diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index ad4220266..46f682385 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-high-tpt-megamoe.yaml
new file mode 100644
index 000000000..6d754905b
--- /dev/null
+++ b/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"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-latency.yaml
new file mode 100644
index 000000000..4f5cf12bd
--- /dev/null
+++ b/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"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-middle-curve.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-middle-curve.yaml
new file mode 100644
index 000000000..4cb7e15dd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-low-middle-curve.yaml
@@ -0,0 +1,125 @@
+name: "svf-vllm-disagg-b300-low-middle-curve"
+
+# 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: 4
+  prefill_workers: 1
+  decode_workers: 4
+  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: 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: "256x512"
+  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"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-max-tpt-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-max-tpt-megamoe.yaml
new file mode 100644
index 000000000..47d98e9af
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-max-tpt-megamoe.yaml
@@ -0,0 +1,130 @@
+name: "svf-vllm-disagg-b300-max-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: 3
+  decode_nodes: 1
+  prefill_workers: 3
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+
+backend:
+  type: vllm
+  connector: null
+  prefill_environment:
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    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: "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: "0.20.1"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml
new file mode 100644
index 000000000..30cfdacd8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b300-mid-curve-megamoe.yaml
@@ -0,0 +1,130 @@
+name: "svf-vllm-disagg-b300-mid-curve-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: 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
+      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: "128x256x512x1024"
+  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"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 12a5e4da5..bba8a15f5 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -2316,3 +2316,10 @@
     - "Adapt the existing DSV4 GB200 vLLM disagg recipes to B200 by mapping each worker to one full 8-GPU B200 node"
     - "Update the B200 DGXC Slurm launcher to support dsv4/fp4 with dynamo-vllm and local recipe overlays"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1303
+
+- config-keys:
+    - dsv4-fp4-b300-dynamo-vllm
+  description:
+    - "Turn off explicit deep_gemm_mega_moe backend selection in B300 Dynamo vLLM throughput recipes"
+    - "Let vLLM choose the MoE backend automatically to avoid CUDA symmetric-memory rendezvous failures during startup"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1304
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index e81bf91a3..0bf203022 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -9,8 +9,8 @@ set -x
 if [[ "$IS_MULTINODE" == "true" ]]; then
 
 # Validate framework
-if [[ $FRAMEWORK != "dynamo-sglang" && $FRAMEWORK != "dynamo-trt" ]]; then
-    echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang"
+if [[ $FRAMEWORK != "dynamo-sglang" && $FRAMEWORK != "dynamo-trt" && $FRAMEWORK != "dynamo-vllm" ]]; then
+    echo "Unsupported framework: $FRAMEWORK. Supported frameworks are: dynamo-trt, dynamo-sglang, dynamo-vllm"
     exit 1
 fi
 
@@ -25,8 +25,22 @@ elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH="/data/models/dsr1-fp8"
     export SERVED_MODEL_NAME="deepseek-r1-fp8"
     export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
+elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $FRAMEWORK == "dynamo-vllm" ]]; then
+    SELECTED_MODEL_PATH=""
+    if [[ -n "${MODEL_PATH:-}" && -d "${MODEL_PATH}" ]]; then
+        SELECTED_MODEL_PATH="$MODEL_PATH"
+    else
+        for candidate in /data/models/dsv4-pro /data/models/deepseek-v4-pro /data/models/DeepSeek-V4-Pro; do
+            if [[ -d "$candidate" ]]; then
+                SELECTED_MODEL_PATH="$candidate"
+                break
+            fi
+        done
+    fi
+    export MODEL_PATH="${SELECTED_MODEL_PATH:-/data/models/dsv4-pro}"
+    export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm"
     exit 1
 fi
 
@@ -41,6 +55,12 @@ fi
 if [[ "$IS_AGENTIC" == "1" ]]; then
     git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
+elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR" || exit 1
+    git checkout aflowers/vllm-gb200-v0.20.0
+    mkdir -p recipes/vllm/deepseek-v4
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
 else
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR" || exit 1
@@ -96,6 +116,8 @@ model_paths:
 containers:
   dynamo-trtllm: "${SQUASH_FILE}"
   dynamo-sglang: "${SQUASH_FILE}"
+  dynamo-vllm: "${SQUASH_FILE}"
+  "${IMAGE}": "${SQUASH_FILE}"
   nginx-sqsh: "${NGINX_SQUASH_FILE}"
 use_exclusive_sbatch_directive: true
 default_mounts: