Add DSV4 B300 Dynamo vLLM disagg

.github/configs/nvidia-master.yaml

@@ -8178,6 +8178,62 @@ dsv4-fp4-gb200-dynamo-vllm-mtp2:
           ep: 8
           dp-attn: true
 
+dsv4-fp4-b300-dynamo-vllm:
+  image: vllm/vllm-openai:v0.20.1
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b300
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # B300 adaptation of the DSV4 B200/GB200 vLLM disagg recipes. Each
+      # prefill/decode worker maps to one full 8-GPU B300 node.
+      - conc-list: [1, 64, 128]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+      - conc-list: [128, 256, 512, 1024]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - conc-list: [4096, 8192]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 dsv4-fp4-gb300-dynamo-vllm:
   image: vllm/vllm-openai:v0.20.0-ubuntu2404
   model: deepseek-ai/DeepSeek-V4-Pro

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml

@@ -0,0 +1,130 @@
+name: "svf-vllm-disagg-b300-high-tpt-megamoe"
+
+# B300 adaptation of the DSV4 GB200/B200 vLLM disagg recipe. Each worker uses
+# one full 8-GPU B300 node, plus a dedicated NATS/etcd infra node.
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+  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
+      enable-ep-weight-filter: true
+      attention-config: '{"use_fp4_indexer_cache": true}'
+      enforce-eager: true
+      max-model-len: 9280
+      max-num-seqs: 16
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 32768
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      gpu-memory-utilization: 0.85
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: true
+      tokenizer-mode: deepseek_v4
+      reasoning-parser: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      attention-config: '{"use_fp4_indexer_cache": true}'
+      max-model-len: 9280
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      gpu-memory-utilization: 0.85
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+      reasoning-parser: deepseek_v4
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096x8192"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+    revision: "0366e4e064385807ea86b088a5c6c878ff23343b"
+  container:
+    image: "vllm/vllm-openai:v0.20.1"
+  frameworks:
+    dynamo: "1.2.0.dev20260426"
+    vllm: "0.20.1"

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml

@@ -0,0 +1,124 @@
+name: "svf-vllm-disagg-b300-low-latency"
+
+# B300 adaptation of the DSV4 GB200/B200 vLLM disagg recipe. Each worker uses
+# one full 8-GPU B300 node, plus a dedicated NATS/etcd infra node.
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:v0.20.1"
+  precision: "fp4"
+
+dynamo:
+  install: true
+  wheel: "1.2.0.dev20260426"
+
+setup_script: vllm-container-deps.sh
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+  decode_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+  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
+      attention-config: '{"use_fp4_indexer_cache": true}'
+      enforce-eager: true
+      max-model-len: 16384
+      max-num-seqs: 16
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 32768
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      no-async-scheduling: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      gpu-memory-utilization: 0.8
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: true
+      offload-group-size: 3
+      offload-num-in-group: 1
+      offload-prefetch-step: 2
+      tokenizer-mode: deepseek_v4
+      reasoning-parser: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      max-model-len: 16384
+      max-num-seqs: 256
+      max-cudagraph-capture-size: 2048
+      max-num-batched-tokens: 256
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+      reasoning-parser: deepseek_v4
+
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1x64x128"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+identity:
+  container:
+    image: "vllm/vllm-openai:v0.20.1"
+  frameworks:
+    dynamo: "1.2.0.dev20260426"
+    vllm: "0.20.1"