diff --git a/.github/workflows/test-process-result.yml b/.github/workflows/test-process-result.yml
index 3d3147fab9..2e2af1d32e 100644
--- a/.github/workflows/test-process-result.yml
+++ b/.github/workflows/test-process-result.yml
@@ -9,6 +9,8 @@ on:
       - '.github/workflows/test-process-result.yml'
       - 'benchmarks/benchmark_lib.sh'
       - 'benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/**/*.yaml'
+      - 'benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/**/*.yaml'
+      - 'benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/**/*.yaml'
       - 'benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-h200-tp8-mtp-kvoffload.yaml'
       - 'benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-2p2d-pcp8-tp8-dp8-mtp.yaml'
       - 'benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-*/**/*.yaml'
@@ -16,6 +18,9 @@ on:
       - 'configs/nvidia-master.yaml'
       - 'runners/launch_gb200-nv.sh'
       - 'runners/launch_gb300-nv.sh'
+      - 'runners/launch_b200-dgxc.sh'
+      - 'runners/launch_b200-nscale-slurm.sh'
+      - 'runners/launch_b300-nv.sh'
       - 'runners/launch_h200-dgxc-slurm.sh'
       - 'runners/inject_srt_power_concurrencies.py'
       - 'utils/aggregate_power.py'
@@ -30,6 +35,7 @@ on:
       - 'utils/process_result.py'
       - 'utils/test_aggregate_power.py'
       - 'utils/test_aggregate_power_multinode.py'
+      - 'utils/test_b200_b300_power_official_contract.py'
       - 'utils/test_gb200_power_official_contract.py'
       - 'utils/test_gb300_power_official_contract.py'
       - 'utils/test_h200_power_official_contract.py'
@@ -62,4 +68,4 @@ jobs:
       - name: Run pytest
         run: |
           cd utils
-          python -m pytest test_aggregate_power.py test_aggregate_power_multinode.py agentic/aggregation/test_power_adapter.py agentic/aggregation/test_power_lifecycle.py agentic/aggregation/test_process_agentic_result.py test_gb200_power_official_contract.py test_gb300_power_official_contract.py test_h200_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_power_adapter.py agentic/aggregation/test_power_lifecycle.py agentic/aggregation/test_process_agentic_result.py test_b200_b300_power_official_contract.py test_gb200_power_official_contract.py test_gb300_power_official_contract.py test_h200_power_official_contract.py test_inject_srt_power_concurrencies.py test_process_result.py -v
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p1d-tp8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p1d-tp8-tp8-mtp.yaml
index 685537c6a6..2e2016a15b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p1d-tp8-tp8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p1d-tp8-tp8-mtp.yaml
@@ -118,3 +118,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 9401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p6d-dep8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p6d-dep8-tp8-mtp.yaml
index 18b377b863..bfb83c48dd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p6d-dep8-tp8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-low-latency-1p6d-dep8-tp8-mtp.yaml
@@ -125,3 +125,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 9401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p1d-dep8-dep8-mtp.yaml
index ea30d07e55..b2291ca000 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p1d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p1d-dep8-dep8-mtp.yaml
@@ -126,3 +126,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 9401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p2d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p2d-dep8-dep8-mtp.yaml
index 9a6254e989..e709c508c8 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p2d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-1p2d-dep8-dep8-mtp.yaml
@@ -137,3 +137,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 9401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-5p3d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-5p3d-dep8-dep8-mtp.yaml
index e5d9a708a1..b1f2e1eba7 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-5p3d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b200-mid-curve-5p3d-dep8-dep8-mtp.yaml
@@ -126,3 +126,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 9401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
index a07d5bb0a7..3ea1d3e722 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8-mtp.yaml
@@ -132,3 +132,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
index 6612aae62f..42bd30e552 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-dep4-dep8.yaml
@@ -143,3 +143,16 @@ 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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
index 75c1815c8a..2293fa9a81 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4-mtp.yaml
@@ -123,3 +123,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml
index 6f0c0f7f68..e55d8dc2f5 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p1d-tp4-tp4.yaml
@@ -118,3 +118,16 @@ 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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
index fef5285ceb..5737b5d734 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p2d-dep4-dep8-mtp.yaml
