Day 0 GB300 DeepSeek-V4-Pro FP4 vLLM disagg

.github/configs/nvidia-master.yaml

@@ -7666,3 +7666,102 @@ dsv4-fp4-gb200-dynamo-vllm:
         tp: 16
         ep: 16
         dp-attn: true
+
+dsv4-fp4-gb300-dynamo-vllm:
+  image: vllm/vllm-openai:deepseekv4-cu130
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: gb300-cw
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  # Same topology + tuning as dsv4-fp4-gb300-dynamo-vllm's gb200 sibling, just
+  # pointed at the gb300 recipe variants. Cluster gb300-cw is 2x 18-node
+  # racks; each job is rack-pinned via srtctl's auto `#SBATCH --segment={N}`.
+  seq-len-configs:
+  - isl: 1024
+    osl: 1024
+    search-space:
+    - conc-list: [1, 4, 8, 16, 32, 64]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/1k1k/disagg-gb300-1p1d-dep8-tep8.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+    - conc-list: [128, 256, 1024, 2048, 4096]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/1k1k/disagg-gb300-1p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true
+    - conc-list: [4096, 8192]
+      prefill:
+        num-worker: 3
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/1k1k/disagg-gb300-3p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true
+
+  - isl: 8192
+    osl: 1024
+    search-space:
+    - conc-list: [1, 4, 8, 16, 32, 64]
+      prefill:
+        num-worker: 1
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-gb300-1p1d-dep8-tep8.yaml"
+      decode:
+        num-worker: 1
+        tp: 8
+        ep: 1
+        dp-attn: false
+    - conc-list: [512, 1024]
+      prefill:
+        num-worker: 3
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-gb300-3p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true
+    - conc-list: [4096, 8192]
+      prefill:
+        num-worker: 7
+        tp: 8
+        ep: 8
+        dp-attn: true
+        additional-settings:
+        - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-gb300-7p1d-dep8-dep16.yaml"
+      decode:
+        num-worker: 1
+        tp: 16
+        ep: 16
+        dp-attn: true

.github/configs/runners.yaml

@@ -131,3 +131,8 @@ gb300:
 - 'gb300-nv_0'
 - 'gb300-nv_1'
 - 'gb300-nv_2'
+gb300-cw:
+- 'gb300-cw_0'
+- 'gb300-cw_1'
+- 'gb300-cw_2'
+- 'gb300-cw_3'

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/1k1k/disagg-gb300-1p1d-dep8-dep16.yaml

@@ -0,0 +1,139 @@
+name: "dsv4-vllm-disagg-gb300-1p1d-dep8-dep16"
+
+# GB300 mirror of disagg-gb200-1p1d-dep8-dep16.yaml. Same tuning at FP4
+# (288 GB HBM/GPU on GB300 vs 184 GB on GB200 — extra headroom for KV).
+# Cluster: gb300-cw (2x 18-node racks); job pins to one rack via the
+# explicit sbatch_directives.segment="6" below (cw's srtslurm.yaml turns
+# off srtctl's auto-segment so each recipe owns its segment value).
+#
+# 1k/1k mid-to-high throughput topology. Single prefill worker feeding a
+# wide DP=16 decode handles conc 256-4096 cleanly for 1k prompts.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:deepseekv4-cu130"
+  precision: "fp4"
+
+dynamo:
+  hash: 6a159fedd8e4a1563aa647c31f622aedbf254b5b
+  # Install handled by our custom vllm-container-deps.sh, which builds
+  # the dynamo wheel ONCE on /mnt/vast and lets every rank pip-install
+  # from cache. See runners/gb300-cw-vllm-container-deps.sh.
+  install: false
+
+setup_script: vllm-container-deps.sh
+
+# Mount /mnt/vast/dynamo_cache into every worker container so each
+# rank can pip-install from the wheel that launch_gb300-cw.sh
+# pre-built there. Without this only /mnt/vast/models/<model> is
+# in scope and our setup script errors out with 'prebuilt cache
+# missing'.
+extra_mount:
+  - "/mnt/vast/dynamo_cache:/mnt/vast/dynamo_cache"
+
+# Pin all 6 nodes to the same rack on cw.
+sbatch_directives:
+  segment: "6"
+  # Use all node memory; cw default was too tight.
+  mem: "0"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 16
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+
+  prefill_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    CARGO_BUILD_JOBS: "4"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    NCCL_DEBUG: "INFO"
+    NCCL_DEBUG_SUBSYS: "INIT,BOOTSTRAP,ENV,NET,GRAPH,NVLS"
+    VLLM_SERVER_DEV_MODE: "1"
+
+  decode_environment:
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    CARGO_BUILD_JOBS: "4"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    NCCL_DEBUG: "INFO"
+    NCCL_DEBUG_SUBSYS: "INIT,BOOTSTRAP,ENV,NET,GRAPH,NVLS"
+    VLLM_SERVER_DEV_MODE: "1"
+
+  vllm_config:
+    prefill:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enforce-eager: true
+      max-model-len: 3072
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      gpu-memory-utilization: 0.88
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 16
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      max-model-len: 3072
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+
+benchmark:
+  type: "sa-bench"
+  isl: 1024
+  osl: 1024
+  concurrencies: "128x256x1024x2048x4096"
+  req_rate: "inf"
+  use_chat_template: false