From 245dcda3cc0ad6e2831659477c2eaf4b22b6e63c Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 13:09:07 -0500
Subject: [PATCH 01/21] feat(agentx): add Kimi K3 GB200 day-0 support
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add srt-slurm TP, TEP, and DEP AgentX profiles plus GB200 runner routing and recipe overlay plumbing.

中文：新增 Kimi K3 GB200 首日 AgentX 支持，包括 srt-slurm TP、TEP、DEP 配置，以及 GB200 运行器路由和配方覆盖逻辑。
---
 .../agg-gb200-dep16-throughput-agentic.yaml   | 122 ++++++++++++++++++
 .../agg-gb200-tep8-balanced-agentic.yaml      | 121 +++++++++++++++++
 .../agg-gb200-tp8-latency-agentic.yaml        | 119 +++++++++++++++++
 configs/nvidia-master.yaml                    |  73 +++++++++++
 perf-changelog.yaml                           |   9 ++
 runners/launch_gb200-nv.sh                    |  22 +++-
 6 files changed, 461 insertions(+), 5 deletions(-)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
new file mode 100644
index 0000000000..5a1721fdd8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -0,0 +1,122 @@
+name: "kimi-k3-vllm-agg-gb200-dep16-throughput-agentic"
+
+# Day-0 GB200 translation of the official throughput-oriented multi_node_dep
+# profile. The 16-way data/expert-parallel worker spans four four-GPU GB200
+# nodes, with four local DP ranks on each node.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
+
+model:
+  path: "kimi-k3"
+  container: "vllm/vllm-openai:kimi-k3"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:kimi-k3"
+  frameworks:
+    dynamo: "1.3.0.dev20260618"
+
+dynamo:
+  install: true
+  wheel: "1.3.0.dev20260618"
+
+environment:
+  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
+  ETCD_LEASE_TTL: "7200"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 4
+  agg_workers: 1
+  gpus_per_agg: 16
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: "kv"
+    router-reset-states: true
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
+    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-dep16-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 16
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      moe-backend: "deep_gemm_mega_moe"
+      all2all-backend: "flashinfer_nvlink_one_sided"
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      max-num-seqs: 128
+      max-num-batched-tokens: 16384
+      language-model-only: true
+      enable-auto-tool-choice: true
+      tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
new file mode 100644
index 0000000000..3993e3f6ac
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
@@ -0,0 +1,121 @@
+name: "kimi-k3-vllm-agg-gb200-tep8-balanced-agentic"
+
+# Day-0 GB200 translation of the official balanced multi_node_tep profile.
+# Dense layers use TP8 while the MoE experts are sharded across the same eight
+# GPUs on two four-GPU GB200 nodes.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tep
+
+model:
+  path: "kimi-k3"
+  container: "vllm/vllm-openai:kimi-k3"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:kimi-k3"
+  frameworks:
+    dynamo: "1.3.0.dev20260618"
+
+dynamo:
+  install: true
+  wheel: "1.3.0.dev20260618"
+
+environment:
+  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
+  ETCD_LEASE_TTL: "7200"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: "kv"
+    router-reset-states: true
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
+    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tep8-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      enable-expert-parallel: true
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      moe-backend: "flashinfer_trtllm"
+      all2all-backend: "flashinfer_nvlink_one_sided"
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      max-num-seqs: 32
+      max-num-batched-tokens: 8192
+      language-model-only: true
+      enable-auto-tool-choice: true
+      tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
new file mode 100644
index 0000000000..464c8fa529
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
@@ -0,0 +1,119 @@
+name: "kimi-k3-vllm-agg-gb200-tp8-latency-agentic"
+
+# Day-0 GB200 translation of the official latency-oriented multi_node_tp
+# profile. Kimi K3 requires at least eight GPUs for TP, so this aggregate
+# worker spans two four-GPU GB200 nodes.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tp
+
+model:
+  path: "kimi-k3"
+  container: "vllm/vllm-openai:kimi-k3"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:kimi-k3"
+  frameworks:
+    dynamo: "1.3.0.dev20260618"
+
+dynamo:
+  install: true
+  wheel: "1.3.0.dev20260618"
+
+environment:
+  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
+  ETCD_LEASE_TTL: "7200"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb200"
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: "kv"
+    router-reset-states: true
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
+    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
+    VLLM_ENGINE_READY_TIMEOUT_S: "7200"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    NCCL_P2P_LEVEL: "NVL"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tp8-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      trust-remote-code: true
+      load-format: "fastsafetensors"
+      moe-backend: "flashinfer_trtllm"
+      gpu-memory-utilization: 0.95
+      max-model-len: 1048576
+      max-num-seqs: 8
+      max-num-batched-tokens: 8192
+      language-model-only: true
+      enable-auto-tool-choice: true
+      tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  aiperf_server_metrics: true
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 36ae816809..fde8856a5b 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7986,6 +7986,79 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
           tp: 8
           ep: 8
           dp-attn: true
+
+# Day-0 Kimi K3 AgentX coverage on GB200. The three aggregate srt-slurm
+# profiles mirror the serving strategies synthesized by the official recipe:
+#   latency:    multi-node TP8  (2 GB200 nodes / 8 GPUs)
+#   balanced:   multi-node TEP8 (2 GB200 nodes / 8 GPUs)
+#   throughput: multi-node DEP16 (4 GB200 nodes / 16 GPUs)
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200
+kimik3-fp4-gb200-dynamo-vllm-agentic:
+  image: vllm/vllm-openai:kimi-k3
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.0.dev20260618" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      # Latency oriented: official multi_node_tp strategy, TP8 across two
+      # four-GPU GB200 nodes.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tp
+      - spec-decoding: none
+        conc-list: [1, 2, 4, 8]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
+      # Balanced: official multi_node_tep strategy, TEP8 across two GB200
+      # nodes. Expert parallelism reduces MoE weight/compute pressure while
+      # retaining TP8 for dense layers.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tep
+      - spec-decoding: none
+        conc-list: [8, 16, 24, 32]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # Throughput oriented: official multi_node_dep strategy, DEP16 across
+      # four GB200 nodes (DP16/EP16, four local DP ranks per node).
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
+      - spec-decoding: none
+        conc-list: [32, 64, 96, 128]
+        prefill:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 16
+          ep: 16
+          dp-attn: true
+
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
   image: lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0524d0c0c9..8b59d9a335 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5116,3 +5116,12 @@
   description:
     - "Bump image from lmsysorg/sglang:v0.5.14-rocm720-mi35x to lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260726"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2349
+
+- config-keys:
+    - kimik3-fp4-gb200-dynamo-vllm-agentic
+  description:
+    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX support through checked-in srt-slurm aggregate recipes and the official vllm/vllm-openai:kimi-k3 image."
+    - "Reference the official GB200 command-builder profiles at https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200: latency-oriented multi-node TP8 on 2 nodes, balanced multi-node TEP8 on 2 nodes, and throughput-oriented multi-node DEP16 on 4 nodes. The upstream recipe is pre-release and does not yet mark GB200 as verified."
+    - "Run the text-only AgentX trace with Kimi K3's full 1M context, kimi_k3 tool-call and reasoning parsers, fastsafetensors loading, K3 latent-MoE tail fusion, FlashInfer TRT-LLM MoE for TP/TEP, and DeepGEMM mega-MoE plus one-sided FlashInfer all-to-all for DEP."
+    - "Teach launch_gb200-nv.sh to map kimik3/fp4 to the pre-staged /mnt/lustre01/models/kimi-k3 checkpoint, register the matching srt-slurm model alias, and overlay the checked-in Kimi K3 AgentX recipes into the pinned agentic srt-slurm checkout."
+  pr-link: TBD
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 426cd2dd87..4cbc852f5c 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -155,6 +155,12 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
     if [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/kimi-k2.5-nvfp4"
         export SRT_SLURM_MODEL_PREFIX="kimi-k2.5-nvfp4"
+    elif [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
+        # Kimi K3's ~1.68 TB MXFP4 checkpoint must be pre-staged on the
+        # compute-visible shared Lustre filesystem. This alias matches
+        # model.path in the checked-in srt-slurm AgentX recipes.
+        export MODEL_PATH="/mnt/lustre01/models/kimi-k3"
+        export SRT_SLURM_MODEL_PREFIX="kimi-k3"
     elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
         # The FP4 checkpoint is staged on compute-visible Lustre. The former
         # /mnt/numa1 path is no longer present on watchtower compute nodes;
@@ -172,7 +178,7 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/MiniMax-M3-MXFP8"
         export SRT_SLURM_MODEL_PREFIX="minimax-m3-mxfp8"
     else
-        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
+        echo "Unsupported model prefix/precision combination: $MODEL_PREFIX/$PRECISION. Supported combinations for dynamo-vllm: kimik2.5/fp4, kimik3/fp4, dsv4/fp4, minimaxm2.5/fp4, minimaxm2.5/fp8, minimaxm3/fp8"
         exit 1
     fi
 else
@@ -183,7 +189,7 @@ NGINX_IMAGE="nginx:1.27.4"
 
 uses_watchtower_shared_fs() {
     case "$MODEL_PREFIX" in
-        minimaxm2.5|minimaxm3|kimik2.5|qwen3.5) return 0 ;;
+        minimaxm2.5|minimaxm3|kimik2.5|kimik3|qwen3.5) return 0 ;;
         *) return 1 ;;
     esac
 }
@@ -364,9 +370,15 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
     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 "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
-        recipes/vllm/deepseek-v4/agentic
+    if [[ "$MODEL_PREFIX" == "kimik3" ]]; then
+        mkdir -p recipes/vllm/kimi-k3/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
+            recipes/vllm/kimi-k3/agentic
+    else
+        mkdir -p recipes/vllm/deepseek-v4/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
+            recipes/vllm/deepseek-v4/agentic
+    fi
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"

From 3e2b96070b7049239f7e56aa9325a2b8d5106414 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 13:09:48 -0500
Subject: [PATCH 02/21] chore: link GB200 Kimi K3 changelog
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Record the draft pull request URL in the append-only benchmark changelog entry.

