tep4-8k1k.yaml
index fd11bf499f..32a04855f7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p1d-tp1-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p1d-tp1-tep4-8k1k.yaml
@@ -66,7 +66,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p2d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p2d-dep2-tep4-8k1k.yaml
index 23c99d3282..95cf8af823 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p2d-dep2-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/4p2d-dep2-tep4-8k1k.yaml
@@ -78,7 +78,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/5p2d-tp1-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/5p2d-tp1-dep8-8k1k.yaml
index 55cc00b611..a8090be5d5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/5p2d-tp1-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/5p2d-tp1-dep8-8k1k.yaml
@@ -67,7 +67,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/8p2d-tp1-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/8p2d-tp1-dep8-8k1k.yaml
index 73525d8c42..e089eec422 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/8p2d-tp1-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/8p2d-tp1-dep8-8k1k.yaml
@@ -67,7 +67,8 @@ health_check:
   max_attempts: 360
   interval_seconds: 10
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "768x1024"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
index 165301c53c..2fdb0701d3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p1d-dep2-tp4-eagle3-8k1k.yaml
@@ -82,7 +82,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
index b32295e325..564b8a214a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p2d-dep2-tp4-eagle3-8k1k.yaml
@@ -82,7 +82,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
index ff1b0c73c7..1877bb1569 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p4d-dep2-tp4-eagle3-8k1k.yaml
@@ -82,7 +82,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
index c5f096d64d..333802b334 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/1p6d-dep2-tp4-eagle3-8k1k.yaml
@@ -82,7 +82,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
index 5943109612..3a6635f992 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p1d-dep2-dep4-eagle3-8k1k.yaml
@@ -86,7 +86,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
index 790e6759a4..477b305010 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/2p3d-dep2-tp4-eagle3-8k1k.yaml
@@ -82,7 +82,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: sa-bench
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   req_rate: inf
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
index 5bbb133628..e61b17525b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tep8-1k1k.yaml
@@ -72,7 +72,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4x16x64x128x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp4-marlin-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp4-marlin-1k1k.yaml
index 49a60981ee..b012823f69 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp4-marlin-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p1d-dep2-tp4-marlin-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "1x4x8x16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
index ef7e66d765..1228e5863f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/1p2d-dep2-dep4-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "2048"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
index 9f5aa341cf..d3c1e0a33e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-dep8-1k1k.yaml
@@ -74,7 +74,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "512x4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
index 42c6e7bbc2..72893e968f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p1d-dep2-tep8-1k1k.yaml
@@ -72,7 +72,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "32"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
index 3e701df05e..668eb5289a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/2p2d-dep2-tep8-1k1k.yaml
@@ -72,7 +72,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
index b9a1d10588..23bf5abf4e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/1k1k/3p2d-dep2-tep8-1k1k.yaml
@@ -72,7 +72,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 1024
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml
index 04aca65862..1b1959900d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p1d-dep2-tp4-marlin-8k1k.yaml
@@ -78,7 +78,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "1x4x8x16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml
index e483108986..7004f72ff8 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p2d-dep2-tep4-8k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml
index 30ac635a9a..fdfb00c4e3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep4-8k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "16x32x64x128"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep8-8k1k.yaml
index 46af72e46c..56e4f0b4ad 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/1p4d-dep2-tep8-8k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
index b1558ae34f..6bd7ffbfa0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-dep8-8k1k.yaml
@@ -78,7 +78,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "256x512"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
index 46aaa045da..8e2e70baa3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/2p2d-dep2-tep8-8k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "16"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
index 3756103eed..1f9b3c759b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-dep8-8k1k.yaml
@@ -78,7 +78,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml
index c9f29f7852..69710288e3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3/b300-fp8/8k1k/4p2d-dep2-tep4-8k1k.yaml
@@ -76,7 +76,8 @@ health_check:
   interval_seconds: 10
 
