[AgentX B200] enable Kimi-K3 power and correct DCP metadata / 启用功耗并修正元数据

.github/workflows/test-process-result.yml

@@ -25,6 +25,8 @@ on:
       - 'runners/launch_b300-dsxe.sh'
       - 'runners/launch_h200-dgxc-slurm.sh'
       - 'runners/inject_srt_power_concurrencies.py'
+      - 'runners/slurm_utils.sh'
+      - 'runners/test_kimik3_bh_power.py'
       - 'utils/aggregate_power.py'
       - 'utils/aggregate_power_multinode.py'
       - 'utils/agentic/aggregation/**'
@@ -57,4 +59,4 @@ jobs:
         run: |
           cd utils
           uv run --no-project --exclude-newer PT12H --python 3.12 --with pytest --with pyyaml \
-            python -m pytest test_aggregate_power.py test_aggregate_power_multinode.py agentic/aggregation/ test_gb300_power_official_contract.py test_inject_srt_power_concurrencies.py test_process_result.py -v
+            python -m pytest test_aggregate_power.py test_aggregate_power_multinode.py agentic/aggregation/ test_gb300_power_official_contract.py test_inject_srt_power_concurrencies.py test_process_result.py ../runners/test_kimik3_bh_power.py -v

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c1-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [1]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c14-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [14]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c24-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [24]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c4-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [4]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c48-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [48]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c8-agentic.yaml

@@ -111,14 +111,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [8]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c96-agentic.yaml

@@ -110,14 +110,27 @@ sbatch_directives:
 srun_options:
   container-remap-root: ""
 
+telemetry:
+  enabled: true
+  provider: dcgm-power
+  default_frequency: 1.0
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+
 benchmark:
   type: custom
+  concurrencies: [96]
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
     AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

configs/nvidia-master.yaml

@@ -10014,103 +10014,103 @@ kimik3-fp4-b200-dynamo-vllm-agentic-dspark:
     - dram-utilization: 0.8
       search-space:
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [1]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c1-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.84"
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [4]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c4-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.84"
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [8]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c8-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.84"
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [14]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c14-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.84"
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [24]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c24-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.36"
       - spec-decoding: mtp
-        kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+        kv-offloading: none
         conc-list: [48]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:
           - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8pp2-mooncake-c48-agentic.yaml"
           - "SYNTHETIC_ACCEPTANCE=true"
           - "SYNTHETIC_ACCEPTANCE_LENGTH=3.36"
-      - kv-offloading: dram
-        kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }
+      - kv-offloading: none
         conc-list: [96]
         num-nodes: 2
         worker:
           num-worker: 1
           tp: 8
           pp: 2
+          dcp-size: 8
           ep: 1
           dp-attn: false
           additional-settings:

perf-changelog.yaml

@@ -7399,3 +7399,12 @@
     - "Disable adaptive verification in the EVAL_ONLY DSpark config as well. It trims verification requests on device, which the ROCm DeepseekV4IndexerBackend does not support, so the eval-only engine refused to start (run 34651830283, c32). Evals keep real block rejection; throughput settings are unchanged."
     - "EVAL_ONLY 的 DSpark 配置同样关闭自适应验证：它会在设备端裁剪验证请求，而 ROCm 的 DeepseekV4IndexerBackend 不支持该操作，导致仅评测引擎拒绝启动（运行 34651830283，c32）。评测仍保留真实块拒绝采样；吞吐设置不变。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2962
+
+- config-keys:
+    - kimik3-fp4-b200-dynamo-vllm-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Enable required power for the seven existing B200 Kimi-K3 recipes, use the pinned AgentX producer and shared collector, fix the exporter import URI, and record DCP8 with offload disabled to match the serving commands."
+    - "启用 Kimi-K3 B200 实测功耗并修正 DCP 与 Offload 元数据。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3043