SemiAnalysisAI · ivanium · Aug 10, 2026
diff --git a/.github/workflows/benchmark-multinode-tmpl.yml b/.github/workflows/benchmark-multinode-tmpl.yml
@@ -6,6 +6,11 @@ on:
       runner:
         required: true
         type: string
+      runner-node:
+        description: "Optional concrete self-hosted runner label used only for scheduling"
+        required: false
+        type: string
+        default: ''
       priority:
         description: "Higher-is-sooner CI priority score"
         required: true
@@ -269,19 +274,19 @@ jobs:
           inputs.skip-queue-pr != '' &&
           format(
             '["self-hosted",{0},{1},{2},{3}]',
-            toJSON(inputs.runner),
+            toJSON(inputs.runner-node != '' && inputs.runner-node || inputs.runner),
             toJSON(format('ci-job-{0}-{1}', inputs.priority, inputs.queue-token)),
             toJSON(format('ci-attempt-{0}', github.run_attempt)),
             toJSON(format('ci-skip-queue-pr-{0}', inputs.skip-queue-pr))
           ) ||
           format(
             '["self-hosted",{0},{1},{2}]',
-            toJSON(inputs.runner),
+            toJSON(inputs.runner-node != '' && inputs.runner-node || inputs.runner),
             toJSON(format('ci-job-{0}-{1}', inputs.priority, inputs.queue-token)),
             toJSON(format('ci-attempt-{0}', github.run_attempt))
           )
         ) ||
-        format('[{0}]', toJSON(inputs.runner))
+        format('[{0}]', toJSON(inputs.runner-node != '' && inputs.runner-node || inputs.runner))
       ) }}
     # Full-context AgentX warmup can legitimately exceed the fixed-sequence
     # eight-hour envelope at high concurrency. Keep one hour beyond the

diff --git a/.github/workflows/e2e-tests.yml b/.github/workflows/e2e-tests.yml
@@ -478,6 +478,7 @@ jobs:
             osl: '0'
             max-model-len: '0'
             runner: ${{ matrix.config.runner }}
+            runner-node: ${{ matrix.config['runner-node'] || '' }}
             priority: ${{ matrix.config.priority }}
             queue-token: ${{ matrix.config['queue-token'] }}
             image: ${{ matrix.config.image }}

diff --git a/.github/workflows/run-sweep.yml b/.github/workflows/run-sweep.yml
@@ -652,6 +652,7 @@ jobs:
             osl: '0'
             max-model-len: '0'
             runner: ${{ matrix.config.runner }}
+            runner-node: ${{ matrix.config['runner-node'] || '' }}
             priority: ${{ matrix.config.priority }}
             queue-token: ${{ matrix.config['queue-token'] }}
             skip-queue-pr: ${{ matrix.config['skip-queue-pr'] || '' }}

diff --git a/...arks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml b/...arks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml
@@ -1,8 +1,8 @@
 name: "svf-vllm-agg-gb300-tp4-mtp-agentic"
 
 # GB300 AgentX aggregate topology: one TP4 worker occupies one four-GPU node
-# and serves both prefill and decode at concurrency 4. Scheduler, CUDA-graph,
-# and memory settings match the B300 vLLM TP4 MTP agentic configuration.
+# and serves both prefill and decode at concurrency 8. Size max-num-seqs at
+# 4x concurrency and expand the MTP CUDA-graph envelope to match.
 
 model:
   path: "deepseek-v4-pro"
@@ -77,7 +77,7 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
-    VLLM_V2_WARMUP_MAX_NUM_SEQS: "16"
+    VLLM_V2_WARMUP_MAX_NUM_SEQS: "32"
     TILELANG_CLEANUP_TEMP_FILES: "1"
     VLLM_USE_NCCL_SYMM_MEM: "0"
     TORCH_SYMMMEM: "NVSHMEM"
@@ -110,23 +110,24 @@ backend:
       tensor-parallel-size: 4
       pipeline-parallel-size: 1
       disable-custom-all-reduce: true
-      enable-cumem-allocator: true
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       max-model-len: 1048576
-      max-num-seqs: 16
+      max-num-seqs: 32
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
-      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64],"mode":0}'
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128],"mode":0}'
       speculative-config: '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.49}'
-      gpu-memory-utilization: 0.93
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      gpu-memory-utilization: 0.94
       stream-interval: 10
       no-disable-hybrid-kv-cache-manager: true
       tokenizer-mode: "deepseek_v4"
 