 benchmark:
-  type: "sa-bench"
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
   isl: 8192
   osl: 1024
   concurrencies: "4096"
diff --git a/benchmarks/multi_node/srt_bench_serving_entrypoint.sh b/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
new file mode 100755
index 0000000000..89e9004a5e
--- /dev/null
+++ b/benchmarks/multi_node/srt_bench_serving_entrypoint.sh
@@ -0,0 +1,94 @@
+#!/usr/bin/env bash
+
+# InferenceX benchmark_serving adapter for srt-slurm's custom benchmark runner.
+
+set -eo pipefail
+
+if [[ $# -ne 15 ]]; then
+    echo "Expected 15 benchmark arguments, received $#" >&2
+    exit 64
+fi
+
+readonly MODEL=$1
+readonly TOKENIZER=$2
+readonly ISL=$3
+readonly OSL=$4
+readonly CONC_LIST=$5
+readonly REQ_RATE=$6
+readonly DISAGG=$7
+readonly PREFILL_GPUS=$8
+readonly DECODE_GPUS=$9
+readonly TOTAL_GPUS=${10}
+readonly RANDOM_RANGE_RATIO=${11}
+readonly NUM_PROMPTS_MULT=${12}
+readonly NUM_WARMUP_MULT=${13}
+readonly USE_CHAT_TEMPLATE=${14}
+readonly DSV4=${15}
+
+readonly INFERENCEX_WORKSPACE=/infmax-workspace
+readonly BENCH_HOST=127.0.0.1
+readonly BENCH_PORT=8000
+readonly BENCH_BACKEND=openai
+readonly TRUST_REMOTE_CODE=true
+readonly RESULT_DIR="/logs/inferencex-bench_isl_${ISL}_osl_${OSL}"
+export EVAL_ONLY=false
+export PROFILE=0
+
+# shellcheck disable=SC1091
+source "$INFERENCEX_WORKSPACE/benchmarks/benchmark_lib.sh"
+
+ensure_bench_serving_deps() {
+    local deps=(aiohttp numpy tqdm transformers huggingface_hub)
+    if python3 -c \
+        "import aiohttp, numpy, tqdm, transformers, huggingface_hub" \
+        2>/dev/null; then
+        return
+    fi
+
+    local venv="/tmp/inferencex-srt-bench-venv"
+    [[ -d "$venv" ]] || python3 -m venv --system-site-packages "$venv"
+    # shellcheck disable=SC1091
+    source "$venv/bin/activate"
+    pip install --quiet "${deps[@]}"
+}
+
+ensure_bench_serving_deps
+mkdir -p "$RESULT_DIR"
+
+curl -fsS "http://${BENCH_HOST}:${BENCH_PORT}/v1/models" >/dev/null || {
+    echo "InferenceX benchmark could not reach the srt-slurm frontend" >&2
+    exit 66
+}
+ulimit -n 65536 2>/dev/null || true
+
+read -r -a concurrencies <<< "${CONC_LIST//x/ }"
+for concurrency in "${concurrencies[@]}"; do
+    if [[ "$DISAGG" == "true" ]]; then
+        result_filename="results_concurrency_${concurrency}_gpus_${TOTAL_GPUS}_ctx_${PREFILL_GPUS}_gen_${DECODE_GPUS}"
+    else
+        result_filename="results_concurrency_${concurrency}_gpus_${TOTAL_GPUS}"
+    fi
+
+    optional_args=()
+    [[ "$USE_CHAT_TEMPLATE" == "true" ]] && optional_args+=(--use-chat-template)
+    [[ "$DSV4" == "true" ]] && optional_args+=(--dsv4)
+    [[ "$TRUST_REMOTE_CODE" == "true" ]] && optional_args+=(--trust-remote-code)
+
+    run_benchmark_serving \
+        --model "$MODEL" \
+        --tokenizer "$TOKENIZER" \
+        --host "$BENCH_HOST" \
+        --port "$BENCH_PORT" \
+        --backend "$BENCH_BACKEND" \
+        --input-len "$ISL" \
+        --output-len "$OSL" \
+        --random-range-ratio "$RANDOM_RANGE_RATIO" \
+        --num-prompts "$((concurrency * NUM_PROMPTS_MULT))" \
+        --num-warmups "$((concurrency * NUM_WARMUP_MULT))" \
+        --request-rate "$REQ_RATE" \
+        --max-concurrency "$concurrency" \
+        --result-filename "$result_filename" \
+        --result-dir "$RESULT_DIR" \
+        --bench-serving-dir "$INFERENCEX_WORKSPACE" \
+        "${optional_args[@]}"
+done
diff --git a/configs/srt-slurm-version.txt b/configs/srt-slurm-version.txt
new file mode 100644
index 0000000000..0b8957ae43
--- /dev/null
+++ b/configs/srt-slurm-version.txt
@@ -0,0 +1 @@
+v1.0.36
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index a276644575..721a6fe4d1 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -6,6 +6,8 @@ SLURM_ACCOUNT="benchmark"
 
 set -x
 
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
+
 # MODEL_PATH: Override with pre-downloaded paths on the shared Lustre tree.
 # Bench scripts and srt-slurm yaml configs specify HuggingFace model IDs for
 # portability, but we resolve to /lustre/fsw/models/* here to avoid repeated
@@ -96,49 +98,9 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
 
     export SERVED_MODEL_NAME=$MODEL
 
-    echo "Cloning srt-slurm repository..."
     SRT_REPO_DIR="srt-slurm"
-    if [ -d "$SRT_REPO_DIR" ]; then
-        echo "Removing existing $SRT_REPO_DIR..."
-        rm -rf "$SRT_REPO_DIR"
-    fi
-
-    # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" ]]; then
-        git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-    elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-        git checkout aflowers/vllm-gb200-v0.20.0
-        mkdir -p recipes/vllm/deepseek-v4
-        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
-    elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" && $PRECISION == "fp8" ]]; then
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-        git checkout main
-    elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-        # Pin srt-slurm: newer commits stopped honoring the hash-pinned dynamo
-        # build and fall back to a dynamo release that is incompatible with this
-        # sglang image (worker fails at import). This is the last commit before
-        # that change. Do not float on main -- the srtctl + dynamo-install
-        # toolchain is unpinned there.
-        git checkout a98738de9b2233459b5456e9ed71af09ce893f92
-        mkdir -p recipes/sglang/dsr1/b200-fp4
-        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/dsr1/b200-fp4" recipes/sglang/dsr1/b200-fp4
-    elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-        git checkout v1.0.29
-        mkdir -p recipes/trtllm/kimi-k25-nvfp4/b200-fp4
-        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4" recipes/trtllm/kimi-k25-nvfp4/b200-fp4
-    else
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR" || exit 1
-        git checkout sa-submission-q2-2026
-    fi
+    prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+    cd "$SRT_REPO_DIR" || exit 1
 
     echo "Installing srtctl..."
     export UV_INSTALL_DIR="$GITHUB_WORKSPACE/.local/bin"
@@ -261,6 +223,9 @@ EOF
     # so large-model loads (e.g. DSR1-FP8 ~680GB off shared FS) finish in time.
     # Uses ${CONFIG_FILE%%:*} because CONFIG_FILE may carry an :override[N] suffix.
     sed -i 's/^  max_attempts: [0-9]*/  max_attempts: 720/' "${CONFIG_FILE%%:*}"
+    CONFIG_FILE="$(
+        prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+    )" || exit 1
     SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "b200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
     echo "$SRTCTL_OUTPUT"
 
