[WIP] Test DeepSeek-V4 GB300 Dynamo AgentX recipes / 测试 DeepSeek-V4 GB300 Dynamo AgentX 配置

.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

.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 }}

.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'] || '' }}

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp4-mtp-agentic.yaml

@@ -1,7 +1,7 @@
 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 serves both prefill and decode at concurrency 8. Scheduler, CUDA-graph,
 # and memory settings match the B300 vLLM TP4 MTP agentic configuration.
 
 model:
@@ -114,7 +114,7 @@ backend:
       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
+      max-num-batched-tokens: 32768
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
@@ -127,6 +127,7 @@ backend:
 
 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"

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb300-tp8-mtp-agentic.yaml

@@ -1,7 +1,7 @@
 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 1 and 4.
 
 model:
   path: "deepseek-v4-pro"
@@ -114,7 +114,7 @@ backend:
       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-batched-tokens: 8192
+      max-num-batched-tokens: 32768
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
@@ -128,6 +128,7 @@ backend:
 
 sbatch_directives:
   cpus-per-task: "144"
+  exclude: "im-gb300-r01-c002,im-gb300-r01-c003"
   mem: "0"
 
 srun_options:

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep8-c128-mtp-agentic.yaml → benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep4-dep12-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-dep12-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
+# DEP12 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
@@ -35,11 +35,11 @@ resources:
   het_jobs: false
   spread_workers: false
   prefill_nodes: 1
-  decode_nodes: 2
+  decode_nodes: 3
   prefill_workers: 1
   decode_workers: 1
   gpus_per_prefill: 4
-  gpus_per_decode: 8
+  gpus_per_decode: 12
 
 infra:
   etcd_nats_dedicated_node: false
@@ -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"
@@ -104,7 +106,7 @@ backend:
       kv-cache-dtype: "fp8"
       tensor-parallel-size: 1
       pipeline-parallel-size: 1
-      data-parallel-size: 8
+      data-parallel-size: 12
       data-parallel-rpc-port: 13345
       enable-cumem-allocator: true
       enable-expert-parallel: true
@@ -149,7 +151,7 @@ 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-dep12-c256-{job_id}"
     MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_STORE_CLIENT_METRIC: "1"
     MC_STORE_CLIENT_METRIC_INTERVAL: "5"
@@ -176,14 +178,15 @@ 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-dep12-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 +201,8 @@ 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"

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic.yaml → benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb300-1p1d-dep8-dep12-c576-mtp-agentic.yaml

@@ -1,7 +1,7 @@
-name: "svf-vllm-disagg-gb300-1p1d-dep8-dep8-c384-mtp-agentic"
+name: "svf-vllm-disagg-gb300-1p1d-dep8-dep12-c576-mtp-agentic"
 
-# Validated GB300 AgentX MTP3 efficiency topology: one DEP8 prefill worker and
-# one DEP8 decode worker at concurrency 384.
+# GB300 AgentX MTP3 topology: one DEP8 prefill worker feeds one
+# DEP12 decode worker at concurrency 576.
 
 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
@@ -35,11 +35,11 @@ resources:
   het_jobs: false
   spread_workers: false
   prefill_nodes: 2
-  decode_nodes: 2
+  decode_nodes: 3
   prefill_workers: 1
   decode_workers: 1
   gpus_per_prefill: 8
-  gpus_per_decode: 8
+  gpus_per_decode: 12
 
 infra:
   etcd_nats_dedicated_node: false
@@ -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"
@@ -104,7 +106,7 @@ backend:
       kv-cache-dtype: "fp8"
       tensor-parallel-size: 1
       pipeline-parallel-size: 1
-      data-parallel-size: 8
+      data-parallel-size: 12
       data-parallel-rpc-port: 13345
       enable-cumem-allocator: true
       enable-expert-parallel: true
@@ -149,7 +151,7 @@ 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-mtp-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep8-dep12-c576-{job_id}"
     MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_STORE_CLIENT_METRIC: "1"
     MC_STORE_CLIENT_METRIC_INTERVAL: "5"
@@ -176,14 +178,15 @@ 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-mtp-{job_id}"
+    DG_JIT_CACHE_DIR: "/tmp/dg-cache-dsv4-gb300-1p1d-dep8-dep12-c576-{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"
   mem: "0"
 
 srun_options:
@@ -199,6 +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"