@@ -133,3 +133,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
index d47b469945..fe5819b0af 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-1p6d-dep4-tp4-mtp.yaml
@@ -127,3 +127,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
index e796e2c680..96912af606 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8-mtp.yaml
@@ -133,3 +133,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml
index 565a74d739..4c5c5ef4db 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-2p1d-dep4-dep8.yaml
@@ -143,3 +143,16 @@ 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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
index 681f350d90..2b9713b708 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8-mtp.yaml
@@ -133,3 +133,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
index 9a8123c439..8ad309af71 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-4p1d-dep4-dep8.yaml
@@ -143,3 +143,16 @@ 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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
index 4813ac793f..f7ff198115 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8-mtp.yaml
@@ -133,3 +133,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml
index c47335237a..6ba7519714 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-6p1d-dep4-dep8.yaml
@@ -143,3 +143,16 @@ 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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml
index ed3b27e906..7bdc43dd73 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/8k1k/disagg-b300-8p1d-dep4-dep8-mtp.yaml
@@ -133,3 +133,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   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
+    port: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml
index d317a218a7..72e6dd1b16 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-dep8-mtp.yaml
@@ -31,6 +31,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -127,3 +128,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml
index b73afe1adf..dc4e403072 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p1d-dep8-tp8-mtp.yaml
@@ -31,6 +31,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -112,3 +113,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml
index 5c9eb254e3..1bebd09a27 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-1p2d-dep8-dep8-mtp.yaml
@@ -31,6 +31,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -128,3 +129,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml
index f71c4b3a3e..2c9e9e2feb 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-2p1d-dep8-dep8-mtp.yaml
@@ -31,6 +31,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -127,3 +128,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml
index 7eb6c47a34..0b7841266e 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-3p1d-dep8-dep8-mtp.yaml
@@ -31,6 +31,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -127,3 +128,16 @@ benchmark:
   req_rate: "inf"
   use_chat_template: true
   custom_tokenizer: "sa_bench_tokenizers.vllm_deepseek_v4.VLLMDeepseekV4Tokenizer"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
index e81ef8631c..26805d4efd 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c1.yaml
@@ -29,6 +29,7 @@ frontend:
   enable_multiple_frontends: false
   type: dynamo
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   connector: null
   decode_environment:
     NCCL_CUMEM_ENABLE: '1'
@@ -111,3 +112,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
index b296440158..dc2fd3bdaa 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c32-c128.yaml
@@ -29,6 +29,7 @@ frontend:
   enable_multiple_frontends: false
   type: dynamo
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   connector: null
   decode_environment:
     NCCL_CUMEM_ENABLE: '1'
@@ -111,3 +112,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
index b566c05fae..a1c93b179d 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-latency-c64.yaml
@@ -29,6 +29,7 @@ frontend:
   enable_multiple_frontends: false
   type: dynamo
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   connector: null
   decode_environment:
     NCCL_CUMEM_ENABLE: '1'
@@ -111,3 +112,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
index 43190836a1..fe4b49077b 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c256.yaml
@@ -29,6 +29,7 @@ frontend:
   enable_multiple_frontends: false
   type: dynamo
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   connector: null
   decode_environment:
     NCCL_CUMEM_ENABLE: '1'
@@ -115,3 +116,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
index 283bb839f3..8fe69b472f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/8k1k/disagg-b200-low-middle-c512.yaml
@@ -29,6 +29,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -116,3 +117,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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
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
index 9247a31daf..b16d38d33d 100644
--- 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
@@ -38,6 +38,7 @@ frontend:
   enable_multiple_frontends: false
 
