diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml
new file mode 100644
index 0000000000..9a382b0dae
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml
@@ -0,0 +1,131 @@
+name: "svf-vllm-disagg-b200-1p1d-dep8-dep8-mtp"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+health_check:
+  max_attempts: 90
+  interval_seconds: 10
+sbatch_directives:
+  segment: "1"
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NVSHMEM_INFO: "1"
+    NVSHMEM_DEBUG: "INFO"
+    NCCL_NVLS_ENABLE: "1"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NCCL_NVLS_ENABLE: "1"
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enforce-eager: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: false
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "512"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml
new file mode 100644
index 0000000000..f5096f712a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml
@@ -0,0 +1,116 @@
+name: "svf-vllm-disagg-b200-1p1d-dep8-tp8-mtp"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+sbatch_directives:
+  segment: "1"
+health_check:
+  max_attempts: 90
+  interval_seconds: 10
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NVSHMEM_INFO: "1"
+    NVSHMEM_DEBUG: "INFO"
+    NCCL_NVLS_ENABLE: "1"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enforce-eager: true
+      max-model-len: 9280
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      # numa-bind: true
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      max-model-len: 16384
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      max-num-seqs: 256
+      max-cudagraph-capture-size: 64
+      max-num-batched-tokens: 256
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x8x16"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml
new file mode 100644
index 0000000000..219fb9c491
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml
@@ -0,0 +1,132 @@
+name: "svf-vllm-disagg-b200-1p2d-dep8-dep8-mtp"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+sbatch_directives:
+  segment: "1"
+health_check:
+  max_attempts: 90
+  interval_seconds: 10
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 2
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NVSHMEM_INFO: "1"
+    NVSHMEM_DEBUG: "INFO"
+    NCCL_NVLS_ENABLE: "1"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NCCL_NVLS_ENABLE: "1"
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enforce-eager: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      # numa-bind: true
+      tokenizer-mode: deepseek_v4
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 128
+      max-cudagraph-capture-size: 128
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "128x256"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml
new file mode 100644
index 0000000000..0e96fff2d5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml
@@ -0,0 +1,131 @@
+name: "svf-vllm-disagg-b200-2p1d-dep8-dep8-mtp"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+sbatch_directives:
+  segment: "1"
+health_check:
+  max_attempts: 90
+  interval_seconds: 10
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NVSHMEM_INFO: "1"
+    NVSHMEM_DEBUG: "INFO"
+    NCCL_NVLS_ENABLE: "1"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NCCL_NVLS_ENABLE: "1"
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enforce-eager: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      # numa-bind: true
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1024"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml
new file mode 100644
index 0000000000..b07bbed4c9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml
@@ -0,0 +1,131 @@
+name: "svf-vllm-disagg-b200-3p1d-dep8-dep8-mtp"
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+health_check:
+  max_attempts: 90
+  interval_seconds: 10
+sbatch_directives:
+  segment: "1"
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 3
+  decode_nodes: 1
+  prefill_workers: 3
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    VLLM_DSV4_MEGA_FP8_COMBINE: "1"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NVSHMEM_INFO: "1"
+    NVSHMEM_DEBUG: "INFO"
+    NCCL_NVLS_ENABLE: "1"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_NET_DEVICES: mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_10:1,mlx5_11:1
+    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_DISABLE_CUDA_VMM: "1"
+    NVSHMEM_SYMMETRIC_SIZE: "16g"
+    NCCL_NVLS_ENABLE: "1"
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      enforce-eager: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":1}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      # numa-bind: true
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      data-parallel-hybrid-lb: true
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      speculative-config: '{"method":"mtp","num_speculative_tokens":1}'
+      attention-config: '{"use_fp4_indexer_cache":true}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      max-model-len: 9280
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a30af59641..78260b442e 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6147,6 +6147,107 @@ dsv4-fp4-gb300-dynamo-vllm:
           ep: 16
           dp-attn: true
 