 sbatch_directives:
   cpus-per-task: "72"
+  exclude: "im-gb300-r01-c002,im-gb300-r01-c003"
   mem: "0"
 
 srun_options:
@@ -141,6 +142,8 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    # Avoid concurrent readers observing a mismatched mmap data/index pair.
+    AIPERF_DATASET_MMAP_CACHE_ENABLED: "false"
     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/...s/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-c1-mtp-agentic.yaml b/...s/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-c1-mtp-agentic.yaml
@@ -0,0 +1,150 @@
+name: "svf-vllm-agg-gb300-tp8-c1-mtp-agentic"
+
+# GB300 AgentX aggregate topology: one TP8 worker spans two four-GPU
+# nodes and serves both prefill and decode at concurrency 1. Keep at least
+# 16 sequence slots and size the MTP CUDA-graph envelope from that limit.
+
+model:
+  path: "deepseek-v4-pro"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f"
+  frameworks:
+    dynamo: "1.2.1"
+
+dynamo:
+  wheel: "1.2.1"
+  install: true
+
+environment:
+  DYNAMO_WHEEL_DIRS: "/srtctl-wheels"
+  # The frontend shares Grace CPU capacity with the long TP8 cold start.
+  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: "gb300"
+  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
+    router-temperature: 0.0
+    router-queue-threshold: 65536
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+    tokenizer: "fastokens"
+
+backend:
+  type: vllm
+  connector: null
+  mooncake_kv_store:
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  aggregated_environment:
+    HF_HUB_CACHE: "/hf_hub_cache"
+    HUGGINGFACE_HUB_CACHE: "/hf_hub_cache"
+    TRANSFORMERS_CACHE: "/hf_hub_cache"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_V2_WARMUP_MAX_NUM_SEQS: "16"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    TORCH_SYMMMEM: "NVSHMEM"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_FLASHINFER_ALLREDUCE_BACKEND: "auto"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "20"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
+    VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    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"
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    MC_STORE_CLIENT_METRIC: "1"
+    MC_STORE_CLIENT_METRIC_INTERVAL: "5"
+    MC_TE_METRIC: "0"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-agg-tp8-c1-mtp-{job_id}"
+  vllm_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
+      kv-cache-dtype: "fp8"
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      disable-custom-all-reduce: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 16
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64],"mode":0}'
+      speculative-config: '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.49}'
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      gpu-memory-utilization: 0.94
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: "deepseek_v4"
+
+sbatch_directives:
+  cpus-per-task: "144"
+  exclude: "im-gb300-r01-c002,im-gb300-r01-c003"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    # Keep aggregate workers in the multinode result schema so ingestion uses
+    # the zero decode-worker count instead of duplicating TP into P and D.
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    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/...arks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-mtp-agentic.yaml b/...arks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-mtp-agentic.yaml
@@ -1,7 +1,8 @@
 name: "svf-vllm-agg-gb300-tp8-mtp-agentic"
 
-# Validated GB300 AgentX aggregate topology: one TP8 worker spans two
-# four-GPU nodes and serves both prefill and decode at concurrency 1.
+# GB300 AgentX aggregate topology: one TP8 worker spans two four-GPU
+# nodes and serves both prefill and decode at concurrency 4. Keep at least
+# 16 sequence slots and otherwise size the scheduler at 4x concurrency.
 