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -128,3 +129,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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: 19401
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
index 260e16fc7d..8f2a01f5c5 100644
--- 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
@@ -38,6 +38,7 @@ frontend:
   enable_multiple_frontends: false
 
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -122,3 +123,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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: 19401
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
index 98b701790a..557f6b1e95 100644
--- 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
@@ -38,6 +38,7 @@ frontend:
   enable_multiple_frontends: false
 
 backend:
+  set_cuda_visible_devices: true  # keep the aflowers/vllm-gb200-v0.20.0 provenance: bind GPUs via CVD, not --device-ids
   type: vllm
   connector: null
   prefill_environment:
@@ -128,3 +129,16 @@ identity:
   frameworks:
     dynamo: "1.2.0.dev20260426"
     vllm: "0.23.0"
+
+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: 19401
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-dep8-c2048.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-dep8-c2048.yaml
index e3c08dc477..b7378fe35f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-dep8-c2048.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-dep8-c2048.yaml
@@ -110,3 +110,16 @@ benchmark:
   concurrencies: '2048'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-tp8-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-tp8-c1.yaml
index 1c0abe2fa0..cf8d37ff59 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-tp8-c1.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p1d-dep8-tp8-c1.yaml
@@ -108,3 +108,16 @@ benchmark:
   concurrencies: '1'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p4d-dep4-tp4-c512.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p4d-dep4-tp4-c512.yaml
index 1ca87ad9b8..f17b45287c 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p4d-dep4-tp4-c512.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p4d-dep4-tp4-c512.yaml
@@ -107,3 +107,16 @@ benchmark:
   concurrencies: '512'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c128.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c128.yaml
index 44f8ff7849..e33c6b773f 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c128.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c128.yaml
@@ -107,3 +107,16 @@ benchmark:
   concurrencies: '128'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c32.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c32.yaml
index c748107fec..07f3c13792 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c32.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-1p8d-dep4-tp4-c32.yaml
@@ -107,3 +107,16 @@ benchmark:
   concurrencies: '32'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-2p1d-dep8-dep8-c8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-2p1d-dep8-dep8-c8192.yaml
index 54e623d767..9c80baf673 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-2p1d-dep8-dep8-c8192.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/8k1k/disagg-b200-2p1d-dep8-dep8-c8192.yaml
@@ -110,3 +110,16 @@ benchmark:
   concurrencies: '8192'
   req_rate: inf
   use_chat_template: true
+
+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
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d51ee4c22b..de9533314a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6475,3 +6475,18 @@
     - "Replace the TP2/EP2 arm with TP2/EP1 so the 2-GPU point matches qwen3.5-fp4-b200-sglang-agentic-mtp."
     - "Add HiCache host-DRAM KV tier arms at TP4 concurrency 40, 48, 56, and 64 and TP2 concurrency 20, 24, 28, and 32, using hicache ratio 1.5 with write_through, direct io, and page_first_direct layout."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2693
+
+
+- config-keys:
+    - kimik2.6-fp4-b200-dynamo-vllm
+    - dsv4-fp4-b200-dynamo-vllm
+    - dsv4-fp4-b300-dynamo-vllm
+    - dsv4-fp4-b300-dynamo-sglang
+    - dsv4-fp4-b200-dynamo-sglang-mtp
+    - dsv4-fp4-b200-dynamo-vllm-mtp
+    - dsv4-fp4-b300-dynamo-sglang-mtp
+  description:
+    - "Enable required dcgm-power telemetry on 37 eligible recipes: 21 B200 and 16 B300."
+    - "Required lanes promote exporter startup timeouts, launch failures, and endpoint resolution failures to blocking validation failures, so a recipe cannot publish a result whose power collection never started."
+    - "Route only enabled recipes through the immutable producer fork and preserve non-power launcher revisions."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2688
diff --git a/runners/launch_b200-dgxc.sh b/runners/launch_b200-dgxc.sh
index 498ad7d1c1..f23701f1f5 100644
--- a/runners/launch_b200-dgxc.sh
+++ b/runners/launch_b200-dgxc.sh
@@ -3,6 +3,8 @@
 # System-specific configuration for B200 DGXC Slurm cluster
 SLURM_PARTITION="${SLURM_PARTITION:-gpu-2}"
 SLURM_ACCOUNT="${SLURM_ACCOUNT:-benchmark}"
+POWER_SRT_SLURM_URL="https://github.com/edwingao28/srt-slurm.git"
+POWER_SRT_SLURM_PIN="e5c837f06a362dc888dfea2ee588e9f19c298270"
 
 set -x
 
@@ -158,6 +160,32 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
         exit 1
     fi
 