中文：在仅追加的基准测试变更日志条目中记录草稿 PR 链接。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 8b59d9a335..55d223058d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5124,4 +5124,4 @@
     - "Reference the official GB200 command-builder profiles at https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200: latency-oriented multi-node TP8 on 2 nodes, balanced multi-node TEP8 on 2 nodes, and throughput-oriented multi-node DEP16 on 4 nodes. The upstream recipe is pre-release and does not yet mark GB200 as verified."
     - "Run the text-only AgentX trace with Kimi K3's full 1M context, kimi_k3 tool-call and reasoning parsers, fastsafetensors loading, K3 latent-MoE tail fusion, FlashInfer TRT-LLM MoE for TP/TEP, and DeepGEMM mega-MoE plus one-sided FlashInfer all-to-all for DEP."
     - "Teach launch_gb200-nv.sh to map kimik3/fp4 to the pre-staged /mnt/lustre01/models/kimi-k3 checkpoint, register the matching srt-slurm model alias, and overlay the checked-in Kimi K3 AgentX recipes into the pinned agentic srt-slurm checkout."
-  pr-link: TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354

From 18b1af1deddf666683f7fe0e85d6b6bc7125bd2d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 18:36:21 -0500
Subject: [PATCH 03/21] fix(agentx): align Kimi K3 DEP metadata
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Represent the throughput arm as TP1 x DP16 = EP16 in the master configuration, matching the vLLM recipe.

中文：将吞吐量配置修正为 TP1 × DP16 = EP16，与 vLLM 配方保持一致。
---
 configs/nvidia-master.yaml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index fde8856a5b..082bc6e69e 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8042,20 +8042,20 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           ep: 8
           dp-attn: false
       # Throughput oriented: official multi_node_dep strategy, DEP16 across
-      # four GB200 nodes (DP16/EP16, four local DP ranks per node).
+      # four GB200 nodes (TP1 x DP16 = EP16, four local DP ranks per node).
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
       - spec-decoding: none
         conc-list: [32, 64, 96, 128]
         prefill:
           num-worker: 1
-          tp: 16
+          tp: 1
           ep: 16
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml"
         decode:
           num-worker: 0
-          tp: 16
+          tp: 1
           ep: 16
           dp-attn: true
 

From acfd9f3ba33591a3e3bdc6e198abaa5060441e5b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 18:42:49 -0500
Subject: [PATCH 04/21] perf(gb200): load Kimi K3 from local NVMe
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use the pre-staged /mnt/numa1/models/Kimi-K3 checkpoint on every GB200 compute node while retaining shared-filesystem orchestration and caches.

中文：在每个 GB200 计算节点上使用预置的 /mnt/numa1/models/Kimi-K3 检查点，同时保留共享文件系统上的编排与缓存。
---
 perf-changelog.yaml        | 2 +-
 runners/launch_gb200-nv.sh | 8 ++++----
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 55d223058d..99844d9f66 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5123,5 +5123,5 @@
     - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX support through checked-in srt-slurm aggregate recipes and the official vllm/vllm-openai:kimi-k3 image."
     - "Reference the official GB200 command-builder profiles at https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200: latency-oriented multi-node TP8 on 2 nodes, balanced multi-node TEP8 on 2 nodes, and throughput-oriented multi-node DEP16 on 4 nodes. The upstream recipe is pre-release and does not yet mark GB200 as verified."
     - "Run the text-only AgentX trace with Kimi K3's full 1M context, kimi_k3 tool-call and reasoning parsers, fastsafetensors loading, K3 latent-MoE tail fusion, FlashInfer TRT-LLM MoE for TP/TEP, and DeepGEMM mega-MoE plus one-sided FlashInfer all-to-all for DEP."
-    - "Teach launch_gb200-nv.sh to map kimik3/fp4 to the pre-staged /mnt/lustre01/models/kimi-k3 checkpoint, register the matching srt-slurm model alias, and overlay the checked-in Kimi K3 AgentX recipes into the pinned agentic srt-slurm checkout."
+    - "Teach launch_gb200-nv.sh to map kimik3/fp4 to the pre-staged node-local NVMe checkpoint at /mnt/numa1/models/Kimi-K3 on every GB200 node, register the matching srt-slurm model alias, and overlay the checked-in Kimi K3 AgentX recipes into the pinned agentic srt-slurm checkout."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 4cbc852f5c..a8c50c7d9a 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -156,10 +156,10 @@ elif [[ $FRAMEWORK == "dynamo-vllm" ]]; then
         export MODEL_PATH="/mnt/lustre01/models/kimi-k2.5-nvfp4"
         export SRT_SLURM_MODEL_PREFIX="kimi-k2.5-nvfp4"
     elif [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
-        # Kimi K3's ~1.68 TB MXFP4 checkpoint must be pre-staged on the
-        # compute-visible shared Lustre filesystem. This alias matches
-        # model.path in the checked-in srt-slurm AgentX recipes.
-        export MODEL_PATH="/mnt/lustre01/models/kimi-k3"
+        # Load Kimi K3 from node-local NVMe for faster startup. The checkpoint
+        # must be pre-staged at this exact path on every allocated GB200 node.
+        # This alias matches model.path in the checked-in AgentX recipes.
+        export MODEL_PATH="/mnt/numa1/models/Kimi-K3"
         export SRT_SLURM_MODEL_PREFIX="kimi-k3"
     elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
         # The FP4 checkpoint is staged on compute-visible Lustre. The former

From fd088cdeea0436c995c37a7eeb4584bb8a6d4128 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 18:45:52 -0500
Subject: [PATCH 05/21] chore: simplify Kimi K3 performance config
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Keep the Day-0 changelog focused on the vLLM recipe performance points and use the model default context length.

中文：精简 Kimi K3 首日性能配置，仅保留与 vLLM 配方一致的性能测试点，并使用模型默认上下文长度。
---
 .../kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml  | 1 -
 .../kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml     | 1 -
 .../vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml  | 1 -
 perf-changelog.yaml                                          | 5 +----
 4 files changed, 1 insertion(+), 7 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 5a1721fdd8..2607ea1572 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -90,7 +90,6 @@ backend:
       moe-backend: "deep_gemm_mega_moe"
       all2all-backend: "flashinfer_nvlink_one_sided"
       gpu-memory-utilization: 0.95
-      max-model-len: 1048576
       max-num-seqs: 128
       max-num-batched-tokens: 16384
       language-model-only: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
index 3993e3f6ac..11053ad3e0 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
@@ -88,7 +88,6 @@ backend:
       moe-backend: "flashinfer_trtllm"
       all2all-backend: "flashinfer_nvlink_one_sided"
       gpu-memory-utilization: 0.95
-      max-model-len: 1048576
       max-num-seqs: 32
       max-num-batched-tokens: 8192
       language-model-only: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
index 464c8fa529..3685e14653 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
@@ -86,7 +86,6 @@ backend:
       load-format: "fastsafetensors"
       moe-backend: "flashinfer_trtllm"
       gpu-memory-utilization: 0.95
-      max-model-len: 1048576
       max-num-seqs: 8
       max-num-batched-tokens: 8192
       language-model-only: true
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 99844d9f66..c94fee0cf9 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5120,8 +5120,5 @@
 - config-keys:
     - kimik3-fp4-gb200-dynamo-vllm-agentic
   description:
-    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX support through checked-in srt-slurm aggregate recipes and the official vllm/vllm-openai:kimi-k3 image."
-    - "Reference the official GB200 command-builder profiles at https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200: latency-oriented multi-node TP8 on 2 nodes, balanced multi-node TEP8 on 2 nodes, and throughput-oriented multi-node DEP16 on 4 nodes. The upstream recipe is pre-release and does not yet mark GB200 as verified."
-    - "Run the text-only AgentX trace with Kimi K3's full 1M context, kimi_k3 tool-call and reasoning parsers, fastsafetensors loading, K3 latent-MoE tail fusion, FlashInfer TRT-LLM MoE for TP/TEP, and DeepGEMM mega-MoE plus one-sided FlashInfer all-to-all for DEP."
-    - "Teach launch_gb200-nv.sh to map kimik3/fp4 to the pre-staged node-local NVMe checkpoint at /mnt/numa1/models/Kimi-K3 on every GB200 node, register the matching srt-slurm model alias, and overlay the checked-in Kimi K3 AgentX recipes into the pinned agentic srt-slurm checkout."
+    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP8 on 2 nodes (conc 1-8), TEP8 on 2 nodes (conc 8-32), and DEP16 on 4 nodes (conc 32-128). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354

From 6c335b500197f239905b7fecbe546a4fd0df6854 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 19:11:49 -0500
Subject: [PATCH 06/21] perf(gb200): size Kimi K3 AgentX sweep from HBM
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Move latency and balanced profiles to 16-GPU TP/TEP, use TP2 x DP8 for EP16 throughput, and extend the throughput concurrency sweep through 256.

