[Power] feat: extend dcgm energy lanes to gb dsv4/qwen3.5 fp4 / 扩展 dcgm 能耗采集到 gb dsv4 与 qwen3.5 fp4

.github/AGENT_OPERATIONS.md

@@ -96,7 +96,7 @@ Multinode disaggregated results add `prefill_gpu_energy_j`, `decode_gpu_energy_j
 
 Every power result — valid or invalid, single-node or multinode — carries `power_metric_schema_version`. Version 2 defines each unprefixed `joules_per_*` field as whole-deployment GPU-board energy over the named denominator; role-scoped energy uses the explicit `prefill_*` / `decode_*` keys. Rows without the field predate the whole-deployment switch and their unprefixed joules are not comparable across topologies.
 
-For srt-slurm recipes, `telemetry: {provider: dcgm-power}` enables official energy collection. `runners/launch_gb200-nv.sh` and `runners/launch_gb300-nv.sh` are the source of truth for `POWER_SRT_SLURM_PIN`. CI derives `POWER_PRODUCER_SHA` from the launcher stamp. `utils/test_gb200_power_official_contract.py` and `utils/test_gb300_power_official_contract.py` enforce the recipe/launcher contract. Only `PRECISION=fp8` dcgm-power lanes are validated.
+For srt-slurm recipes, `telemetry: {provider: dcgm-power}` enables official energy collection. `runners/launch_gb200-nv.sh` and `runners/launch_gb300-nv.sh` are the source of truth for `POWER_SRT_SLURM_PIN`. CI derives `POWER_PRODUCER_SHA` from the launcher stamp. `utils/test_gb200_power_official_contract.py` and `utils/test_gb300_power_official_contract.py` enforce the recipe/launcher contract. Eligible recipe-gated `dynamo-sglang` dcgm-power lanes are validated.
 
 Power audit artifacts are named `power_audit_<result>` and contain `power_validation_<result>.json` for single-node runs or `power_validation_<result>_*.json` for multinode runs. They are uploaded even when validation fails.
 

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml

@@ -154,3 +154,16 @@ benchmark:
   concurrencies: "512"
   req_rate: "inf"
   use_chat_template: false
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml

@@ -115,3 +115,16 @@ benchmark:
   concurrencies: "1"
   req_rate: "inf"
   use_chat_template: false
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml

@@ -154,3 +154,16 @@ benchmark:
   concurrencies: "256"
   req_rate: "inf"
   use_chat_template: false
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-10p1d-dep4-dep32-18-c2500.yaml

@@ -6,7 +6,7 @@ name: "disagg-gb300-10p1d-dep4-dep32-18-c2500"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-12p1d-dep4-dep24-18-c3000.yaml

@@ -6,7 +6,7 @@ name: "disagg-gb300-12p1d-dep4-dep24-18-c3000"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-14p1d-dep4-dep16-18-c8192.yaml

@@ -6,7 +6,7 @@ name: "disagg-gb300-14p1d-dep4-dep16-18-c8192"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-15p1d-dep4-dep12-18-c12000.yaml

@@ -6,7 +6,7 @@ name: "disagg-gb300-15p1d-dep4-dep12-18-c12000"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml

@@ -33,7 +33,7 @@ name: "disagg-gb300-1p1d-dep4-dep16-5-c1024"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:
@@ -177,3 +177,18 @@ benchmark:
   req_rate: "inf"
   use_chat_template: false
   custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    # 9401 is already bound by the cluster-level exporter on im-gb300 nodes;
+    # use a port outside that range.
+    port: 19401

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml

@@ -33,7 +33,7 @@ name: "disagg-gb300-1p1d-tp4-tp4-2-c1"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 # See ../1k1k/disagg-gb200-1p1d-dep8-tep8.yaml for the dynamo pin
@@ -160,3 +160,18 @@ benchmark:
   req_rate: "inf"
   use_chat_template: false
   custom_tokenizer: "sa_bench_tokenizers.sglang_deepseek_v4.SGLangDeepseekV4Tokenizer"
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    # 9401 is already bound by the cluster-level exporter on im-gb300 nodes;
+    # use a port outside that range.
+    port: 19401

benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-8p1d-dep4-dep40-18-c2048.yaml

@@ -6,7 +6,7 @@ name: "disagg-gb300-8p1d-dep4-dep40-18-c2048"
 
 model:
   path: "deepseek-v4-pro"
-  container: "lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233"
+  container: "lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3"
   precision: "fp4"
 
 dynamo:

benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml

@@ -173,3 +173,18 @@ benchmark:
   concurrencies: "1x4x8x16x32x64x256"
   req_rate: "inf"
   random_range_ratio: 0.8
+
+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: 10
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    # 9401 is already bound by the cluster-level exporter on im-gb300 nodes;
+    # use a port outside that range.
+    port: 19401

configs/nvidia-master.yaml

@@ -5875,7 +5875,7 @@ dsv4-fp4-gb300-dynamo-trt-mtp:
           dp-attn: true
 
 dsv4-fp4-gb300-dynamo-sglang:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
+  image: lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: gb300

perf-changelog.yaml

@@ -5926,6 +5926,18 @@
     - "Rides on the NVFP4-V2 checkpoint switch from #2205"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2550
 
+- config-keys:
+    - dsv4-fp4-gb200-dynamo-sglang
+    - dsv4-fp4-gb300-dynamo-sglang
+    - qwen3.5-fp4-gb300-dynamo-sglang
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Extend official DCGM power collection to the recipe-gated FP4 GB200 and GB300 Dynamo-SGLang lanes."
+    - "Synchronize all DSV4 GB300 recipes on lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3."
+    - "Pin the srt-slurm power producer and record its exact checkout SHA for strict provenance validation."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2507
+
 - config-keys:
     - qwen3.5-fp8-b300-sglang-agentic-power-ab
     - qwen3.5-fp4-b300-sglang-agentic-power-ab

runners/launch_gb200-nv.sh

@@ -302,11 +302,12 @@ if [[ -n "$CONFIG_FILE" && -f "$_RECIPE_SRC" ]] && awk '
     USES_DCGM_POWER=1
 fi
 
-# Note (wenyao): the producer pin descends from the fp8 srt-slurm lineage
-# (cargo/maturin bootstrap); a non-fp8 power recipe would silently clone the
-# wrong lineage, so fail fast instead.
-if [[ "$USES_DCGM_POWER" == "1" && "$PRECISION" != "fp8" ]]; then
-    echo "Error: dcgm-power lanes are only validated for PRECISION=fp8, got: $PRECISION" >&2
+# Note (wenyao): the producer pin follows the srt-slurm main lineage that the
+# dynamo-sglang lanes run on (fp8 validated end-to-end, fp4 recipes
+# parse-verified against the pin); other frameworks clone diverging refs
+# (aflowers branch, sa-submission), so fail fast for them instead.
+if [[ "$USES_DCGM_POWER" == "1" && "$FRAMEWORK" != "dynamo-sglang" ]]; then
+    echo "Error: dcgm-power lanes are only validated for FRAMEWORK=dynamo-sglang, got: $FRAMEWORK" >&2
     exit 1
 fi
 
@@ -475,10 +476,25 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     mkdir -p recipes/vllm/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
-    # Stay on NVIDIA/srt-slurm:main (default) — submission branch no
-    # longer needed; overlay our hand-rolled DSV4 sglang recipes onto it.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
+    if [[ "$USES_DCGM_POWER" == "1" ]]; then
+        # Note (wenyao): on this cluster the DSV4-Pro checkpoint lives on the
+        # compute-node /mnt/numa1 NVMe (same staging the agentic path and the
+        # llm-d sweeps load from); the lustre alias target the shared dsv4
+        # block exports is not present here. Scoped to the power lane so the
+        # non-power lane keeps whatever the external-cluster staging expects.
+        export MODEL_PATH="/mnt/numa1/models/DeepSeek-V4-Pro"
+        git clone "$POWER_SRT_SLURM_URL" "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+        git checkout "$POWER_SRT_SLURM_PIN" || exit 1
+        # The power lane must run the exact pinned producer SHA, never a moving branch.
+        test "$(git rev-parse HEAD)" = "$POWER_SRT_SLURM_PIN" || { echo "Error: srt-slurm HEAD does not match POWER_SRT_SLURM_PIN=$POWER_SRT_SLURM_PIN" >&2; exit 1; }
+        git rev-parse HEAD > "$GITHUB_WORKSPACE/power-producer-sha.txt"
+    else
+        # Stay on NVIDIA/srt-slurm:main (default) — submission branch no
+        # longer needed; overlay our hand-rolled DSV4 sglang recipes onto it.
+        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+    fi
     mkdir -p recipes/sglang/deepseek-v4
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.1" ]]; then