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 098e984be5..d9d2e69802 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -8,6 +8,8 @@ MINIMAX_M3_SLURM_EXCLUDED_NODELIST="${MINIMAX_M3_SLURM_EXCLUDED_NODELIST-b300-01
 
 set -x
 
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
+
 if [[ "$IS_MULTINODE" == "true" ]]; then
 
 # Validate framework
@@ -58,57 +60,13 @@ else
     exit 1
 fi
 
-echo "Cloning srt-slurm repository..."
 SRT_REPO_DIR="srt-slurm"
 SRTCTL_SETUP_SCRIPT=""
-if [ -d "$SRT_REPO_DIR" ]; then
-    echo "Removing existing $SRT_REPO_DIR..."
-    rm -rf "$SRT_REPO_DIR"
-fi
-
-# TODO(CJQ): make first class upon srt-slurm upstream refactor
-if [[ "$IS_AGENTIC" == "1" ]]; then
-    git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout aflowers/vllm-gb200-v0.20.0
-    mkdir -p recipes/vllm/deepseek-v4
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml ]]; then
-    git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout c1b6b5c97f323baefad577d70c4e8392b6f537d9
-    mkdir -p recipes/vllm/minimax-m3
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3" recipes/vllm/minimax-m3
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/*-tp1-*.yaml ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout c1fb6989fc5aca803b4ca0f2d17d8be85fad9732
-    mkdir -p recipes/vllm/minimax-m3
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3" recipes/vllm/minimax-m3
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && ( $PRECISION == "fp4" || $PRECISION == "fp8" ) ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout sa-submission-q2-2026
-    mkdir -p recipes/vllm/minimax-m3
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3" recipes/vllm/minimax-m3
-    if [[ $PRECISION == "fp8" ]]; then
-        SRTCTL_SETUP_SCRIPT="minimax-m3-vllm-fixes.sh"
-    fi
-    # NVIDIA/srt-slurm#38
-    git show 22d46ba9971615016d2339c9ffbc7b4597accfad --format= -- src/srtctl/core/ip_utils/get_node_ip.sh | git apply - || exit 1
-    if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
-        cp \
-            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
-            "configs/$SRTCTL_SETUP_SCRIPT"
-    fi
-else
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout sa-submission-q2-2026
+if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
+    SRTCTL_SETUP_SCRIPT="minimax-m3-vllm-fixes.sh"
 fi
+prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+cd "$SRT_REPO_DIR" || exit 1
 
 echo "Installing srtctl..."
 export UV_INSTALL_DIR="$GITHUB_WORKSPACE/.local/bin"
@@ -189,18 +147,23 @@ sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
 if [[ "$MODEL_PREFIX" == "minimaxm3" && -n "$MINIMAX_M3_SLURM_EXCLUDED_NODELIST" ]]; then
     sed -i "/^name:.*/a sbatch_directives:\n  exclude: \"${MINIMAX_M3_SLURM_EXCLUDED_NODELIST}\"" "$CONFIG_FILE"
 fi
-SRTCTL_APPLY_ARGS=(
-    -f "$CONFIG_FILE"
-    --tags "b300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
-)
 # The MTP and TP1 8k1k recipes use newer srt-slurm revisions whose preflight checks
 # model.path on this GHA login host. MiniMax-M3 NVFP4 is intentionally staged
 # under compute-node-local /scratch (as in the original B300 submission), so
 # the login host cannot stat it even though workers can. Keep this bypass
 # scoped to those recipe sets; runtime model loading still validates the path.
+PREFLIGHT_ARGS=()
 if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && ( "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/mtp/*.yaml || "$CONFIG_FILE" == recipes/vllm/minimax-m3/b300-fp4/8k1k/*-tp1-*.yaml ) ]]; then
-    SRTCTL_APPLY_ARGS+=(--no-preflight)
+    PREFLIGHT_ARGS=(--no-preflight)
 fi
+CONFIG_FILE="$(
+    prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+)" || exit 1
+SRTCTL_APPLY_ARGS=(
+    "${PREFLIGHT_ARGS[@]}"
+    -f "$CONFIG_FILE"
+    --tags "b300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
+)
 if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then
     SRTCTL_APPLY_ARGS+=(--setup-script "$SRTCTL_SETUP_SCRIPT")
 fi
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 426cd2dd87..48e0768496 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -5,6 +5,7 @@
 set -x
 
 source "$(dirname "${BASH_SOURCE[0]}")/slurm_utils.sh"
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
 
 export SLURM_PARTITION="batch"
 export SLURM_ACCOUNT="benchmark"
@@ -337,7 +338,6 @@ if [[ -z "$CONFIG_FILE" ]]; then
     exit 1
 fi
 