中文：根据 GB200 HBM 容量调整 Kimi K3 AgentX 扫描。将延迟与均衡配置扩展为 16 GPU TP/TEP，吞吐配置改为 TP2 × DP8（EP16），并将吞吐并发扫描扩展至 256。
---
 .../agg-gb200-dep16-throughput-agentic.yaml   |  8 ++--
 ... => agg-gb200-tep16-balanced-agentic.yaml} | 16 +++----
 ...ml => agg-gb200-tp16-latency-agentic.yaml} | 16 +++----
 configs/nvidia-master.yaml                    | 42 +++++++++----------
 perf-changelog.yaml                           |  2 +-
 5 files changed, 41 insertions(+), 43 deletions(-)
 rename benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/{agg-gb200-tep8-balanced-agentic.yaml => agg-gb200-tep16-balanced-agentic.yaml} (89%)
 rename benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/{agg-gb200-tp8-latency-agentic.yaml => agg-gb200-tp16-latency-agentic.yaml} (89%)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 2607ea1572..d89336be4e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -1,8 +1,8 @@
 name: "kimi-k3-vllm-agg-gb200-dep16-throughput-agentic"
 
 # Day-0 GB200 translation of the official throughput-oriented multi_node_dep
-# profile. The 16-way data/expert-parallel worker spans four four-GPU GB200
-# nodes, with four local DP ranks on each node.
+# profile. TP2 x DP8 gives EP16 across four four-GPU GB200 nodes. TP2 shards
+# the dense weights so each DP rank has room for K3's default BF16 KV cache.
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
 
 model:
@@ -80,9 +80,9 @@ backend:
   vllm_config:
     aggregated:
       served-model-name: "moonshotai/Kimi-K3"
-      tensor-parallel-size: 1
+      tensor-parallel-size: 2
       pipeline-parallel-size: 1
-      data-parallel-size: 16
+      data-parallel-size: 8
       data-parallel-rpc-port: 13345
       enable-expert-parallel: true
       trust-remote-code: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
similarity index 89%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index 11053ad3e0..880806478d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -1,9 +1,9 @@
-name: "kimi-k3-vllm-agg-gb200-tep8-balanced-agentic"
+name: "kimi-k3-vllm-agg-gb200-tep16-balanced-agentic"
 
 # Day-0 GB200 translation of the official balanced multi_node_tep profile.
-# Dense layers use TP8 while the MoE experts are sharded across the same eight
-# GPUs on two four-GPU GB200 nodes.
-# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tep
+# Dense layers and MoE experts are sharded across 16 GPUs on four GB200 nodes,
+# leaving enough HBM for the default BF16 KV cache.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep
 
 model:
   path: "kimi-k3"
@@ -38,9 +38,9 @@ health_check:
 resources:
   gpu_type: "gb200"
   gpus_per_node: 4
-  agg_nodes: 2
+  agg_nodes: 4
   agg_workers: 1
-  gpus_per_agg: 8
+  gpus_per_agg: 16
 
 infra:
   etcd_nats_dedicated_node: false
@@ -76,11 +76,11 @@ backend:
     UCX_TLS: "rc,cuda_copy"
     NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
     NCCL_P2P_LEVEL: "NVL"
-    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tep8-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tep16-{job_id}"
   vllm_config:
     aggregated:
       served-model-name: "moonshotai/Kimi-K3"
-      tensor-parallel-size: 8
+      tensor-parallel-size: 16
       pipeline-parallel-size: 1
       enable-expert-parallel: true
       trust-remote-code: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
similarity index 89%
rename from benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index 3685e14653..f2e0345e56 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -1,9 +1,9 @@
-name: "kimi-k3-vllm-agg-gb200-tp8-latency-agentic"
+name: "kimi-k3-vllm-agg-gb200-tp16-latency-agentic"
 
 # Day-0 GB200 translation of the official latency-oriented multi_node_tp
-# profile. Kimi K3 requires at least eight GPUs for TP, so this aggregate
-# worker spans two four-GPU GB200 nodes.
-# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tp
+# profile. TP16 across four GB200 nodes leaves enough HBM for the default
+# BF16 KV cache with the 2.8T MXFP4 checkpoint.
+# https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
 
 model:
   path: "kimi-k3"
@@ -38,9 +38,9 @@ health_check:
 resources:
   gpu_type: "gb200"
   gpus_per_node: 4
-  agg_nodes: 2
+  agg_nodes: 4
   agg_workers: 1
-  gpus_per_agg: 8
+  gpus_per_agg: 16
 
 infra:
   etcd_nats_dedicated_node: false
@@ -76,11 +76,11 @@ backend:
     UCX_TLS: "rc,cuda_copy"
     NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
     NCCL_P2P_LEVEL: "NVL"
-    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tp8-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tp16-{job_id}"
   vllm_config:
     aggregated:
       served-model-name: "moonshotai/Kimi-K3"
-      tensor-parallel-size: 8
+      tensor-parallel-size: 16
       pipeline-parallel-size: 1
       trust-remote-code: true
       load-format: "fastsafetensors"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 082bc6e69e..718cef0591 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7989,9 +7989,9 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
 
 # Day-0 Kimi K3 AgentX coverage on GB200. The three aggregate srt-slurm
 # profiles mirror the serving strategies synthesized by the official recipe:
-#   latency:    multi-node TP8  (2 GB200 nodes / 8 GPUs)
-#   balanced:   multi-node TEP8 (2 GB200 nodes / 8 GPUs)
-#   throughput: multi-node DEP16 (4 GB200 nodes / 16 GPUs)
+#   latency:    multi-node TP16       (4 GB200 nodes / 16 GPUs)
+#   balanced:   multi-node TEP16      (4 GB200 nodes / 16 GPUs)
+#   throughput: multi-node TP2 x DP8  (4 GB200 nodes / EP16)
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200
 kimik3-fp4-gb200-dynamo-vllm-agentic:
   image: vllm/vllm-openai:kimi-k3
@@ -8006,56 +8006,54 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
   scenarios:
     agentic-coding:
     - search-space:
-      # Latency oriented: official multi_node_tp strategy, TP8 across two
-      # four-GPU GB200 nodes.
-      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tp
+      # Latency oriented: multi_node_tp strategy, TP16 across four GB200
+      # nodes. TP8 does not leave enough HBM for K3's default BF16 KV cache.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
       - spec-decoding: none
         conc-list: [1, 2, 4, 8]
         prefill:
           num-worker: 1
-          tp: 8
+          tp: 16
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tp8-latency-agentic.yaml"
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml"
         decode:
           num-worker: 0
-          tp: 8
+          tp: 16
           ep: 1
           dp-attn: false
-      # Balanced: official multi_node_tep strategy, TEP8 across two GB200
-      # nodes. Expert parallelism reduces MoE weight/compute pressure while
-      # retaining TP8 for dense layers.
-      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=2&strategy=multi_node_tep
+      # Balanced: multi_node_tep strategy, TEP16 across four GB200 nodes.
+      # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep
       - spec-decoding: none
         conc-list: [8, 16, 24, 32]
         prefill:
           num-worker: 1
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tep8-balanced-agentic.yaml"
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml"
         decode:
           num-worker: 0
-          tp: 8
-          ep: 8
+          tp: 16
+          ep: 16
           dp-attn: false
       # Throughput oriented: official multi_node_dep strategy, DEP16 across
-      # four GB200 nodes (TP1 x DP16 = EP16, four local DP ranks per node).
+      # four GB200 nodes (TP2 x DP8 = EP16, two local DP ranks per node).
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
       - spec-decoding: none
-        conc-list: [32, 64, 96, 128]
+        conc-list: [32, 64, 96, 128, 192, 256]
         prefill:
           num-worker: 1
-          tp: 1
+          tp: 2
           ep: 16
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml"
         decode:
           num-worker: 0
-          tp: 1
+          tp: 2
           ep: 16
           dp-attn: true
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c94fee0cf9..f9eaf3d718 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5120,5 +5120,5 @@
 - config-keys:
     - kimik3-fp4-gb200-dynamo-vllm-agentic
   description:
-    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP8 on 2 nodes (conc 1-8), TEP8 on 2 nodes (conc 8-32), and DEP16 on 4 nodes (conc 32-128). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
+    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP16 (conc 1-8), TEP16 (conc 8-32), and TP2 x DP8 = EP16 (conc 32-256). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354

From dc90e33db5f0152884ea223eea7a6f6ce0482f96 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 19:29:33 -0500
Subject: [PATCH 07/21] fix(gb200): drop unsupported Kimi K3 worker parsers
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The Dynamo vLLM worker entrypoint does not accept vllm serve tool-call or reasoning parser flags; agentic request parsing belongs to the Dynamo frontend.

中文：移除 Dynamo vLLM worker 不支持的 Kimi K3 工具调用与推理解析参数；AgentX 请求解析由 Dynamo frontend 负责。
---
 .../kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml    | 3 ---
 .../vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml | 3 ---
 .../vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml   | 3 ---
 3 files changed, 9 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index d89336be4e..4b43f34cc7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -93,9 +93,6 @@ backend:
       max-num-seqs: 128
       max-num-batched-tokens: 16384
       language-model-only: true
-      enable-auto-tool-choice: true
-      tool-call-parser: "kimi_k3"
-      reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
 
 sbatch_directives:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index 880806478d..946ed26b4c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -91,9 +91,6 @@ backend:
       max-num-seqs: 32
       max-num-batched-tokens: 8192
       language-model-only: true
-      enable-auto-tool-choice: true
-      tool-call-parser: "kimi_k3"
-      reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
       compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
 
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index f2e0345e56..a970a3fe52 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -89,9 +89,6 @@ backend:
       max-num-seqs: 8
       max-num-batched-tokens: 8192
       language-model-only: true
-      enable-auto-tool-choice: true
-      tool-call-parser: "kimi_k3"
-      reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
       compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
 

From d3b3f053a2a126ce4d71c64a971959f7bc97a18f Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 19:36:31 -0500
Subject: [PATCH 08/21] fix(gb200): let srt-slurm select collective interfaces
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove literal shell placeholders from the vLLM worker environment so Gloo and NCCL use the interfaces selected by the srt-slurm launcher.