runners/launch_gb300-nv.sh

@@ -155,11 +155,12 @@ if [[ -n "$CONFIG_FILE" && -f "$_RECIPE_SRC" ]] && awk '
     USES_DCGM_POWER=1
 fi
 
-# Note (wenyao): the producer pin descends from the fp8 v1.0.25 lineage
-# (cargo/maturin bootstrap); an fp4 power recipe would silently skip the
-# sa-submission branch it needs, so fail fast instead.
-if [[ "$USES_DCGM_POWER" == "1" && "$PRECISION" != "fp8" ]]; then
-    echo "Error: dcgm-power lanes are only validated for PRECISION=fp8, got: $PRECISION" >&2
+# Note (wenyao): the producer pin follows the srt-slurm main lineage that the
+# dynamo-sglang lanes run on (fp8 validated end-to-end, fp4 recipes
+# parse-verified against the pin); other frameworks clone diverging refs
+# (aflowers branch, sa-submission), so fail fast for them instead.
+if [[ "$USES_DCGM_POWER" == "1" && "$FRAMEWORK" != "dynamo-sglang" ]]; then
+    echo "Error: dcgm-power lanes are only validated for FRAMEWORK=dynamo-sglang, got: $FRAMEWORK" >&2
     exit 1
 fi
 
@@ -271,14 +272,25 @@ elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
     # Fixed-length DeepSeek-V4 recipes are version-controlled in this repository;
-    # overlay them onto the srt-slurm release that bootstraps cargo/maturin for
-    # the hash-pinned Dynamo source build before launch.
-    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
-    cd "$SRT_REPO_DIR"
-    git checkout v1.0.25
-    mkdir -p recipes/sglang/deepseek-v4/8k1k
+    # overlay them onto the selected srt-slurm checkout. Power lanes use the
+    # exact producer pin and stamp it for strict result provenance; non-power
+    # lanes retain the v1.0.25 release that bootstraps cargo/maturin for the
+    # hash-pinned Dynamo source build before launch.
+    if [[ "$USES_DCGM_POWER" == "1" ]]; then
+        git clone "$POWER_SRT_SLURM_URL" "$SRT_REPO_DIR" || exit 1
+        cd "$SRT_REPO_DIR" || exit 1
+        git checkout "$POWER_SRT_SLURM_PIN" || exit 1
+        # The power lane must run the exact pinned producer SHA, never a moving branch.
+        test "$(git rev-parse HEAD)" = "$POWER_SRT_SLURM_PIN" || { echo "Error: srt-slurm HEAD does not match POWER_SRT_SLURM_PIN=$POWER_SRT_SLURM_PIN" >&2; exit 1; }
+        git rev-parse HEAD > "$GITHUB_WORKSPACE/power-producer-sha.txt"
+    else
+        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
+        cd "$SRT_REPO_DIR" || exit 1
+        git checkout v1.0.25 || exit 1
+    fi
+    mkdir -p recipes/sglang/deepseek-v4/8k1k || exit 1
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k" \
-        recipes/sglang/deepseek-v4/8k1k
+        recipes/sglang/deepseek-v4/8k1k || exit 1
 elif [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
@@ -472,16 +484,19 @@ inject_synthetic_acceptance "$CONFIG_PATH" "$FRAMEWORK" || exit 1
 #   - glm5.1, whose GLM-5.1-NVFP4 weights are prestaged on the compute-node
 #     /scratch/models, and
 #   - qwen3.5 fp8, whose weights are also on the compute-node /scratch/models
-#     and which runs on srt-slurm:v1.0.25 (the release that has the preflight;
-#     qwen3.5 fp4 runs on v1.0.29, which has none).
+#     and which runs on srt-slurm:v1.0.25 (the release that has the preflight),
+#   - qwen3.5 fp4 dynamo-trt, which runs on v1.0.29 without that preflight, and
+#   - the qwen3.5 fp4 and dsv4 sglang power lanes, which run the pinned
+#     producer (a main-lineage fork that has the preflight) against the same
+#     /scratch checkpoints.
 # The engine still fails loudly at runtime if the path is genuinely missing on
 # the compute node. Other fixed-seq-len recipes resolve model.path to a
 # login-visible location, so keep the precheck enforced for them.
 SRTCTL_APPLY_ARGS=(
     -f "$CONFIG_FILE"
     --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
 )
-if [[ "$IS_AGENTIC" == "1" || "$MODEL_PREFIX" == "glm5.1" || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp8" ) || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-trt" ) ]]; then
+if [[ "$IS_AGENTIC" == "1" || "$MODEL_PREFIX" == "glm5.1" || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp8" ) || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && ( "$FRAMEWORK" == "dynamo-trt" || "$USES_DCGM_POWER" == "1" ) ) || ( "$USES_DCGM_POWER" == "1" && "$MODEL_PREFIX" == "dsv4" && "$FRAMEWORK" == "dynamo-sglang" ) ]]; then
     SRTCTL_APPLY_ARGS+=(--no-preflight)
 fi
 if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then