+    USES_DCGM_POWER=0
+    _POWER_CONFIG_FILE="${CONFIG_FILE:-}"
+    if [[ "${EVAL_ONLY:-false}" == "true" && -n "${EVAL_CONFIG_FILE:-}" ]]; then
+        _POWER_CONFIG_FILE="$EVAL_CONFIG_FILE"
+    fi
+    _RECIPE_REL="${_POWER_CONFIG_FILE%%:*}"
+    _RECIPE_SRC="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/${_RECIPE_REL#recipes/}"
+    if [[ -n "$_POWER_CONFIG_FILE" && -f "$_RECIPE_SRC" ]] && awk '
+        /^telemetry:/ { t = 1; next }
+        t && /^[^ ]/  { t = 0 }
+        t && /^  provider: dcgm-power$/ { p = 1 }
+        t && /^  enabled: true$/        { e = 1 }
+        END { exit !(p && e) }
+    ' "$_RECIPE_SRC"; then
+        USES_DCGM_POWER=1
+    fi
+    if [[ "$USES_DCGM_POWER" == "1" && (
+        "${IS_AGENTIC:-0}" == "1" ||
+        "$MODEL_PREFIX" != "dsv4" ||
+        "$PRECISION" != "fp4" ||
+        "$FRAMEWORK" != "dynamo-vllm"
+    ) ]]; then
+        echo "Error: B200 DGXC dcgm-power is limited to fixed-sequence DSV4 FP4 dynamo-vllm" >&2
+        exit 1
+    fi
+
     export SERVED_MODEL_NAME=$MODEL
 
     echo "Cloning srt-slurm repository..."
@@ -170,7 +198,15 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     # Kimi K3 aggregate profiles use the srt-slurm fork that supports direct
     # multi-node vLLM. Pin the tested renderer so branch movement cannot change
     # generated rank commands between sweep points.
-    if [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
+    if [[ "$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
+        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"
+        mkdir -p recipes/vllm/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+    elif [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
         git clone --branch klaud/direct-vllm-multinode --single-branch https://github.com/functionstackx/srt-slurm-nv.git "$SRT_REPO_DIR" || exit 1
         cd "$SRT_REPO_DIR" || exit 1
         git checkout df5baa93f4caf5169dea2a4236ad2cc742fe40e7 || exit 1
@@ -286,6 +322,17 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     import_squash "$SQUASH_FILE" "$IMAGE" || exit 1
     import_squash "$NGINX_SQUASH_FILE" "$NGINX_IMAGE" || exit 1
 
+    if [[ "$USES_DCGM_POWER" == "1" ]]; then
+        DCGM_EXPORTER_IMAGE="nvcr.io/nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless"
+        # enroot resolves bare paths against Docker Hub; nvcr.io pulls need the registry# form
+        DCGM_EXPORTER_ENROOT_REF="${DCGM_EXPORTER_IMAGE/nvcr.io\//nvcr.io#}"
+        DCGM_EXPORTER_SQSH="$SQUASH_DIR/$(echo "$DCGM_EXPORTER_IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+        import_squash "$DCGM_EXPORTER_SQSH" "$DCGM_EXPORTER_ENROOT_REF" || exit 1
+        test -r "$DCGM_EXPORTER_SQSH" || { echo "Error: DCGM exporter squash not readable: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+        unsquashfs -l "$DCGM_EXPORTER_SQSH" > /dev/null || { echo "Error: DCGM exporter squash invalid: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+        sha256sum "$DCGM_EXPORTER_SQSH" > "$GITHUB_WORKSPACE/exporter-image.sha256"
+    fi
+
     export ISL="$ISL"
     export OSL="$OSL"
     export EVAL_ONLY="${EVAL_ONLY:-false}"
@@ -336,6 +383,11 @@ use_exclusive_sbatch_directive: true
 ${DEFAULT_MOUNTS_BLOCK}
 EOF
 
+    if [[ "$USES_DCGM_POWER" == "1" ]]; then
+        sed -i "/^  nginx-sqsh:/a\\  dcgm-exporter: ${DCGM_EXPORTER_SQSH}" srtslurm.yaml
+        grep -q "^  dcgm-exporter: " srtslurm.yaml || { echo "Error: dcgm-exporter injection failed: nginx-sqsh anchor not found in srtslurm.yaml" >&2; exit 1; }
+    fi
+
     echo "Generated srtslurm.yaml:"
     cat srtslurm.yaml
 