-echo "Cloning srt-slurm repository..."
 SRT_REPO_DIR="srt-slurm"
 SRTCTL_SETUP_SCRIPT=""
 if uses_watchtower_shared_fs; then
@@ -345,95 +345,12 @@ if uses_watchtower_shared_fs; then
     RUN_KEY="${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}-${RUNNER_NAME}-$$"
     SRT_REPO_DIR="${SHARED_BASE}/srt-slurm-${RUN_KEY}"
 fi
-if [ -d "$SRT_REPO_DIR" ]; then
-    echo "Removing existing $SRT_REPO_DIR..."
-    rm -rf "$SRT_REPO_DIR"
-fi
 
-# TODO(CJQ): make first class upon srt-slurm upstream refactor
-if [[ "$IS_AGENTIC" == "1" ]]; then
-    # Agentic multi-node uses the same pinned cquil11/srt-slurm-nv commit as
-    # launch_gb300-nv.sh — everything the agentic recipes need is there:
-    #   - BenchmarkType.CUSTOM + benchmark.command + benchmark.env
-    #     (the hook that hands off to benchmarks/multi_node/agentic_srt.sh)
-    #   - DynamoConfig.wheel (recipes pin the ai-dynamo wheel)
-    #   - srtctl apply --no-preflight (model path /mnt/numa1 is compute-node
-    #     local NVMe, invisible to the login-node runner)
-    #   - benchmark_stage srun_options propagation (container-remap-root
-    #     must reach the agentic_srt.sh srun)
-    git clone https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout de59739b172e507e15ebf145bfe305f606e82fbf
-    mkdir -p recipes/vllm/deepseek-v4/agentic
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
-        recipes/vllm/deepseek-v4/agentic
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout aflowers/vllm-gb200-v0.20.0
-    # Use `cp -rT` so if the upstream branch ever ships a stub
-    # `recipes/vllm/deepseek-v4/` directory, we overlay our recipes onto
-    # it rather than nesting (`cp -r src dst` would create
-    # `recipes/vllm/deepseek-v4/deepseek-v4/...` in that case).
-    mkdir -p recipes/vllm/deepseek-v4
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
-    # Stay on NVIDIA/srt-slurm:main (default) — submission branch no
-    # longer needed; overlay our hand-rolled DSV4 sglang recipes onto it.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    mkdir -p recipes/sglang/deepseek-v4
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.1" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-    mkdir -p recipes/sglang/glm5
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5" recipes/sglang/glm5
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "qwen3.5" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    mkdir -p recipes/sglang/qwen3.5
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5" recipes/sglang/qwen3.5
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.1" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    mkdir -p recipes/sglang/glm5
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5" recipes/sglang/glm5
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout sa-submission-q2-2026 || exit 1
-    mkdir -p recipes/vllm/minimax-m3-gb200-fp8 || exit 1
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb200-fp8" recipes/vllm/minimax-m3-gb200-fp8 || exit 1
+if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
     SRTCTL_SETUP_SCRIPT="minimax-m3-gb200-vllm-fixes.sh"
-    cp \
-        "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
-        "configs/$SRTCTL_SETUP_SCRIPT" || exit 1
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
-    cd "$SRT_REPO_DIR" || exit 1
-    git checkout main || exit 1
-    mkdir -p recipes/vllm/kimi-k2.5-fp4 || exit 1
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4" recipes/vllm/kimi-k2.5-fp4 || exit 1
-elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "kimik2.5" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "glm5" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout v1.0.26
-    mkdir -p recipes/trtllm/glm5
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/trtllm/glm5" recipes/trtllm/glm5
-else
-    git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
 fi
+prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+cd "$SRT_REPO_DIR" || exit 1
 
 echo "Installing srtctl..."
 curl -LsSf https://astral.sh/uv/install.sh | sh
@@ -578,11 +495,11 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
     PREFLIGHT_ARGS=(--no-preflight)
 fi
 
+CONFIG_FILE="$(
+    prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+)" || exit 1
 SRTCTL_APPLY_ARGS=(
     "${PREFLIGHT_ARGS[@]}"
-    # Pass the full CONFIG_FILE (not the stripped CONFIG_PATH): srtctl needs the
-    # ":zip_override_...[i]" selector to pick the recipe block. For plain-file
-    # recipes CONFIG_FILE == CONFIG_PATH, so this is a no-op for them.
     -f "$CONFIG_FILE"
     --tags "gb200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
 )
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index e5a8d059b2..9798df5857 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -4,6 +4,8 @@
 
 set -exo pipefail
 
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
+
 export SLURM_PARTITION="batch_1"
 export SLURM_ACCOUNT="benchmark"
 export ENROOT_ROOTFS_WRITABLE=1
@@ -132,108 +134,19 @@ export EVAL_ONLY="${EVAL_ONLY:-false}"
 export ISL="$ISL"
 export OSL="$OSL"
 