中文：移除 vLLM worker 环境中的字面量 shell 占位符，由 srt-slurm launcher 为 Gloo 与 NCCL 选择网络接口。
---
 .../kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml     | 2 --
 .../vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml  | 2 --
 .../vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml    | 2 --
 3 files changed, 6 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 4b43f34cc7..c21d950963 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -60,8 +60,6 @@ backend:
     HF_HUB_CACHE: "/hf_hub_cache"
     HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
     TRANSFORMERS_CACHE: "/hf_hub_cache"
-    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
-    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index 946ed26b4c..c09a8e82cc 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -60,8 +60,6 @@ backend:
     HF_HUB_CACHE: "/hf_hub_cache"
     HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
     TRANSFORMERS_CACHE: "/hf_hub_cache"
-    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
-    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index a970a3fe52..926f9a95ef 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -60,8 +60,6 @@ backend:
     HF_HUB_CACHE: "/hf_hub_cache"
     HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
     TRANSFORMERS_CACHE: "/hf_hub_cache"
-    GLOO_SOCKET_IFNAME: "$IFACE_NAME"
-    NCCL_SOCKET_IFNAME: "$IFACE_NAME"
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"

From f513599de50e9a09c57f76aece1b02bbf34760b3 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 19:48:50 -0500
Subject: [PATCH 09/21] fix(gb200): launch Kimi K3 TP2 DP8 per node
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Patch the pinned agentic srt-slurm checkout to launch one vLLM process per GB200 node, with two local DP ranks and TP2 per rank. Export every per-node metrics endpoint to AIPerf.

中文：修补 AgentX 固定版本的 srt-slurm，使每个 GB200 节点启动一个 vLLM 进程，每个进程承载两个本地 DP rank，并为每个 rank 使用 TP2；同时将所有节点的 metrics endpoint 传给 AIPerf。
---
 .../configs/kimi-k3-vllm-tp-dp-per-node.patch | 101 ++++++++++++++++++
 .../agg-gb200-dep16-throughput-agentic.yaml   |   1 +
 runners/launch_gb200-nv.sh                    |   2 +
 3 files changed, 104 insertions(+)
 create mode 100644 benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch

diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch b/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch
new file mode 100644
index 0000000000..086d944559
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch
@@ -0,0 +1,101 @@
+diff --git a/src/srtctl/backends/vllm.py b/src/srtctl/backends/vllm.py
+index 828c109..5844d2a 100644
+--- a/src/srtctl/backends/vllm.py
++++ b/src/srtctl/backends/vllm.py
+@@ -33,6 +33,7 @@ if TYPE_CHECKING:
+ 
+ # Type alias for worker modes
+ WorkerMode = Literal["prefill", "decode", "agg"]
++DPLaunchMode = Literal["per_gpu", "per_node"]
+ 
+ 
+ @dataclass(frozen=True)
+@@ -93,6 +94,10 @@ class VLLMProtocol:
+     # dynamo 1.0.0+: translated to --kv-transfer-config (--connector was removed).
+     connector: str | None = "nixl"
+ 
++    # DP process layout. Per-node mode lets vLLM create multiple local DP
++    # ranks while retaining tensor parallelism within each rank.
++    dp_launch_mode: DPLaunchMode = "per_gpu"
++
+     Schema: ClassVar[builtins.type[Schema]] = Schema
+ 
+     # =========================================================================
+@@ -211,6 +216,9 @@ class VLLMProtocol:
+             # Standard TP mode: one process per node
+             return endpoints_to_processes(endpoints, base_sys_port=base_sys_port, port_allocator=port_allocator)
+ 
++        if self.dp_launch_mode == "per_node":
++            return endpoints_to_processes(endpoints, base_sys_port=base_sys_port, port_allocator=port_allocator)
++
+         # DP+EP mode: one process per GPU
+         processes: list[Process] = []
+         current_sys_port = base_sys_port
+@@ -352,7 +360,51 @@ class VLLMProtocol:
+         # Check if this is DP+EP mode (data-parallel-size set)
+         is_dp_mode = self._is_dp_mode(mode)
+ 
+-        if is_dp_mode:
++        if is_dp_mode and self.dp_launch_mode == "per_node":
++            dp_size = self._get_dp_size(mode)
++            tp_size = config.get("tensor-parallel-size") or config.get("tensor_parallel_size") or 1
++            if dp_size is None:
++                raise ValueError(f"{mode} is missing data-parallel-size")
++
++            total_gpus = sum(len(item.gpu_indices) for item in endpoint_processes)
++            if dp_size * tp_size != total_gpus:
++                raise ValueError(
++                    f"{mode} data-parallel-size={dp_size} x tensor-parallel-size={tp_size} "
++                    f"does not match the endpoint's {total_gpus} allocated GPUs"
++                )
++            if len(process.gpu_indices) % tp_size != 0:
++                raise ValueError(
++                    f"{mode} has {len(process.gpu_indices)} local GPUs, which is not divisible "
++                    f"by tensor-parallel-size={tp_size}"
++                )
++
++            local_dp_size = len(process.gpu_indices) // tp_size
++            dp_start_rank = endpoint_nodes.index(process.node) * local_dp_size
++            dp_rpc_port = config.pop("data-parallel-rpc-port", None) or config.pop(
++                "data_parallel_rpc_port",
++                VLLM_DATA_PARALLEL_RPC_PORT,
++            )
++            config.pop("data-parallel-size-local", None)
++            config.pop("data_parallel_size_local", None)
++            config.pop("data-parallel-start-rank", None)
++            config.pop("data_parallel_start_rank", None)
++            config.pop("data-parallel-hybrid-lb", None)
++            config.pop("data_parallel_hybrid_lb", None)
++
++            cmd.extend(
++                [
++                    "--data-parallel-size-local",
++                    str(local_dp_size),
++                    "--data-parallel-start-rank",
++                    str(dp_start_rank),
++                    "--data-parallel-address",
++                    leader_ip,
++                    "--data-parallel-rpc-port",
++                    str(dp_rpc_port),
++                    "--data-parallel-hybrid-lb",
++                ]
++            )
++        elif is_dp_mode:
+             # DP+EP mode: each GPU runs its own process
+             # process.node_rank is the dp_rank (set in endpoints_to_processes)
+             dp_rank = process.node_rank
+diff --git a/src/srtctl/cli/mixins/benchmark_stage.py b/src/srtctl/cli/mixins/benchmark_stage.py
+index 9a0f24a..bc7bc23 100644
+--- a/src/srtctl/cli/mixins/benchmark_stage.py
++++ b/src/srtctl/cli/mixins/benchmark_stage.py
+@@ -346,8 +346,9 @@ class BenchmarkStageMixin:
+         engine and therefore must not appear as AIPerf workers.
+         """
+         urls: list[str] = []
++        include_dp_nodes = getattr(self.config.backend, "dp_launch_mode", None) == "per_node"
+         for process in self.backend_processes:
+-            if process.is_leader and process.sys_port > 0:
++            if (process.is_leader or include_dp_nodes) and process.sys_port > 0:
+                 host = get_hostname_ip(process.node, self.runtime.network_interface)
+                 urls.append(f"http://{host}:{process.sys_port}/metrics")
+ 
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index c21d950963..35bd33f02a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -56,6 +56,7 @@ frontend:
 backend:
   type: vllm
   connector: null
+  dp_launch_mode: per_node
   aggregated_environment:
     HF_HUB_CACHE: "/hf_hub_cache"
     HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index a8c50c7d9a..b0f59de5f8 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -374,6 +374,8 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
         mkdir -p recipes/vllm/kimi-k3/agentic
         cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
             recipes/vllm/kimi-k3/agentic
+        git apply \
+            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch"
     else
         mkdir -p recipes/vllm/deepseek-v4/agentic
         cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \

From bc21aac81aed861d89b4740a594331c149d26846 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 19:58:16 -0500
Subject: [PATCH 10/21] fix(gb200): rebalance Kimi K3 EP16 for HBM
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use TP4 x DP4 for the throughput-oriented EP16 profile after TP2 x DP8 exceeded GB200 HBM during MXFP4 weight loading.

中文：TP2 × DP8 在 MXFP4 权重加载阶段超过 GB200 HBM 容量，因此将吞吐优先的 EP16 配置调整为 TP4 × DP4。
---
 .../agentic/agg-gb200-dep16-throughput-agentic.yaml       | 4 ++--
 configs/nvidia-master.yaml                                | 8 ++++----
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 35bd33f02a..6c391c3362 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -79,9 +79,9 @@ backend:
   vllm_config:
     aggregated:
       served-model-name: "moonshotai/Kimi-K3"
-      tensor-parallel-size: 2
+      tensor-parallel-size: 4
       pipeline-parallel-size: 1
-      data-parallel-size: 8
+      data-parallel-size: 4
       data-parallel-rpc-port: 13345
       enable-expert-parallel: true
       trust-remote-code: true
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 718cef0591..f616b60deb 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7991,7 +7991,7 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
 # profiles mirror the serving strategies synthesized by the official recipe:
 #   latency:    multi-node TP16       (4 GB200 nodes / 16 GPUs)
 #   balanced:   multi-node TEP16      (4 GB200 nodes / 16 GPUs)
-#   throughput: multi-node TP2 x DP8  (4 GB200 nodes / EP16)
+#   throughput: multi-node TP4 x DP4  (4 GB200 nodes / EP16)
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200
 kimik3-fp4-gb200-dynamo-vllm-agentic:
   image: vllm/vllm-openai:kimi-k3
@@ -8040,20 +8040,20 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           ep: 16
           dp-attn: false
       # Throughput oriented: official multi_node_dep strategy, DEP16 across
-      # four GB200 nodes (TP2 x DP8 = EP16, two local DP ranks per node).
+      # four GB200 nodes (TP4 x DP4 = EP16, one local DP rank per node).
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
       - spec-decoding: none
         conc-list: [32, 64, 96, 128, 192, 256]
         prefill:
           num-worker: 1
-          tp: 2
+          tp: 4
           ep: 16
           dp-attn: true
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml"
         decode:
           num-worker: 0
-          tp: 2
+          tp: 4
           ep: 16
           dp-attn: true
 

From 819af419c455ac66f8931a7170c4098ff60f39d6 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 20:08:08 -0500
Subject: [PATCH 11/21] fix(gb200): use native srt-slurm TPDP support
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Pin a tested srt-slurm revision that natively groups TP GPUs within each DP rank and exports every DP metrics endpoint. Remove the vendored patch and reserve enough HBM for one native 1M-context request.