@@ -435,6 +487,12 @@ EOF
 
     echo "Found logs directory: $LOGS_DIR"
 
+    if [[ "$USES_DCGM_POWER" == "1" ]]; then
+        mkdir -p "$LOGS_DIR/power"
+        cp "$GITHUB_WORKSPACE/exporter-image.sha256" "$LOGS_DIR/power/exporter-image.sha256"
+        cp "$GITHUB_WORKSPACE/power-producer-sha.txt" "$LOGS_DIR/power/power-producer-sha.txt"
+    fi
+
     cp -r "$LOGS_DIR" "$GITHUB_WORKSPACE/LOGS"
     tar czf "$GITHUB_WORKSPACE/multinode_server_logs.tar.gz" -C "$LOGS_DIR" .
 
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index 4bf9531d0f..e9fbd9a5e6 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -14,6 +14,8 @@
 
 SLURM_PARTITION="batch_1"
 SLURM_ACCOUNT="benchmark"
+POWER_SRT_SLURM_URL="https://github.com/edwingao28/srt-slurm.git"
+POWER_SRT_SLURM_PIN="e5c837f06a362dc888dfea2ee588e9f19c298270"
 
 # Node-local NVMe, not a shared filesystem: much faster for the ~1.6T
 # DeepSeek-V4-Pro load, and already pre-staged on every nscale compute node.
@@ -59,12 +61,55 @@ if [[ $FRAMEWORK != "dynamo-vllm" ]] &&
     exit 1
 fi
 
+USES_DCGM_POWER=0
+_POWER_CONFIG_FILE="${CONFIG_FILE:-}"
+if [[ "${EVAL_ONLY:-false}" == "true" && -n "${EVAL_CONFIG_FILE:-}" ]]; then
+    _POWER_CONFIG_FILE="$EVAL_CONFIG_FILE"
+fi
+_RECIPE_REL="${_POWER_CONFIG_FILE%%:*}"
+_RECIPE_SRC="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/${_RECIPE_REL#recipes/}"
+if [[ -n "$_POWER_CONFIG_FILE" && -f "$_RECIPE_SRC" ]] && awk '
+    /^telemetry:/ { t = 1; next }
+    t && /^[^ ]/  { t = 0 }
+    t && /^  provider: dcgm-power$/ { p = 1 }
+    t && /^  enabled: true$/        { e = 1 }
+    END { exit !(p && e) }
+' "$_RECIPE_SRC"; then
+    USES_DCGM_POWER=1
+fi
+if [[ "$USES_DCGM_POWER" == "1" && (
+    "${IS_AGENTIC:-0}" == "1" ||
+    "$PRECISION" != "fp4" ||
+    ( "$MODEL_PREFIX" == "dsv4" && "$FRAMEWORK" != "dynamo-sglang" && "$FRAMEWORK" != "dynamo-vllm" ) ||
+    ( "$MODEL_PREFIX" == "kimik2.6" && "$FRAMEWORK" != "dynamo-vllm" ) ||
+    ( "$MODEL_PREFIX" != "dsv4" && "$MODEL_PREFIX" != "kimik2.6" )
+) ]]; then
+    echo "Error: B200 nscale dcgm-power is limited to fixed-sequence DSV4/Kimi-K2.6 FP4 lanes" >&2
+    exit 1
+fi
+
 export SERVED_MODEL_NAME=$MODEL
 