+dsv4-fp4-b200-dynamo-vllm-mtp:
+  image: vllm/vllm-openai:vllm-x86_64-cu13-0.25.1-7a33ba9
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b200-new
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  kv-p2p-transfer: nixl
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # B200 MTP disagg recipes. Each worker maps to one full 8-GPU B200 node
+      # (prefill dep8; decode tp8, except the low-latency tp8 variant).
+      # SYNTHETIC_ACCEPTANCE is a throughput-only control (injected into the
+      # recipe by runners/inject_synthetic_acceptance.py); EVAL_ONLY runs keep
+      # the checked-in real-MTP recipe so verification is measured normally.
+      - spec-decoding: "mtp"
+        conc-list: [1, 8, 16]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [128, 256]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml"
+        decode:
+          num-worker: 2
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - spec-decoding: "mtp"
+        conc-list: [512]
+        prefill:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - spec-decoding: "mtp"
+        conc-list: [1024]
+        prefill:
+          num-worker: 2
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+      - spec-decoding: "mtp"
+        conc-list: [4096]
+        prefill:
+          num-worker: 3
+          tp: 1
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=1.79"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml"
+        decode:
+          num-worker: 1
+          tp: 1
+          ep: 8
+          dp-attn: true
+
 dsv4-fp4-gb300-dynamo-trt:
   image: nvcr.io#nvidia/ai-dynamo/tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6eebd97070..95a4f5a8aa 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5355,3 +5355,9 @@
     - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path"
     - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213
+
+- config-keys:
+    - dsv4-fp4-b200-dynamo-vllm-mtp
+  description:
+    - "Add DeepSeek-V4-Pro FP4 B200 disaggregated Dynamo vLLM MTP benchmarks for 8k/1k"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2390
diff --git a/runners/inject_synthetic_acceptance.py b/runners/inject_synthetic_acceptance.py
index 4bda75a9a4..80d2740e79 100644
--- a/runners/inject_synthetic_acceptance.py
+++ b/runners/inject_synthetic_acceptance.py
@@ -47,6 +47,11 @@ def _log(msg):
     print(f"[Synthetic AR] {msg}")
 
 
+def _enabled(name: str) -> bool:
+    """Return whether an environment flag is explicitly enabled."""
+    return os.environ.get(name, "false").strip().lower() == "true"
+
+
 def _yaml_key(model_prefix):
     return MODEL_PREFIX_TO_YAML_KEY.get(model_prefix, model_prefix)
 
@@ -120,7 +125,11 @@ def _resolve_al(config_text, injector, ref_yaml):
 
 
 def inject(config_file, framework):
-    if os.environ.get("SYNTHETIC_ACCEPTANCE", "false") != "true":
+    if _enabled("EVAL_ONLY"):
+        print("[Synthetic AL] EVAL_ONLY=true: keeping real MTP recipe")
+        return 0
+
+    if not _enabled("SYNTHETIC_ACCEPTANCE"):
         return 0
 
     injector = get_injector(framework)
@@ -144,8 +153,10 @@ def inject(config_file, framework):
     new_content, count = injector.rewrite(content, al, _log)
 
     if count == 0:
-        _log("WARNING: No speculative-config entries found to modify; leaving recipe unchanged")
-        return 0
+        sys.exit(
+            "ERROR: SYNTHETIC_ACCEPTANCE=true but no speculative-config "
+            f"entries were found in {config_file}"
+        )
 
     with open(config_file, "w") as f:
         f.write(new_content)
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index be68ce073b..ed24a61d13 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -300,12 +300,21 @@ EOF
         exit 1
     fi
 
+    # Strip any :override[N] selector so sed and the injector operate on the file.
+    CONFIG_PATH="${CONFIG_FILE%%:*}"
+
     # Override the job name in the config file with the runner name
-    sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "${CONFIG_FILE%%:*}"
+    sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_PATH"
     # Bump recipe health-check timeout from 360×10s=3600s to 720×10s=7200s
     # 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%%:*}"
+    sed -i 's/^  max_attempts: [0-9]*/  max_attempts: 720/' "$CONFIG_PATH"
+
+    # Optionally inject synthetic acceptance into the recipe's speculative-config
+    # when SYNTHETIC_ACCEPTANCE=true (no-op otherwise). Must run after the name
+    # override and before srtctl apply so the rendered job picks it up. Fail
+    # fast so a broken opt-in never reaches srtctl with an unrewritten recipe.
+    python3 "$GITHUB_WORKSPACE/runners/inject_synthetic_acceptance.py" \
+        "$CONFIG_PATH" "$FRAMEWORK" || exit 1
 
     SRTCTL_PREFLIGHT_ARGS=()
     # Kimi K2.6 weights are staged on the Slurm compute nodes, not the login node.
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
new file mode 120000
index 0000000000..f45b7edcb3
--- /dev/null
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -0,0 +1 @@
+launch_b200-dgxc.sh
\ No newline at end of file