 model:
   path: "deepseek-v4-pro"
@@ -77,7 +78,7 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
     VLLM_LOG_STATS_INTERVAL: "1"
-    VLLM_V2_WARMUP_MAX_NUM_SEQS: "32"
+    VLLM_V2_WARMUP_MAX_NUM_SEQS: "16"
     TILELANG_CLEANUP_TEMP_FILES: "1"
     VLLM_USE_NCCL_SYMM_MEM: "0"
     TORCH_SYMMMEM: "NVSHMEM"
@@ -110,24 +111,24 @@ backend:
       tensor-parallel-size: 8
       pipeline-parallel-size: 1
       disable-custom-all-reduce: true
-      enable-cumem-allocator: true
       attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
       max-model-len: 1048576
-      max-num-seqs: 32
+      max-num-seqs: 16
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
-      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
-      max-cudagraph-capture-size: 128
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64],"mode":0}'
       speculative-config: '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":2.49}'
-      gpu-memory-utilization: 0.90
+      moe-backend: "deep_gemm_amxf4_mega_moe"
+      gpu-memory-utilization: 0.94
       stream-interval: 10
       no-disable-hybrid-kv-cache-manager: true
       tokenizer-mode: "deepseek_v4"
 
 sbatch_directives:
   cpus-per-task: "144"
+  exclude: "im-gb300-r01-c002,im-gb300-r01-c003"
   mem: "0"
 
 srun_options:

diff --git a/...b300-1p1d-dep4-dep8-c128-mtp-agentic.yaml → ...b300-1p1d-dep4-dep8-c256-mtp-agentic.yaml b/...b300-1p1d-dep4-dep8-c128-mtp-agentic.yaml → ...b300-1p1d-dep4-dep8-c256-mtp-agentic.yaml
@@ -1,7 +1,7 @@
-name: "svf-vllm-disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic"
+name: "svf-vllm-disagg-gb300-1p1d-dep4-dep8-c256-mtp-agentic"
 
-# Validated GB300 AgentX MTP3 low-latency topology: one DEP4 prefill worker
-# feeds one DEP8 decode worker at concurrency 128.
+# GB300 AgentX MTP3 topology: one DEP4 prefill worker feeds one
+# DEP8 decode worker at concurrency 256.
 
 model:
   path: "deepseek-v4-pro"
@@ -17,7 +17,7 @@ identity:
     dynamo: "1.3.0.dev20260720"
 
 dynamo:
-  version: "1.3.0.dev20260720"
+  wheel: "1.3.0.dev20260720"
   install: true
 
 setup_script: vllm-container-deps.sh
@@ -57,6 +57,8 @@ frontend:
     router-session-affinity-ttl-secs: 900
   env:
     DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TCP_CHANNEL_BUFFER: "128"
+    DYN_TCP_REQUEST_TIMEOUT: "60"
 
 backend:
   type: vllm
@@ -65,7 +67,7 @@ backend:
   mooncake_kv_store:
     store_config:
       metadata_server: "P2PHANDSHAKE"
-      global_segment_size: "150GB"
+      global_segment_size: "180GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
       device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
@@ -86,6 +88,7 @@ backend:
       max-model-len: 1048576
       max-num-seqs: 64
       max-num-batched-tokens: 8192
+      long-prefill-token-threshold: 1024
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
@@ -149,7 +152,7 @@ backend:
     VLLM_USE_BREAKABLE_CUDAGRAPH: "0"
     VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
     VLLM_DSV4_MEGA_FP8_COMBINE: "1"
-    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-c128-mtp-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep4-dep8-c256-{job_id}"
     MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_STORE_CLIENT_METRIC: "1"
     MC_STORE_CLIENT_METRIC_INTERVAL: "5"
@@ -176,14 +179,15 @@ backend:
     VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
     VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
     VLLM_DSV4_MEGA_FP8_COMBINE: "1"
-    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-c128-mtp-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep4-dep8-c256-{job_id}"
     MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_STORE_CLIENT_METRIC: "1"
     MC_STORE_CLIENT_METRIC_INTERVAL: "5"
     MC_TE_METRIC: "0"
 
 sbatch_directives:
   cpus-per-task: "72"
+  exclude: "im-gb300-r01-c002,im-gb300-r01-c003"
 
 srun_options:
   container-remap-root: ""
@@ -198,6 +202,8 @@ benchmark:
     IS_MULTINODE: "true"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
     AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    # Avoid concurrent readers observing a mismatched mmap data/index pair.
+    AIPERF_DATASET_MMAP_CACHE_ENABLED: "false"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
     WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"