-echo "Cloning srt-slurm repository..."
 RUN_KEY=$(printf "%s" "${RESULT_FILENAME:-${RUNNER_NAME:-gb300-nv}}" | sha1sum | cut -c1-12)
 SRT_REPO_DIR="${GITHUB_WORKSPACE}/srt-slurm-${GITHUB_RUN_ID:-manual}-${GITHUB_RUN_ATTEMPT:-0}-${RUN_KEY}"
 SRTCTL_SETUP_SCRIPT=""
-rm -rf "$SRT_REPO_DIR"
-
-if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
-    # DSv4 GB300 sglang agentic: NVIDIA/srt-slurm v1.0.10 has the nginx
-    # client_max_body_size fix (>1 MiB agentic warmup bodies), the
-    # session-affinity frontend, and the BenchmarkType.CUSTOM / extra_mount
-    # schema these recipes need.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout v1.0.10
-    mkdir -p recipes/sglang/deepseek-v4/agentic
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic" \
-        recipes/sglang/deepseek-v4/agentic
-elif [[ "$IS_AGENTIC" == "1" ]]; then
-    # Agentic recipes use NVIDIA/srt-slurm v1.0.36. This is the upstream
-    # version validated in InferenceX PR #2302 and includes per-node DP,
-    # matching Dynamo health counts, multi-node TP port handling, and
-    # Mooncake compatibility. Keep it pinned so sweeps are reproducible.
-    git clone --branch v1.0.36 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
-    cd "$SRT_REPO_DIR" || exit 1
-
-    mkdir -p recipes/vllm/deepseek-v4/agentic || exit 1
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
-        recipes/vllm/deepseek-v4/agentic || exit 1
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout aflowers/gb200-dsv4-recipes
-    mkdir -p recipes/vllm/deepseek-v4
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout main
-    if [[ $PRECISION == "fp4" ]]; then
-        mkdir -p recipes/sglang/glm5/gb300-fp4
-        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5/gb300-fp4" recipes/sglang/glm5/gb300-fp4
-    fi
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.1" && $PRECISION == "fp8" ]]; then
-    # GLM-5.1 FP8 (gb300) recipes are version-controlled in-repo; overlay them
-    # onto the pinned submission branch.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-    mkdir -p recipes/sglang/glm5.1
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.1" recipes/sglang/glm5.1
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.1" ]]; then
-    # GLM-5.1 MTP recipe (recipes/gb300-fp4/glm5-mtp.yaml) lives on
-    # NVIDIA/srt-slurm:main — check it out; no in-repo overlay needed.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout main
-elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "qwen3.5" ]]; then
-    # Overlay our version-controlled Qwen3.5 recipes onto the srt-slurm checkout.
-    # fp8 recipes pin dynamo by commit hash (source install), which needs the
-    # cargo/maturin bootstrap included in the srt-slurm v1.0.25 release — the
-    # sa-submission-q2-2026 sglang install path assumes maturin ships in the
-    # image, and the lmsysorg/sglang nightly-dev-cu13 image doesn't include it.
-    # Same branch the identical gb200-fp8 recipes run on. fp4 recipes pin
-    # dynamo by version (pip install) and stay on the submission branch they
-    # were validated against.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    if [[ $PRECISION == "fp8" ]]; then
-        git checkout v1.0.25
-    else
-        git checkout sa-submission-q2-2026
-    fi
-    mkdir -p recipes/sglang/qwen3.5
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5" recipes/sglang/qwen3.5
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-    mkdir -p recipes/vllm/minimax-m3-gb300-fp8
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/minimax-m3-gb300-fp8" recipes/vllm/minimax-m3-gb300-fp8
+if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "minimaxm3" ]]; then
     SRTCTL_SETUP_SCRIPT="minimax-m3-gb300-vllm-fixes.sh"
-    cp \
-        "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/$SRTCTL_SETUP_SCRIPT" \
-        "configs/$SRTCTL_SETUP_SCRIPT"
-elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout main
-    mkdir -p recipes/vllm/kimi-k2.5-fp4
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.5-fp4" recipes/vllm/kimi-k2.5-fp4
-elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "dsv4" ]]; then
+fi
+if [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "dsv4" ]]; then
     # DSv4 dynamo-trt recipes use the HuggingFace model ID as model.path,
     # so override SRT_SLURM_MODEL_PREFIX to match the recipe's model path key.
     SRT_SLURM_MODEL_PREFIX="deepseek-ai/DeepSeek-V4-Pro"
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
-else
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout sa-submission-q2-2026
 fi
+prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+cd "$SRT_REPO_DIR" || exit 1
 
 echo "Installing srtctl..."
 export UV_INSTALL_DIR="$GITHUB_WORKSPACE/.local/bin"
@@ -330,12 +243,13 @@ sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_PATH"
 #   - the agentic path (DSv4-Pro checkpoint),
 #   - glm5.1, whose GLM-5.1-NVFP4 weights are prestaged on the compute-node
 #     /scratch/models, and
-#   - qwen3.5 fp8, whose weights are also on the compute-node /scratch/models
-#     and which runs on srt-slurm:v1.0.25 (the release that has the preflight;
-#     qwen3.5 fp4 runs on sa-submission-q2-2026, which has none).
+#   - qwen3.5 fp8, whose weights are also on the compute-node /scratch/models.
 # The engine still fails loudly at runtime if the path is genuinely missing on
 # the compute node. Other fixed-seq-len recipes resolve model.path to a
 # login-visible location, so keep the precheck enforced for them.
+CONFIG_FILE="$(
+    prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+)" || exit 1
 SRTCTL_APPLY_ARGS=(
     -f "$CONFIG_FILE"
     --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
diff --git a/runners/launch_h100-dgxc-slurm.sh b/runners/launch_h100-dgxc-slurm.sh
index 1334c95542..42642ab444 100644
--- a/runners/launch_h100-dgxc-slurm.sh
+++ b/runners/launch_h100-dgxc-slurm.sh
@@ -11,6 +11,8 @@ SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
 
 set -x
 
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
+
 if [[ "$IS_MULTINODE" == "true" ]]; then
 
     # MODEL_PATH: Override with pre-downloaded paths on H100 runner
@@ -38,22 +40,9 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         exit 1
     fi
 