中文：固定经过测试的 srt-slurm 版本，原生支持每个 DP rank 内的 TP GPU 分组，并导出全部 DP metrics endpoint。删除仓库内补丁，同时为原生 1M 上下文请求预留足够 HBM。
---
 .../configs/kimi-k3-vllm-tp-dp-per-node.patch | 101 ------------------
 .../agg-gb200-dep16-throughput-agentic.yaml   |   3 +-
 runners/launch_gb200-nv.sh                    |  10 +-
 3 files changed, 6 insertions(+), 108 deletions(-)
 delete mode 100644 benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch

diff --git a/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch b/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch
deleted file mode 100644
index 086d944559..0000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch
+++ /dev/null
@@ -1,101 +0,0 @@
-diff --git a/src/srtctl/backends/vllm.py b/src/srtctl/backends/vllm.py
-index 828c109..5844d2a 100644
---- a/src/srtctl/backends/vllm.py
-+++ b/src/srtctl/backends/vllm.py
-@@ -33,6 +33,7 @@ if TYPE_CHECKING:
- 
- # Type alias for worker modes
- WorkerMode = Literal["prefill", "decode", "agg"]
-+DPLaunchMode = Literal["per_gpu", "per_node"]
- 
- 
- @dataclass(frozen=True)
-@@ -93,6 +94,10 @@ class VLLMProtocol:
-     # dynamo 1.0.0+: translated to --kv-transfer-config (--connector was removed).
-     connector: str | None = "nixl"
- 
-+    # DP process layout. Per-node mode lets vLLM create multiple local DP
-+    # ranks while retaining tensor parallelism within each rank.
-+    dp_launch_mode: DPLaunchMode = "per_gpu"
-+
-     Schema: ClassVar[builtins.type[Schema]] = Schema
- 
-     # =========================================================================
-@@ -211,6 +216,9 @@ class VLLMProtocol:
-             # Standard TP mode: one process per node
-             return endpoints_to_processes(endpoints, base_sys_port=base_sys_port, port_allocator=port_allocator)
- 
-+        if self.dp_launch_mode == "per_node":
-+            return endpoints_to_processes(endpoints, base_sys_port=base_sys_port, port_allocator=port_allocator)
-+
-         # DP+EP mode: one process per GPU
-         processes: list[Process] = []
-         current_sys_port = base_sys_port
-@@ -352,7 +360,51 @@ class VLLMProtocol:
-         # Check if this is DP+EP mode (data-parallel-size set)
-         is_dp_mode = self._is_dp_mode(mode)
- 
--        if is_dp_mode:
-+        if is_dp_mode and self.dp_launch_mode == "per_node":
-+            dp_size = self._get_dp_size(mode)
-+            tp_size = config.get("tensor-parallel-size") or config.get("tensor_parallel_size") or 1
-+            if dp_size is None:
-+                raise ValueError(f"{mode} is missing data-parallel-size")
-+
-+            total_gpus = sum(len(item.gpu_indices) for item in endpoint_processes)
-+            if dp_size * tp_size != total_gpus:
-+                raise ValueError(
-+                    f"{mode} data-parallel-size={dp_size} x tensor-parallel-size={tp_size} "
-+                    f"does not match the endpoint's {total_gpus} allocated GPUs"
-+                )
-+            if len(process.gpu_indices) % tp_size != 0:
-+                raise ValueError(
-+                    f"{mode} has {len(process.gpu_indices)} local GPUs, which is not divisible "
-+                    f"by tensor-parallel-size={tp_size}"
-+                )
-+
-+            local_dp_size = len(process.gpu_indices) // tp_size
-+            dp_start_rank = endpoint_nodes.index(process.node) * local_dp_size
-+            dp_rpc_port = config.pop("data-parallel-rpc-port", None) or config.pop(
-+                "data_parallel_rpc_port",
-+                VLLM_DATA_PARALLEL_RPC_PORT,
-+            )
-+            config.pop("data-parallel-size-local", None)
-+            config.pop("data_parallel_size_local", None)
-+            config.pop("data-parallel-start-rank", None)
-+            config.pop("data_parallel_start_rank", None)
-+            config.pop("data-parallel-hybrid-lb", None)
-+            config.pop("data_parallel_hybrid_lb", None)
-+
-+            cmd.extend(
-+                [
-+                    "--data-parallel-size-local",
-+                    str(local_dp_size),
-+                    "--data-parallel-start-rank",
-+                    str(dp_start_rank),
-+                    "--data-parallel-address",
-+                    leader_ip,
-+                    "--data-parallel-rpc-port",
-+                    str(dp_rpc_port),
-+                    "--data-parallel-hybrid-lb",
-+                ]
-+            )
-+        elif is_dp_mode:
-             # DP+EP mode: each GPU runs its own process
-             # process.node_rank is the dp_rank (set in endpoints_to_processes)
-             dp_rank = process.node_rank
-diff --git a/src/srtctl/cli/mixins/benchmark_stage.py b/src/srtctl/cli/mixins/benchmark_stage.py
-index 9a0f24a..bc7bc23 100644
---- a/src/srtctl/cli/mixins/benchmark_stage.py
-+++ b/src/srtctl/cli/mixins/benchmark_stage.py
-@@ -346,8 +346,9 @@ class BenchmarkStageMixin:
-         engine and therefore must not appear as AIPerf workers.
-         """
-         urls: list[str] = []
-+        include_dp_nodes = getattr(self.config.backend, "dp_launch_mode", None) == "per_node"
-         for process in self.backend_processes:
--            if process.is_leader and process.sys_port > 0:
-+            if (process.is_leader or include_dp_nodes) and process.sys_port > 0:
-                 host = get_hostname_ip(process.node, self.runtime.network_interface)
-                 urls.append(f"http://{host}:{process.sys_port}/metrics")
- 
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 6c391c3362..f4f5c99fed 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -56,7 +56,6 @@ frontend:
 backend:
   type: vllm
   connector: null
-  dp_launch_mode: per_node
   aggregated_environment:
     HF_HUB_CACHE: "/hf_hub_cache"
     HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
@@ -88,7 +87,7 @@ backend:
       load-format: "fastsafetensors"
       moe-backend: "deep_gemm_mega_moe"
       all2all-backend: "flashinfer_nvlink_one_sided"
-      gpu-memory-utilization: 0.95
+      gpu-memory-utilization: 0.98
       max-num-seqs: 128
       max-num-batched-tokens: 16384
       language-model-only: true
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index b0f59de5f8..4309d087b1 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -358,8 +358,7 @@ fi
 
 # TODO(CJQ): make first class upon srt-slurm upstream refactor
 if [[ "$IS_AGENTIC" == "1" ]]; then
-    # Agentic multi-node uses the same pinned cquil11/srt-slurm-nv commit as
-    # launch_gb300-nv.sh — everything the agentic recipes need is there:
+    # Agentic multi-node pins cquil11/srt-slurm-nv revisions that provide:
     #   - BenchmarkType.CUSTOM + benchmark.command + benchmark.env
     #     (the hook that hands off to benchmarks/multi_node/agentic_srt.sh)
     #   - DynamoConfig.wheel (recipes pin the ai-dynamo wheel)
@@ -369,14 +368,15 @@ if [[ "$IS_AGENTIC" == "1" ]]; then
     #     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
     if [[ "$MODEL_PREFIX" == "kimik3" ]]; then
+        # Kimi K3 additionally needs vLLM TP groups within data parallel and
+        # every DP engine metrics endpoint exposed to AIPerf.
+        git checkout b1fb626fbdbfe3306dcb51cb181ab35861ec3b1c
         mkdir -p recipes/vllm/kimi-k3/agentic
         cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
             recipes/vllm/kimi-k3/agentic
-        git apply \
-            "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/configs/kimi-k3-vllm-tp-dp-per-node.patch"
     else
+        git checkout de59739b172e507e15ebf145bfe305f606e82fbf
         mkdir -p recipes/vllm/deepseek-v4/agentic
         cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
             recipes/vllm/deepseek-v4/agentic

From 458d6f669c87187750adab6fe0150cb7645c0fd7 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 20:10:58 -0500
Subject: [PATCH 12/21] docs(perf): correct Kimi K3 throughput topology
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Record the tested TP4 x DP4 = EP16 throughput profile.

