Add DSV4 B200 Dynamo vLLM disagg

.github/configs/nvidia-master.yaml

@@ -7942,6 +7942,75 @@ kimik2.5-fp4-gb200-dynamo-vllm:
           ep: 16
           dp-attn: true
 
+dsv4-fp4-b200-dynamo-vllm:
+  image: vllm/vllm-openai:v0.20.1
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: b200-multinode
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # B200 adaptation of the DSV4 GB200 vLLM disagg recipes. Each worker
+      # maps to one full 8-GPU B200 node.
+      - conc-list: [1,64]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+      - conc-list: [512]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-curve.yaml"
+        decode:
+          num-worker: 4
+          tp: 8
+          ep: 1
+          dp-attn: false
+      - conc-list: [1024, 2048, 4096]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-high-tpt-megamoe.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - conc-list: [8192]
+        prefill:
+          num-worker: 3
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/8k1k/disagg-b200-max-tpt-megamoe.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 dsv4-fp4-gb200-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-b200-high-tpt-megamoe.yaml

@@ -0,0 +1,138 @@
+name: "svf-vllm-disagg-b200-high-tpt-megamoe"
+
+# Mirrored from NVIDIA/srt-slurm aflowers/vllm-gb200-v0.20.0 branch:
+#   recipes/vllm/deepseek-v4-pro/GB200/8k1k/disagg-gb200-high-tpt-megamoe.yaml
+#
+# B200 adaptation of the GB200 recipe below. Each prefill/decode worker uses
+# one full 8-GPU B200 node, plus a dedicated NATS/etcd infra node.
+#
+# Local deltas vs upstream:
+#   * model.path alias renamed deepseekv4-fp4 -> deepseek-v4-pro to match
+#     SRT_SLURM_MODEL_PREFIX in the launch script.
+#   * model.container set to vllm/vllm-openai:v0.20.1 to
+#     match nvidia-master.yaml image (which the launch script registers as
+#     the alias key in srtslurm.yaml). Upstream variants ship either the
+#     non-dynamo floating tag or a sha256 pin.
+#   * slurm.time_limit + health_check set to 8h / 1440 attempts to
+#     absorb cold-cache model loads.
+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: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 2
+  decode_nodes: 1
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    # VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    # VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    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"
+    # VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    # VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    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
+      enforce-eager: true
+      max-model-len: 9280
+      max-num-seqs: 16
+      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
+      gpu-memory-utilization: 0.95
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      numa-bind: true
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      data-parallel-rpc-port: 13345
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      max-model-len: 9280
+      max-num-seqs: 512
+      max-cudagraph-capture-size: 512
+      max-num-batched-tokens: 512
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4096"
+  req_rate: "inf"
+  use_chat_template: true
+  custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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: "0.20.0"

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency.yaml

@@ -0,0 +1,137 @@
+name: "svf-vllm-disagg-b200-low-latency"
+
+# Mirrored from NVIDIA/srt-slurm aflowers/vllm-gb200-v0.20.0 branch:
+#   recipes/vllm/deepseek-v4-pro/GB200/8k1k/disagg-gb200-low-latency.yaml
+#
+# B200 adaptation of the GB200 recipe below. Each prefill/decode worker uses
+# one full 8-GPU B200 node, plus a dedicated NATS/etcd infra node.
+#
+# Local deltas vs upstream:
+#   * model.path alias renamed deepseekv4-fp4 -> deepseek-v4-pro to match
+#     SRT_SLURM_MODEL_PREFIX in the launch script.
+#   * model.container set to vllm/vllm-openai:v0.20.1 to
+#     match nvidia-master.yaml image (which the launch script registers as
+#     the alias key in srtslurm.yaml). Upstream variants ship either the
+#     non-dynamo floating tag or a sha256 pin.
+#   * slurm.time_limit + health_check set to 8h / 1440 attempts to
+#     absorb cold-cache model loads.
+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: "b200"
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
+    # VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    # VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    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"
+    # VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    # VLLM_MOE_ROUTING_SIMULATION_STRATEGY: "uniform_random"
+    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
+      enforce-eager: true
+      max-model-len: 16384
+      max-num-seqs: 16
+      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
+      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
+      # offload-params: "w13_weight w2_weight w13_weight_scale w2_weight_scale wq_b wo_a wo_b shared_experts"
+      tokenizer-mode: deepseek_v4
+    decode:
+      kv-transfer-config: '{"kv_connector": "NixlConnector", "kv_role": "kv_both"}'
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+#      data-parallel-size: 8
+#      data-parallel-rpc-port: 13345
+#      enable-expert-parallel: true
+      max-model-len: 16384
+      max-num-seqs: 256
+      max-cudagraph-capture-size: 256
+      max-num-batched-tokens: 256
+      trust-remote-code: true
+      no-enable-prefix-caching: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      gpu-memory-utilization: 0.9
+      stream-interval: 50
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1"
+  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: "0.20.0"