 echo "Cloning srt-slurm repository..."
 SRT_REPO_DIR="srt-slurm"
 rm -rf "$SRT_REPO_DIR"
-if [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
+if [[ "$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
+    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"
+    if [[ "$MODEL_PREFIX" == "dsv4" && "$FRAMEWORK" == "dynamo-sglang" ]]; then
+        mkdir -p recipes/sglang/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
+    elif [[ "$MODEL_PREFIX" == "dsv4" ]]; then
+        mkdir -p recipes/vllm/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+    else
+        mkdir -p recipes/vllm/kimi-k2.6
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6" recipes/vllm/kimi-k2.6
+    fi
+elif [[ "$IS_AGENTIC" == "1" && $MODEL_PREFIX == "kimik3" ]]; then
     # Pin the tested renderer so branch movement cannot change generated rank
     # commands between sweep points.
     git clone --branch main --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR" || exit 1
@@ -149,6 +194,17 @@ import_squash() {
 import_squash "$SQUASH_FILE" "$IMAGE" || exit 1
 import_squash "$NGINX_SQUASH_FILE" "$NGINX_IMAGE" || exit 1
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    DCGM_EXPORTER_IMAGE="nvcr.io/nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless"
+    # enroot resolves bare paths against Docker Hub; nvcr.io pulls need the registry# form
+    DCGM_EXPORTER_ENROOT_REF="${DCGM_EXPORTER_IMAGE/nvcr.io\//nvcr.io#}"
+    DCGM_EXPORTER_SQSH="$SQUASH_DIR/$(echo "$DCGM_EXPORTER_IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    import_squash "$DCGM_EXPORTER_SQSH" "$DCGM_EXPORTER_ENROOT_REF" || exit 1
+    test -r "$DCGM_EXPORTER_SQSH" || { echo "Error: DCGM exporter squash not readable: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+    unsquashfs -l "$DCGM_EXPORTER_SQSH" > /dev/null || { echo "Error: DCGM exporter squash invalid: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+    sha256sum "$DCGM_EXPORTER_SQSH" > "$GITHUB_WORKSPACE/exporter-image.sha256"
+fi
+
 export ISL="$ISL"
 export OSL="$OSL"
 export EVAL_ONLY="${EVAL_ONLY:-false}"
@@ -193,6 +249,11 @@ use_exclusive_sbatch_directive: true
 ${DEFAULT_MOUNTS_BLOCK}
 EOF
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    sed -i "/^  nginx-sqsh:/a\\  dcgm-exporter: ${DCGM_EXPORTER_SQSH}" srtslurm.yaml
+    grep -q "^  dcgm-exporter: " srtslurm.yaml || { echo "Error: dcgm-exporter injection failed: nginx-sqsh anchor not found in srtslurm.yaml" >&2; exit 1; }
+fi
+
 echo "Generated srtslurm.yaml:"
 cat srtslurm.yaml
 
@@ -235,7 +296,7 @@ SRTCTL_PREFLIGHT_ARGS=()
 # These weights are staged on the Slurm compute nodes, not the login node.
 if [[ $MODEL_PREFIX == "kimik2.6" ]] ||
    [[ $MODEL_PREFIX == "kimik3" ]] ||
-   [[ $MODEL_PREFIX == "dsv4" && $FRAMEWORK == "dynamo-sglang" ]]; then
+   [[ $MODEL_PREFIX == "dsv4" ]]; then
     SRTCTL_PREFLIGHT_ARGS+=(--no-preflight)
 fi
 
@@ -270,6 +331,12 @@ if [ ! -d "$LOGS_DIR" ]; then
     exit 1
 fi
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    mkdir -p "$LOGS_DIR/power"
+    cp "$GITHUB_WORKSPACE/exporter-image.sha256" "$LOGS_DIR/power/exporter-image.sha256"
+    cp "$GITHUB_WORKSPACE/power-producer-sha.txt" "$LOGS_DIR/power/power-producer-sha.txt"
+fi
+
 cp -r "$LOGS_DIR" "$GITHUB_WORKSPACE/LOGS"
 bundle_server_logs "$LOGS_DIR" "$GITHUB_WORKSPACE/multinode_server_logs.tar.gz"
 
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index f7878a8ad8..37d35c7b3d 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -3,6 +3,8 @@
 # System-specific configuration for B300 NV Slurm cluster (sa-shared)
 SLURM_PARTITION="batch_1"
 SLURM_ACCOUNT="benchmark"
+POWER_SRT_SLURM_URL="https://github.com/edwingao28/srt-slurm.git"
+POWER_SRT_SLURM_PIN="e5c837f06a362dc888dfea2ee588e9f19c298270"
 # b300-018 repeatedly times out UCX/NIXL transfers; allow an empty override to disable this.
 MINIMAX_M3_SLURM_EXCLUDED_NODELIST="${MINIMAX_M3_SLURM_EXCLUDED_NODELIST-b300-018}"
 