中文：记录经过测试的 TP4 x DP4 = EP16 吞吐配置。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f9eaf3d718..1549d55202 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5120,5 +5120,5 @@
 - config-keys:
     - kimik3-fp4-gb200-dynamo-vllm-agentic
   description:
-    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP16 (conc 1-8), TEP16 (conc 8-32), and TP2 x DP8 = EP16 (conc 32-256). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
+    - "Add Day-0 Kimi K3 MXFP4 GB200 AgentX performance points following the official vLLM recipe: TP16 (conc 1-8), TEP16 (conc 8-32), and TP4 x DP4 = EP16 (conc 32-256). https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2354

From 6d10fbc53353946b3c8a5b382a448093049717e8 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 20:30:36 -0500
Subject: [PATCH 13/21] fix(gb200): use official Kimi K3 Dynamo runtime
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use NVIDIA’s Kimi K3 runtime with native tokenizer, parser, FP8 KV-cache, and FlashInfer support. Keep the model-default 1M context and remove the vLLM-mutating setup script.\n\n中文：改用 NVIDIA 官方 Kimi K3 Dynamo 运行时，启用原生 tokenizer、解析器、FP8 KV 缓存和 FlashInfer 支持。保留模型默认的 1M 上下文，并移除会修改 vLLM 文件的环境配置脚本。
---
 .../agg-gb200-dep16-throughput-agentic.yaml   | 43 +++++++++++++------
 .../agg-gb200-tep16-balanced-agentic.yaml     | 42 ++++++++++++------
 .../agg-gb200-tp16-latency-agentic.yaml       | 41 ++++++++++++------
 configs/nvidia-master.yaml                    |  6 +--
 4 files changed, 89 insertions(+), 43 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index f4f5c99fed..7404aee47e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -1,33 +1,28 @@
 name: "kimi-k3-vllm-agg-gb200-dep16-throughput-agentic"
 
 # Day-0 GB200 translation of the official throughput-oriented multi_node_dep
-# profile. TP2 x DP8 gives EP16 across four four-GPU GB200 nodes. TP2 shards
-# the dense weights so each DP rank has room for K3's default BF16 KV cache.
+# profile. TP4 x DP4 gives EP16 across four four-GPU GB200 nodes.
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
 
 model:
   path: "kimi-k3"
-  container: "vllm/vllm-openai:kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:kimi-k3"
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.3.0.dev20260618"
+    dynamo: "1.4.0-kimi-k3-dev.1"
 
 dynamo:
-  install: true
-  wheel: "1.3.0.dev20260618"
+  install: false
 
 environment:
-  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
   ETCD_LEASE_TTL: "7200"
 
-setup_script: vllm-container-deps.sh
-
 slurm:
   time_limit: "8:00:00"
 
@@ -50,8 +45,12 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
+    dyn-chat-processor: "dynamo"
     router-mode: "kv"
-    router-reset-states: true
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
 
 backend:
   type: vllm
@@ -63,9 +62,10 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
-    VLLM_USE_V2_MODEL_RUNNER: "1"
-    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
     VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
     NCCL_CUMEM_ENABLE: "1"
     NCCL_MNNVL_ENABLE: "1"
     NCCL_NVLS_ENABLE: "1"
@@ -74,6 +74,8 @@ backend:
     UCX_TLS: "rc,cuda_copy"
     NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
     NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
     DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-dep16-{job_id}"
   vllm_config:
     aggregated:
@@ -85,12 +87,25 @@ backend:
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       moe-backend: "deep_gemm_mega_moe"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
       all2all-backend: "flashinfer_nvlink_one_sided"
-      gpu-memory-utilization: 0.98
+      gpu-memory-utilization: 0.95
       max-num-seqs: 128
       max-num-batched-tokens: 16384
+      block-size: 64
       language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
 
 sbatch_directives:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index c09a8e82cc..3cced99c6b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -1,33 +1,29 @@
 name: "kimi-k3-vllm-agg-gb200-tep16-balanced-agentic"
 
 # Day-0 GB200 translation of the official balanced multi_node_tep profile.
-# Dense layers and MoE experts are sharded across 16 GPUs on four GB200 nodes,
-# leaving enough HBM for the default BF16 KV cache.
+# Dense layers and MoE experts are sharded across 16 GPUs on four GB200 nodes
+# with the official FP8 KV cache.
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep
 
 model:
   path: "kimi-k3"
-  container: "vllm/vllm-openai:kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:kimi-k3"
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.3.0.dev20260618"
+    dynamo: "1.4.0-kimi-k3-dev.1"
 
 dynamo:
-  install: true
-  wheel: "1.3.0.dev20260618"
+  install: false
 
 environment:
-  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
   ETCD_LEASE_TTL: "7200"
 
-setup_script: vllm-container-deps.sh
-
 slurm:
   time_limit: "8:00:00"
 
@@ -50,8 +46,12 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
+    dyn-chat-processor: "dynamo"
     router-mode: "kv"
-    router-reset-states: true
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
 
 backend:
   type: vllm
@@ -63,9 +63,10 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
-    VLLM_USE_V2_MODEL_RUNNER: "1"
-    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
     VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
     NCCL_CUMEM_ENABLE: "1"
     NCCL_MNNVL_ENABLE: "1"
     NCCL_NVLS_ENABLE: "1"
@@ -74,6 +75,8 @@ backend:
     UCX_TLS: "rc,cuda_copy"
     NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
     NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
     DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tep16-{job_id}"
   vllm_config:
     aggregated:
@@ -83,12 +86,25 @@ backend:
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       moe-backend: "flashinfer_trtllm"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
       all2all-backend: "flashinfer_nvlink_one_sided"
       gpu-memory-utilization: 0.95
       max-num-seqs: 32
       max-num-batched-tokens: 8192
+      block-size: 64
       language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
       compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
 
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index 926f9a95ef..b381ecd1a7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -1,33 +1,28 @@
 name: "kimi-k3-vllm-agg-gb200-tp16-latency-agentic"
 
 # Day-0 GB200 translation of the official latency-oriented multi_node_tp
-# profile. TP16 across four GB200 nodes leaves enough HBM for the default
-# BF16 KV cache with the 2.8T MXFP4 checkpoint.
+# profile. TP16 spans four GB200 nodes and uses the official FP8 KV cache.
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
 
 model:
   path: "kimi-k3"
-  container: "vllm/vllm-openai:kimi-k3"
+  container: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:kimi-k3"
+    image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.3.0.dev20260618"
+    dynamo: "1.4.0-kimi-k3-dev.1"
 
 dynamo:
-  install: true
-  wheel: "1.3.0.dev20260618"
+  install: false
 
 environment:
-  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
   ETCD_LEASE_TTL: "7200"
 
-setup_script: vllm-container-deps.sh
-
 slurm:
   time_limit: "8:00:00"
 
@@ -50,8 +45,12 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
+    dyn-chat-processor: "dynamo"
     router-mode: "kv"
-    router-reset-states: true
+    router-kv-events: true
+    router-temperature: "0"
+    router-min-initial-workers: 1
+    kv-cache-block-size: 64
 
 backend:
   type: vllm
@@ -63,9 +62,10 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "7200"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
-    VLLM_USE_V2_MODEL_RUNNER: "1"
-    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "0"
     VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "mnnvl"
     NCCL_CUMEM_ENABLE: "1"
     NCCL_MNNVL_ENABLE: "1"
     NCCL_NVLS_ENABLE: "1"
@@ -74,6 +74,8 @@ backend:
     UCX_TLS: "rc,cuda_copy"
     NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
     NCCL_P2P_LEVEL: "NVL"
+    NVIDIA_GDRCOPY: "1"
+    PYTORCH_ALLOC_CONF: "expandable_segments:True"
     DG_JIT_CACHE_DIR: "/tmp/dg-cache-kimi-k3-gb200-tp16-{job_id}"
   vllm_config:
     aggregated:
@@ -82,11 +84,24 @@ backend:
       pipeline-parallel-size: 1
       trust-remote-code: true
       load-format: "fastsafetensors"
+      safetensors-load-strategy: "lazy"
+      kv-cache-dtype: "fp8"
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       moe-backend: "flashinfer_trtllm"
+      kda-prefill-backend: "flashkda"
+      kernel-config: '{"enable_cutedsl_warmup":true}'
       gpu-memory-utilization: 0.95
       max-num-seqs: 8
       max-num-batched-tokens: 8192
+      block-size: 64
       language-model-only: true
+      disable-custom-all-reduce: true
+      enable-prefix-caching: true
+      scheduler-cls: "vllm.v1.core.sched.async_scheduler.AsyncScheduler"
+      dyn-tool-call-parser: "kimi_k3"
+      reasoning-parser: "kimi_k3"
+      dyn-reasoning-parser: "kimi_k3"
       no-enable-flashinfer-autotune: true
       compilation-config: '{"pass_config":{"fuse_allreduce_rms":false}}'
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f616b60deb..240ade1b55 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7994,20 +7994,20 @@ dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
 #   throughput: multi-node TP4 x DP4  (4 GB200 nodes / EP16)
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200
 kimik3-fp4-gb200-dynamo-vllm-agentic:
-  image: vllm/vllm-openai:kimi-k3
+  image: nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:gb200-nv
   precision: fp4
   framework: dynamo-vllm
-  router: { name: dynamo-router, version: "1.3.0.dev20260618" }
+  router: { name: dynamo-router, version: "1.4.0-kimi-k3-dev.1" }
   multinode: true
   disagg: false
   scenarios:
     agentic-coding:
     - search-space:
       # Latency oriented: multi_node_tp strategy, TP16 across four GB200
-      # nodes. TP8 does not leave enough HBM for K3's default BF16 KV cache.
+      # nodes.
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tp
       - spec-decoding: none
         conc-list: [1, 2, 4, 8]

From 41580415773dd1ea79fc106d13c03d57684a5cc4 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 20:53:02 -0500
Subject: [PATCH 14/21] fix(gb200): align Kimi K3 runtime metadata
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Match identity verification and result metadata to the Dynamo 1.3.0 package reported by the official Kimi K3 runtime image.\n\n中文：将身份校验与结果元数据对齐到官方 Kimi K3 运行时镜像实际报告的 Dynamo 1.3.0 包版本。
---
 .../kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml     | 2 +-
 .../vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml  | 2 +-
 .../vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml    | 2 +-
 configs/nvidia-master.yaml                                      | 2 +-
 4 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 7404aee47e..36984f889e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -15,7 +15,7 @@ identity:
   container:
     image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.4.0-kimi-k3-dev.1"
+    dynamo: "1.3.0"
 
 dynamo:
   install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index 3cced99c6b..fa99537d07 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -16,7 +16,7 @@ identity:
   container:
     image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.4.0-kimi-k3-dev.1"
+    dynamo: "1.3.0"
 
 dynamo:
   install: false
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index b381ecd1a7..61cae58cd6 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -15,7 +15,7 @@ identity:
   container:
     image: "nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1"
   frameworks:
-    dynamo: "1.4.0-kimi-k3-dev.1"
+    dynamo: "1.3.0"
 
 dynamo:
   install: false
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 240ade1b55..2625eb0eef 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8000,7 +8000,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
   runner: cluster:gb200-nv
   precision: fp4
   framework: dynamo-vllm
-  router: { name: dynamo-router, version: "1.4.0-kimi-k3-dev.1" }
+  router: { name: dynamo-router, version: "1.3.0" }
   multinode: true
   disagg: false
   scenarios:

From a32ee0b18e6cb136af60ba93b36b59aee043fbc2 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 20:58:55 -0500
Subject: [PATCH 15/21] fix(gb200): reserve Kimi K3 prefill workspace
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use the official Kimi K3 Dynamo 0.92 GPU memory utilization so native 1M-context MLA prefills retain transient workspace without changing max model length.\n\n中文：采用官方 Kimi K3 Dynamo 配方的 0.92 GPU 显存利用率，为原生 1M 上下文 MLA 预填充保留临时工作区，且不修改最大模型长度。
---
 .../kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml     | 2 +-
 .../vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml  | 2 +-
 .../vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml    | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 36984f889e..d281f9cd5e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -95,7 +95,7 @@ backend:
       kda-prefill-backend: "flashkda"
       kernel-config: '{"enable_cutedsl_warmup":true}'
       all2all-backend: "flashinfer_nvlink_one_sided"
-      gpu-memory-utilization: 0.95
+      gpu-memory-utilization: 0.92
       max-num-seqs: 128
       max-num-batched-tokens: 16384
       block-size: 64
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index fa99537d07..6db5821a1a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -94,7 +94,7 @@ backend:
       kda-prefill-backend: "flashkda"
       kernel-config: '{"enable_cutedsl_warmup":true}'
       all2all-backend: "flashinfer_nvlink_one_sided"
-      gpu-memory-utilization: 0.95
+      gpu-memory-utilization: 0.92
       max-num-seqs: 32
       max-num-batched-tokens: 8192
       block-size: 64
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index 61cae58cd6..8c5389505b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -91,7 +91,7 @@ backend:
       moe-backend: "flashinfer_trtllm"
       kda-prefill-backend: "flashkda"
       kernel-config: '{"enable_cutedsl_warmup":true}'
-      gpu-memory-utilization: 0.95
+      gpu-memory-utilization: 0.92
       max-num-seqs: 8
       max-num-batched-tokens: 8192
       block-size: 64

From 86f4e7009c4a5b406c45b2f47a3e2bc2d783c3b8 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 21:14:24 -0500
Subject: [PATCH 16/21] fix(config): report Kimi K3 TP16 as EP16
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Apply vLLM EP_SIZE = TP_SIZE × DP_SIZE in the master metadata for the DP1 latency profile.\n\n中文：在 DP1 延迟配置的主配置元数据中应用 vLLM 的 EP_SIZE = TP_SIZE × DP_SIZE 规则。
---
 configs/nvidia-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 2625eb0eef..01f09a5387 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8014,14 +8014,14 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
         prefill:
           num-worker: 1
           tp: 16
-          ep: 1
+          ep: 16
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml"
         decode:
           num-worker: 0
           tp: 16
-          ep: 1
+          ep: 16
           dp-attn: false
       # Balanced: multi_node_tep strategy, TEP16 across four GB200 nodes.
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_tep

From 1f052d2387fb12e439973449a992f4c6b705be21 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 21:34:42 -0500
Subject: [PATCH 17/21] fix(gb200): stage srt AgentX aggregates
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Copy per-concurrency AgentX aggregate JSON files from the compute-visible shared workspace back into the GitHub workspace, and cover success and missing-result behavior with regression tests.

中文：将各并发度的 AgentX 聚合 JSON 从计算节点可见的共享工作区复制回 GitHub 工作区，并通过回归测试覆盖成功复制和结果缺失场景。
---
 runners/launch_gb200-nv.sh  | 79 +++++++++++++++++++++----------------
 runners/slurm_utils.sh      | 30 ++++++++++++++
 runners/test_slurm_utils.py | 57 ++++++++++++++++++++++++++
 3 files changed, 132 insertions(+), 34 deletions(-)
 create mode 100644 runners/test_slurm_utils.py

diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 4309d087b1..2f913e692c 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -644,43 +644,54 @@ if [[ "${EVAL_ONLY:-false}" != "true" ]]; then
         exit 1
     fi
 
-    # Find all result subdirectories
-    RESULT_SUBDIRS=$(find "$LOGS_DIR" -maxdepth 1 -type d -name "*isl*osl*" 2>/dev/null)
-
-    if [ -z "$RESULT_SUBDIRS" ]; then
-        echo "Warning: No result subdirectories found in $LOGS_DIR"
+    if [[ "$IS_AGENTIC" == "1" ]]; then
+        # The custom benchmark runs inside the compute-visible
+        # INFMAX_WORKSPACE mount. Its aggregation step writes one
+        # ${RESULT_FILENAME}_conc<N>.json there per point; stage those files
+        # back to GITHUB_WORKSPACE for the workflow guard and artifact upload.
+        copy_agentic_results \
+            "$INFMAX_WORKSPACE" \
+            "$GITHUB_WORKSPACE" \
+            "$RESULT_FILENAME" || exit 1
     else
-        # Process results from all configurations
-        for result_subdir in $RESULT_SUBDIRS; do
-            echo "Processing result subdirectory: $result_subdir"
-
-            # Extract configuration info from directory name
-            CONFIG_NAME=$(basename "$result_subdir")
-
-            # Find all result JSON files
-            RESULT_FILES=$(find "$result_subdir" -name "results_concurrency_*.json" 2>/dev/null)
-
-            for result_file in $RESULT_FILES; do
-                if [ -f "$result_file" ]; then
-                    # Extract metadata from filename
-                    # Files may be "results_concurrency_N_gpus_G_ctx_C_gen_D.json" (disagg) or "results_concurrency_N_gpus_G.json" (non-disagg)
-                    filename=$(basename "$result_file")
-                    concurrency=$(echo "$filename" | sed -n 's/results_concurrency_\([0-9]*\)_gpus_.*/\1/p')
-                    gpus=$(echo "$filename" | sed -n 's/results_concurrency_[0-9]*_gpus_\([0-9][0-9]*\).*/\1/p')
-                    ctx=$(echo "$filename" | sed -n 's/.*_ctx_\([0-9]*\)_gen_.*/\1/p')
-                    gen=$(echo "$filename" | sed -n 's/.*_gen_\([0-9]*\)\.json/\1/p')
-
-                    echo "Processing concurrency $concurrency with $gpus GPUs (ctx: $ctx, gen: $gen): $result_file"
-
-                    if [ -n "$ctx" ] && [ -n "$gen" ]; then
-                        WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}_ctx_${ctx}_gen_${gen}.json"
-                    else
-                        WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}.json"
+        # Find all fixed-sequence result subdirectories.
+        RESULT_SUBDIRS=$(find "$LOGS_DIR" -maxdepth 1 -type d -name "*isl*osl*" 2>/dev/null)
+
+        if [ -z "$RESULT_SUBDIRS" ]; then
+            echo "Warning: No result subdirectories found in $LOGS_DIR"
+        else
+            # Process results from all configurations
+            for result_subdir in $RESULT_SUBDIRS; do
+                echo "Processing result subdirectory: $result_subdir"
+
+                # Extract configuration info from directory name
+                CONFIG_NAME=$(basename "$result_subdir")
+
+                # Find all result JSON files
+                RESULT_FILES=$(find "$result_subdir" -name "results_concurrency_*.json" 2>/dev/null)
+
+                for result_file in $RESULT_FILES; do
+                    if [ -f "$result_file" ]; then
+                        # Extract metadata from filename
+                        # Files may be "results_concurrency_N_gpus_G_ctx_C_gen_D.json" (disagg) or "results_concurrency_N_gpus_G.json" (non-disagg)
+                        filename=$(basename "$result_file")
+                        concurrency=$(echo "$filename" | sed -n 's/results_concurrency_\([0-9]*\)_gpus_.*/\1/p')
+                        gpus=$(echo "$filename" | sed -n 's/results_concurrency_[0-9]*_gpus_\([0-9][0-9]*\).*/\1/p')
+                        ctx=$(echo "$filename" | sed -n 's/.*_ctx_\([0-9]*\)_gen_.*/\1/p')
+                        gen=$(echo "$filename" | sed -n 's/.*_gen_\([0-9]*\)\.json/\1/p')
+
+                        echo "Processing concurrency $concurrency with $gpus GPUs (ctx: $ctx, gen: $gen): $result_file"
+
+                        if [ -n "$ctx" ] && [ -n "$gen" ]; then
+                            WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}_ctx_${ctx}_gen_${gen}.json"
+                        else
+                            WORKSPACE_RESULT_FILE="$GITHUB_WORKSPACE/${RESULT_FILENAME}_${CONFIG_NAME}_conc${concurrency}_gpus_${gpus}.json"
+                        fi
+                        copy_to_workspace "$result_file" "$WORKSPACE_RESULT_FILE" || exit 1
                     fi
-                    copy_to_workspace "$result_file" "$WORKSPACE_RESULT_FILE" || exit 1
-                fi
+                done
             done
-        done
+        fi
     fi
 
     echo "All result files processed"
diff --git a/runners/slurm_utils.sh b/runners/slurm_utils.sh
index c26b4f1690..55c8652b84 100644
--- a/runners/slurm_utils.sh
+++ b/runners/slurm_utils.sh
@@ -41,6 +41,36 @@ copy_to_workspace() {
     echo "Copied $(basename "$source_file") to $destination_file"
 }
 