utils/test_gb200_power_official_contract.py

@@ -61,7 +61,7 @@ def assert_recipe_driven_detection(launcher):
     assert "t && /^  enabled: true$/        { e = 1 }" in launcher
     assert "USES_DCGM_POWER=1" in launcher
     assert (
-        'if [[ "$USES_DCGM_POWER" == "1" && "$PRECISION" != "fp8" ]]; then'
+        'if [[ "$USES_DCGM_POWER" == "1" && "$FRAMEWORK" != "dynamo-sglang" ]]; then'
         in launcher
     )
     # Exporter provisioning, pinned clone, srtslurm.yaml injection, and audit
@@ -149,14 +149,36 @@ def test_pin_literal_lives_only_in_the_two_launchers():
     ]
 
 
-def test_exactly_two_recipes_opt_into_dcgm_power():
+# Every recipe that opts into dcgm-power, with the exporter port its cluster
+# requires (im-gb300 nodes already bind 9401; gb200 keeps the default).
+POWER_RECIPES = {
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-dep8-dep16-6-c512.yaml": 9401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p1d-tp8-tp8-4-c1.yaml": 9401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb200-1p2d-dep8-dep16-10-c256.yaml": 9401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-dep4-dep16-5-c1024.yaml": 19401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-gb300-1p1d-tp4-tp4-2-c1.yaml": 19401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml": 9401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp4/8k1k/disagg/stp/8k1k_stp_lowlat_0.yaml": 19401,
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml": 19401,
+}
+
+
+def test_exactly_the_declared_recipes_opt_into_dcgm_power():
     hits = git_grep_lines(
         "-lF", "provider: dcgm-power", "--", "benchmarks/multi_node/srt-slurm-recipes"
     )
-    assert sorted(hits) == [
-        "benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp8/8k1k/1p1d-tp4-tp4.yaml",
-        "benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb300-fp8/8k1k/1p1d-tp4-tp4.yaml",
-    ]
+    assert sorted(hits) == sorted(POWER_RECIPES)
+
+
+def test_every_power_recipe_declares_the_same_telemetry_contract():
+    for path, port in POWER_RECIPES.items():
+        telemetry = yaml.safe_load((REPO_ROOT / path).read_text())["telemetry"]
+        assert telemetry["enabled"] is True, path
+        assert telemetry["provider"] == "dcgm-power", path
+        assert telemetry["default_frequency"] == 1.0, path
+        assert telemetry["required"] is True, path
+        assert telemetry["dcgm_exporter"]["container_image"] == "dcgm-exporter", path
+        assert telemetry["dcgm_exporter"]["port"] == port, path
 
 
 def test_lane_and_pin_have_no_env_override_backdoor():