-    echo "Cloning srt-slurm repository..."
     SRT_REPO_DIR="srt-slurm"
-    if [ -d "$SRT_REPO_DIR" ]; then
-        echo "Removing existing $SRT_REPO_DIR..."
-        rm -rf "$SRT_REPO_DIR"
-    fi
-
-    # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" ]]; then
-        git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR"
-    else
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR"
-        git checkout sa-submission-q2-2026
-    fi
+    prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+    cd "$SRT_REPO_DIR" || exit 1
 
     echo "Installing srtctl..."
     export UV_INSTALL_DIR="/mnt/nfs/sa-shared/.uv/bin"
@@ -145,6 +134,9 @@ EOF
     sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
     # Raise sglang's torch-distributed TCPStore timeout from the 600s gloo default
     sed -i '/^      watchdog-timeout:/a\      dist-timeout: 1800' "${CONFIG_FILE%%:*}"
+    CONFIG_FILE="$(
+        prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+    )" || exit 1
     SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "h100,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
     echo "$SRTCTL_OUTPUT"
 
diff --git a/runners/launch_h200-dgxc-slurm.sh b/runners/launch_h200-dgxc-slurm.sh
index 00c6cc4977..28cdd21f73 100755
--- a/runners/launch_h200-dgxc-slurm.sh
+++ b/runners/launch_h200-dgxc-slurm.sh
@@ -7,6 +7,8 @@ SLURM_ACCOUNT="sa-shared"
 
 set -x
 
+source "$(dirname "${BASH_SOURCE[0]}")/srt_slurm.sh" || exit 1
+
 if [[ "$IS_MULTINODE" == "true" ]]; then
 
     # MODEL_PATH: Override with pre-downloaded paths on H200 runner
@@ -34,22 +36,9 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         exit 1
     fi
 
-    echo "Cloning srt-slurm repository..."
     SRT_REPO_DIR="srt-slurm"
-    if [ -d "$SRT_REPO_DIR" ]; then
-        echo "Removing existing $SRT_REPO_DIR..."
-        rm -rf "$SRT_REPO_DIR"
-    fi
-
-    # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" ]]; then
-        git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR"
-    else
-        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-        cd "$SRT_REPO_DIR"
-        git checkout sa-submission-q2-2026
-    fi
+    prepare_srt_slurm_checkout "$SRT_REPO_DIR" || exit 1
+    cd "$SRT_REPO_DIR" || exit 1
 
     echo "Installing srtctl..."
     curl -LsSf https://astral.sh/uv/install.sh | sh
@@ -135,6 +124,9 @@ EOF
     sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
     sed -i '/^health_check:/,/^[^ ]/{ /^health_check:/d; /^  /d; }' "${CONFIG_FILE%%:*}"
     printf '\nhealth_check:\n  max_attempts: 720\n  interval_seconds: 10\n' >> "${CONFIG_FILE%%:*}"
+    CONFIG_FILE="$(
+        prepare_inferencex_srt_benchmark_config "$CONFIG_FILE"
+    )" || exit 1
     SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "h200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
     echo "$SRTCTL_OUTPUT"
 
diff --git a/runners/srt_slurm.sh b/runners/srt_slurm.sh
new file mode 100644
index 0000000000..8eb2bb55ea
--- /dev/null
+++ b/runners/srt_slurm.sh
@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+
+# Shared srt-slurm checkout setup for multinode launchers.
+
+SRT_SLURM_INFERENCEX_ROOT="$(
+    cd "$(dirname "${BASH_SOURCE[0]}")/.." >/dev/null 2>&1 || exit 1
+    pwd
+)"
+SRT_SLURM_VERSION_FILE="${SRT_SLURM_INFERENCEX_ROOT}/configs/srt-slurm-version.txt"
+SRT_SLURM_REPOSITORY="https://github.com/NVIDIA/srt-slurm.git"
+
+if [[ ! -r "$SRT_SLURM_VERSION_FILE" ]]; then
+    echo "Unable to read the srt-slurm version from $SRT_SLURM_VERSION_FILE" >&2
+    return 1
+fi
+
+read -r SRT_SLURM_VERSION < "$SRT_SLURM_VERSION_FILE"
+if [[ -z "$SRT_SLURM_VERSION" ]]; then
+    echo "The srt-slurm version file is empty: $SRT_SLURM_VERSION_FILE" >&2
+    return 1
+fi
+
+prepare_srt_slurm_checkout() {
+    local destination=$1
+    local overlay_root="${SRT_SLURM_INFERENCEX_ROOT}/benchmarks/multi_node/srt-slurm-recipes"
+    local source_dir
+    local source_name
+    local target_dir
+
+    if [[ -z "$destination" || "$destination" == "/" ]]; then
+        echo "Refusing to prepare an invalid srt-slurm destination: '$destination'" >&2
+        return 1
+    fi
+
+    echo "Cloning srt-slurm ${SRT_SLURM_VERSION}..."
+    rm -rf "$destination"
+    git clone \
+        --branch "$SRT_SLURM_VERSION" \
+        --single-branch \
+        "$SRT_SLURM_REPOSITORY" \
+        "$destination" || return 1
+
+    # Overlay all InferenceX-owned recipes and setup scripts. Keeping these in
+    # InferenceX lets every launcher use the same tagged srt-slurm code while
+    # preserving configurations that have not been released upstream.
+    for source_dir in "$overlay_root"/*; do
+        [[ -d "$source_dir" ]] || continue
+        source_name=${source_dir##*/}
+        if [[ "$source_name" == "configs" ]]; then
+            target_dir="$destination/configs"
+        else
+            target_dir="$destination/recipes/$source_name"
+        fi
+        mkdir -p "$target_dir"
+        cp -R "$source_dir/." "$target_dir/" || return 1
+    done
+
+    # The B200 TRT-LLM Kimi K2.5 configs predate the upstream recipe naming
+    # convention used by CONFIG_FILE, so expose the checked-in recipes at the
+    # expected path without maintaining a second copy.
+    source_dir="$overlay_root/trtllm/kimi-k2.5/disagg/trtllm_dynamo/b200-fp4"
+    if [[ -d "$source_dir" ]]; then
+        target_dir="$destination/recipes/trtllm/kimi-k25-nvfp4/b200-fp4"
+        mkdir -p "$target_dir"
+        cp -R "$source_dir/." "$target_dir/" || return 1
+    fi
+}
+
+prepare_inferencex_srt_benchmark_config() {
+    local config_spec=$1
+
+    python \
+        "${SRT_SLURM_INFERENCEX_ROOT}/runners/srt_slurm_benchmark_config.py" \
+        "$config_spec"
+}
diff --git a/runners/srt_slurm_benchmark_config.py b/runners/srt_slurm_benchmark_config.py
new file mode 100755
index 0000000000..06d0f52856
--- /dev/null
+++ b/runners/srt_slurm_benchmark_config.py
@@ -0,0 +1,252 @@
+#!/usr/bin/env python3
+"""Hydrate InferenceX srt-slurm custom benchmark recipes."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import shlex
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+from srtctl.core.config import (
+    load_cluster_config,
+    resolve_config_with_defaults,
+    resolve_override_yaml,
+)
+from srtctl.core.schema import SrtConfig
+
+
+CUSTOM_BENCHMARK_ENTRYPOINT = (
+    "/infmax-workspace/benchmarks/multi_node/"
+    "srt_bench_serving_entrypoint.sh"
+)
+CUSTOM_BENCHMARK_COMMAND = f"bash {CUSTOM_BENCHMARK_ENTRYPOINT}"
+
+
+def split_config_spec(config_spec: str) -> tuple[Path, str | None]:
+    """Split srtctl's ``path.yaml:selector`` syntax."""
+    marker = ".yaml:"
+    if marker not in config_spec:
+        return Path(config_spec), None
+    path, selector = config_spec.split(marker, 1)
+    return Path(f"{path}.yaml"), selector
+
+
+def to_plain_data(value: Any) -> Any:
+    """Remove ruamel container and scalar types before PyYAML serialization."""
+    if isinstance(value, dict):
+        return {key: to_plain_data(item) for key, item in value.items()}
+    if isinstance(value, list):
+        return [to_plain_data(item) for item in value]
+    if isinstance(value, str):
+        return str(value)
+    if isinstance(value, bool):
+        return bool(value)
+    if isinstance(value, int):
+        return int(value)
+    if isinstance(value, float):
+        return float(value)
+    return value
+
+
+def load_selected_recipes(
+    config_spec: str,
+) -> tuple[Path, list[tuple[str | None, dict[str, Any]]], bool]:
+    """Load a plain recipe or resolve the requested override variants."""
+    config_path, selector = split_config_spec(config_spec)
+    with config_path.open(encoding="utf-8") as config_file:
+        raw_config = yaml.safe_load(config_file)
+
+    if not isinstance(raw_config, dict):
+        raise ValueError(f"{config_path} is not a YAML mapping")
+
+    if "base" not in raw_config:
+        if selector is not None:
+            raise ValueError(
+                f"{config_path} is not an override recipe, but selector "
+                f"{selector!r} was provided"
+            )
+        return config_path, [(None, raw_config)], False
+
+    if selector is None:
+        variants = [
+            (suffix, to_plain_data(recipe))
+            for suffix, recipe in resolve_override_yaml(config_path)
+        ]
+        return config_path, variants, True
+
+    variants = resolve_override_yaml(config_path, selector)
+    if not variants:
+        raise ValueError(
+            f"{config_spec} did not resolve to any variants"
+        )
+    return (
+        config_path,
+        [(None, to_plain_data(recipe)) for _, recipe in variants],
+        len(variants) > 1,
+    )
+
+
+def stringify_concurrencies(concurrencies: list[int] | str | None) -> str:
+    """Return the entrypoint's whitespace-separated concurrency format."""
+    if isinstance(concurrencies, list):
+        return " ".join(str(value) for value in concurrencies)
+    if concurrencies is None:
+        return ""
+    return str(concurrencies).replace("x", " ")
+
+
+def custom_benchmark_arguments(config: SrtConfig) -> list[str]:
+    """Translate recipe settings to entrypoint arguments."""
+    benchmark = config.benchmark
+    resources = config.resources
+
+    if benchmark.dataset_name not in (None, "random"):
+        raise ValueError(
+            "InferenceX benchmark_serving currently supports only the random "
+            f"dataset, not {benchmark.dataset_name!r}"
+        )
+    if benchmark.reuse_http_connections:
+        raise ValueError(
+            "InferenceX benchmark_serving does not support "
+            "benchmark.reuse_http_connections"
+        )
+    if (
+        benchmark.slow_down_sleep_time is not None
+        or benchmark.slow_down_wait_time is not None
+    ):
+        raise ValueError(
+            "InferenceX benchmark_serving does not support SA-Bench slow_down"
+        )
+
+    concurrencies = stringify_concurrencies(benchmark.concurrencies)
+    if not concurrencies:
+        raise ValueError("benchmark.concurrencies is required")
+    if benchmark.isl is None or benchmark.osl is None:
+        raise ValueError("benchmark.isl and benchmark.osl are required")
+
+    if resources.is_disaggregated:
+        prefill_gpus = resources.prefill_gpus
+        decode_gpus = resources.decode_gpus
+        total_gpus = prefill_gpus + decode_gpus
+    else:
+        prefill_gpus = 0
+        decode_gpus = 0
+        total_gpus = (resources.agg_nodes or 1) * resources.gpus_per_node
+
+    model_path = config.model.path
+    tokenizer = (
+        model_path.removeprefix("hf:")
+        if model_path.startswith("hf:")
+        else "/model"
+    )
+    custom_tokenizer = benchmark.custom_tokenizer or ""
+    is_deepseek_v4 = "deepseek_v4" in custom_tokenizer.lower()
+
+    return [
+        config.served_model_name,
+        tokenizer,
+        str(benchmark.isl),
+        str(benchmark.osl),
+        concurrencies,
+        str(benchmark.req_rate or "inf"),
+        str(resources.is_disaggregated).lower(),
+        str(prefill_gpus),
+        str(decode_gpus),
+        str(total_gpus),
+        str(benchmark.random_range_ratio or 0.8),
+        str(benchmark.num_prompts_mult or 10),
+        str(
+            benchmark.num_warmup_mult
+            if benchmark.num_warmup_mult is not None
+            else 2
+        ),
+        str(benchmark.use_chat_template).lower(),
+        str(
+            is_deepseek_v4 and benchmark.use_chat_template
+        ).lower(),
+    ]
+
+
+def hydrate_recipe(recipe: dict[str, Any]) -> bool:
+    """Hydrate an InferenceX throughput recipe; return whether it changed."""
+    benchmark = recipe.get("benchmark")
+    if not isinstance(benchmark, dict):
+        return False
+
+    benchmark_type = benchmark.get("type")
+    if benchmark_type == "sa-bench":
+        benchmark["type"] = "custom"
+        benchmark["command"] = CUSTOM_BENCHMARK_COMMAND
+    elif not (
+        benchmark_type == "custom"
+        and benchmark.get("command") == CUSTOM_BENCHMARK_COMMAND
+    ):
+        return False
+
+    cluster_config = load_cluster_config()
+    resolved = resolve_config_with_defaults(recipe, cluster_config)
+    config = SrtConfig.Schema().load(resolved)
+    if not isinstance(config, SrtConfig):
+        raise TypeError("srt-slurm did not return a typed SrtConfig")
+
+    benchmark["command"] = shlex.join(
+        [
+            "bash",
+            CUSTOM_BENCHMARK_ENTRYPOINT,
+            *custom_benchmark_arguments(config),
+        ]
+    )
+    benchmark.pop("env", None)
+    return True
+
+
+def render_config(config_spec: str, output_dir: Path) -> str:
+    """Render selected recipes and return their path for ``srtctl apply``."""
+    config_path, variants, keep_override_format = load_selected_recipes(
+        config_spec
+    )
+    changed = False
+    for _, recipe in variants:
+        changed = hydrate_recipe(recipe) or changed
+    if not changed:
+        return config_spec
+
+    if keep_override_format:
+        rendered: dict[str, Any] = {"base": {}}
+        for index, (suffix, recipe) in enumerate(variants):
+            variant_name = suffix or f"inferencex_{index}"
+            rendered[f"override_{variant_name}"] = recipe
+    else:
+        if len(variants) != 1:
+            raise ValueError(
+                f"{config_spec} resolved to {len(variants)} variants but "
+                "cannot preserve their selector"
+            )
+        rendered = variants[0][1]
+
+    digest = hashlib.sha256(config_spec.encode()).hexdigest()[:12]
+    output_dir.mkdir(parents=True, exist_ok=True)
+    output_path = output_dir / f"{config_path.stem}-{digest}.yaml"
+    with output_path.open("w", encoding="utf-8") as output_file:
+        yaml.safe_dump(rendered, output_file, sort_keys=False)
+    return str(output_path)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("config_spec")
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=Path(".inferencex-recipes"),
+    )
+    args = parser.parse_args()
+    print(render_config(args.config_spec, args.output_dir))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/utils/bench_serving/backend_request_func.py b/utils/bench_serving/backend_request_func.py
index 4c8820f8da..662147ada3 100644
--- a/utils/bench_serving/backend_request_func.py
+++ b/utils/bench_serving/backend_request_func.py
@@ -553,11 +553,20 @@ def get_tokenizer(
         return MistralTokenizer.from_pretrained(
             str(pretrained_model_name_or_path))
     else:
-        tokenizer = AutoTokenizer.from_pretrained(
-            pretrained_model_name_or_path,
-            trust_remote_code=trust_remote_code,
-            **kwargs,
-        )
+        try:
+            tokenizer = AutoTokenizer.from_pretrained(
+                pretrained_model_name_or_path,
+                trust_remote_code=trust_remote_code,
+                **kwargs,
+            )
+        except ValueError as error:
+            if "deepseek_v4" not in str(error).lower():
+                raise
+            tokenizer = PreTrainedTokenizerFast.from_pretrained(
+                pretrained_model_name_or_path,
+                trust_remote_code=trust_remote_code,
+                **kwargs,
+            )
         return _fix_tokenizer_for_sglang(tokenizer, pretrained_model_name_or_path)