+copy_agentic_results() {
+    local source_dir="$1"
+    local workspace="$2"
+    local result_filename="$3"
+    local result_file
+    local copied=0
+
+    if [[ ! -d "$source_dir" ]]; then
+        echo "ERROR: agentic result directory not found at $source_dir" >&2
+        return 1
+    fi
+
+    while IFS= read -r -d '' result_file; do
+        copy_to_workspace \
+            "$result_file" \
+            "$workspace/$(basename "$result_file")" || return 1
+        copied=$((copied + 1))
+    done < <(
+        find "$source_dir" -maxdepth 1 -type f \
+            -name "${result_filename}_conc*.json" -print0
+    )
+
+    if [[ "$copied" -eq 0 ]]; then
+        echo "ERROR: no ${result_filename}_conc*.json results found in $source_dir" >&2
+        return 1
+    fi
+
+    echo "Copied $copied agentic result file(s)"
+}
+
 copy_eval_artifacts() {
     local eval_dir="$1"
     local workspace="$2"
diff --git a/runners/test_slurm_utils.py b/runners/test_slurm_utils.py
new file mode 100644
index 0000000000..c027cc9b2f
--- /dev/null
+++ b/runners/test_slurm_utils.py
@@ -0,0 +1,57 @@
+import subprocess
+from pathlib import Path
+
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SLURM_UTILS = REPO_ROOT / "runners" / "slurm_utils.sh"
+
+
+def run_bash(command: str, *args: Path | str) -> subprocess.CompletedProcess[str]:
+    return subprocess.run(
+        ["bash", "-c", command, "bash", *(str(arg) for arg in args)],
+        check=False,
+        capture_output=True,
+        text=True,
+    )
+
+
+def test_copy_agentic_results_stages_only_matching_points(tmp_path: Path) -> None:
+    source = tmp_path / "source"
+    workspace = tmp_path / "workspace"
+    source.mkdir()
+    workspace.mkdir()
+    (source / "run_conc1.json").write_text('{"conc": 1}\n')
+    (source / "run_conc16.json").write_text('{"conc": 16}\n')
+    (source / "other_conc1.json").write_text('{"conc": 1}\n')
+
+    result = run_bash(
+        'source "$1"; copy_agentic_results "$2" "$3" run',
+        SLURM_UTILS,
+        source,
+        workspace,
+    )
+
+    assert result.returncode == 0, result.stderr
+    assert sorted(path.name for path in workspace.iterdir()) == [
+        "run_conc1.json",
+        "run_conc16.json",
+    ]
+
+
+def test_copy_agentic_results_fails_when_aggregate_is_missing(
+    tmp_path: Path,
+) -> None:
+    source = tmp_path / "source"
+    workspace = tmp_path / "workspace"
+    source.mkdir()
+    workspace.mkdir()
+
+    result = run_bash(
+        'source "$1"; copy_agentic_results "$2" "$3" run',
+        SLURM_UTILS,
+        source,
+        workspace,
+    )
+
+    assert result.returncode != 0
+    assert "no run_conc*.json results found" in result.stderr

From 64de5595dfa9d3ed204f428916fb718b31f01414 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 22:31:09 -0500
Subject: [PATCH 18/21] fix(agentx): burst Kimi K3 saturation starts
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Enable AIPerf burst phase starts only for the Kimi K3 GB200 saturation recipes so warmup and profiling reach the requested concurrency within the benchmark window. Inter-turn replay timing remains unchanged.

中文：仅在 Kimi K3 GB200 饱和度测试配置中启用 AIPerf 阶段起始突发，使预热和正式测试能在基准测试窗口内达到目标并发度，同时保持轮次间回放时序不变。
---
 benchmarks/benchmark_lib.sh                                | 7 +++++++
 .../agentic/agg-gb200-dep16-throughput-agentic.yaml        | 1 +
 .../kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml  | 1 +
 .../kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml    | 1 +
 4 files changed, 10 insertions(+)

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index aa9cc6304f..33349852ad 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -1775,6 +1775,13 @@ build_replay_cmd() {
     # least one profile turn after warmup.
     REPLAY_CMD+=" --trajectory-start-min-ratio 0.25"
     REPLAY_CMD+=" --trajectory-start-max-ratio 0.75"
+    # Saturation sweeps need the requested lanes to become active inside the
+    # benchmark window. Agentic replay otherwise preserves the traces' original
+    # phase-start offsets; long-lived traces can spread the initial warmup over
+    # many hours before cache-pressure warmup or profiling begins.
+    if [ "${AIPERF_BURST_PHASE_STARTS:-0}" = "1" ]; then
+        REPLAY_CMD+=" --burst-phase-starts"
+    fi
     # After the normal t* snapshot warmup, continue those exact trajectories
     # with one-token outputs and no idle delays for 10 minutes. Profiling begins
     # only after those requests drain and resumes from the resulting live state.
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index d281f9cd5e..36170adaa3 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -125,6 +125,7 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index 6db5821a1a..bfc1530dac 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -125,6 +125,7 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index 8c5389505b..cffcc2cb6b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -122,6 +122,7 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"

From ff5fb061ba2195a765fa22fb7f752a6fdfcd6ec4 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 27 Jul 2026 22:58:04 -0500
Subject: [PATCH 19/21] refactor(kimik3): depend on shared phase starts
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove the duplicated AIPerf command-builder hook and Kimi-specific opt-ins. The shared default is tracked in InferenceX PR #2375.

中文：移除重复的 AIPerf 命令构建逻辑和 Kimi 专用开关，统一依赖 InferenceX PR #2375 提供的共享默认行为。
---
 benchmarks/benchmark_lib.sh                                | 7 -------
 .../agentic/agg-gb200-dep16-throughput-agentic.yaml        | 1 -
 .../kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml  | 1 -
 .../kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml    | 1 -
 4 files changed, 10 deletions(-)

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 33349852ad..aa9cc6304f 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -1775,13 +1775,6 @@ build_replay_cmd() {
     # least one profile turn after warmup.
     REPLAY_CMD+=" --trajectory-start-min-ratio 0.25"
     REPLAY_CMD+=" --trajectory-start-max-ratio 0.75"
-    # Saturation sweeps need the requested lanes to become active inside the
-    # benchmark window. Agentic replay otherwise preserves the traces' original
-    # phase-start offsets; long-lived traces can spread the initial warmup over
-    # many hours before cache-pressure warmup or profiling begins.
-    if [ "${AIPERF_BURST_PHASE_STARTS:-0}" = "1" ]; then
-        REPLAY_CMD+=" --burst-phase-starts"
-    fi
     # After the normal t* snapshot warmup, continue those exact trajectories
     # with one-token outputs and no idle delays for 10 minutes. Profiling begins
     # only after those requests drain and resumes from the resulting live state.
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index 36170adaa3..d281f9cd5e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -125,7 +125,6 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
index bfc1530dac..6db5821a1a 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tep16-balanced-agentic.yaml
@@ -125,7 +125,6 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
index cffcc2cb6b..8c5389505b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-tp16-latency-agentic.yaml
@@ -122,7 +122,6 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_BURST_PHASE_STARTS: "1"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"

From 235aa5eb4615598e3e69081a216daae786134062 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 28 Jul 2026 02:05:41 -0500
Subject: [PATCH 20/21] fix(agentx): extend Kimi K3 DEP warmup drain
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Allow the full-context DEP saturation sweep up to 3600 seconds to drain accelerated warmup requests. The grace period exits early when the servers become idle.

中文：将 Kimi K3 DEP 全上下文饱和度测试的加速预热排空上限延长至 3600 秒；服务空闲后会提前结束等待。
---
 .../vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml | 1 +
 configs/nvidia-master.yaml                                       | 1 +
 2 files changed, 2 insertions(+)

diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
index d281f9cd5e..04b082bcf9 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb200-dep16-throughput-agentic.yaml
@@ -126,6 +126,7 @@ benchmark:
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
     AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AGENTIC_WARMUP_GRACE_PERIOD: "3600"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
     WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index bd5e3185f6..6b17a8d61d 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8073,6 +8073,7 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           dp-attn: false
       # Throughput oriented: official multi_node_dep strategy, DEP16 across
       # four GB200 nodes (TP4 x DP4 = EP16, one local DP rank per node).
+      # The recipe allows up to 3600s for full-context saturation warmup to drain.
       # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=gb200&nodes=4&strategy=multi_node_dep
       - spec-decoding: none
         conc-list: [32, 64, 96, 128, 192, 256]

From 1187a301fa47df1c6a82e542b6dd58da088fea21 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 28 Jul 2026 17:52:29 -0500
Subject: [PATCH 21/21] chore(runners): register fourth GB200 worker
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add gb200-nv_3 to the GB200 cluster runner pool so the Kimi K3 sweep can use the capacity provisioned for this bring-up.

中文：将 gb200-nv_3 注册到 GB200 集群运行器池，使 Kimi K3 扫描能够使用本次适配新增的容量。
---
 configs/runners.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/configs/runners.yaml b/configs/runners.yaml
index 851b821ba2..bb0156ab9a 100644
--- a/configs/runners.yaml
+++ b/configs/runners.yaml
@@ -247,6 +247,7 @@ labels:
     - gb200-nv_0
     - gb200-nv_1
     - gb200-nv_2
+    - gb200-nv_3
   cluster:gb300-nv:
     - gb300-nv_0
     - gb300-nv_1