@@ -16,6 +18,28 @@ if [[ $FRAMEWORK != "dynamo-sglang" && $FRAMEWORK != "dynamo-trt" && $FRAMEWORK
     exit 1
 fi
 
+USES_DCGM_POWER=0
+_RECIPE_REL="${CONFIG_FILE%%:*}"
+_RECIPE_SRC="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/${_RECIPE_REL#recipes/}"
+if [[ -n "$CONFIG_FILE" && -f "$_RECIPE_SRC" ]] && awk '
+    /^telemetry:/ { t = 1; next }
+    t && /^[^ ]/  { t = 0 }
+    t && /^  provider: dcgm-power$/ { p = 1 }
+    t && /^  enabled: true$/        { e = 1 }
+    END { exit !(p && e) }
+' "$_RECIPE_SRC"; then
+    USES_DCGM_POWER=1
+fi
+if [[ "$USES_DCGM_POWER" == "1" && (
+    "${IS_AGENTIC:-0}" == "1" ||
+    "$MODEL_PREFIX" != "dsv4" ||
+    "$PRECISION" != "fp4" ||
+    ( "$FRAMEWORK" != "dynamo-sglang" && "$FRAMEWORK" != "dynamo-vllm" )
+) ]]; then
+    echo "Error: B300 dcgm-power is limited to fixed-sequence DSV4 FP4 dynamo-sglang/vllm" >&2
+    exit 1
+fi
+
 # MODEL_PATH: Override with pre-downloaded paths on B300 runner
 # The yaml files specify HuggingFace model IDs for portability, but we use
 # local paths to avoid repeated downloading on the shared B300 cluster.
@@ -70,7 +94,20 @@ if [ -d "$SRT_REPO_DIR" ]; then
 fi
 
 # TODO(CJQ): make first class upon srt-slurm upstream refactor
-if [[ "$IS_AGENTIC" == "1" ]]; then
+if [[ "$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
+    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"
+    if [[ "$FRAMEWORK" == "dynamo-sglang" ]]; then
+        mkdir -p recipes/sglang/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4" recipes/sglang/deepseek-v4
+    else
+        mkdir -p recipes/vllm/deepseek-v4
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4" recipes/vllm/deepseek-v4
+    fi
+elif [[ "$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" && $MODEL_PREFIX == "dsv4" ]]; then
@@ -139,6 +176,27 @@ NGINX_SQUASH_FILE="/data/squash/$(echo "$NGINX_IMAGE" | sed 's/[\/:@#]/_/g').sqs
 srun -N 1 -A $SLURM_ACCOUNT -p $SLURM_PARTITION bash -c "enroot import -o $SQUASH_FILE docker://$IMAGE"
 srun -N 1 -A $SLURM_ACCOUNT -p $SLURM_PARTITION bash -c "enroot import -o $NGINX_SQUASH_FILE docker://$NGINX_IMAGE"
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    DCGM_EXPORTER_IMAGE="nvcr.io/nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless"
+    # enroot resolves bare paths against Docker Hub; nvcr.io pulls need the registry# form
+    DCGM_EXPORTER_ENROOT_REF="${DCGM_EXPORTER_IMAGE/nvcr.io\//nvcr.io#}"
+    DCGM_EXPORTER_SQSH="/data/squash/$(echo "$DCGM_EXPORTER_IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+    DCGM_EXPORTER_LOCK="${DCGM_EXPORTER_SQSH}.lock"
+    srun -N 1 -A "$SLURM_ACCOUNT" -p "$SLURM_PARTITION" bash -c "
+        set -euo pipefail
+        exec 9>\"$DCGM_EXPORTER_LOCK\"
+        flock -w 1800 9
+        if unsquashfs -l \"$DCGM_EXPORTER_SQSH\" > /dev/null 2>&1; then
+            exit 0
+        fi
+        rm -f \"$DCGM_EXPORTER_SQSH\"
+        enroot import -o \"$DCGM_EXPORTER_SQSH\" \"docker://$DCGM_EXPORTER_ENROOT_REF\"
+        unsquashfs -l \"$DCGM_EXPORTER_SQSH\" > /dev/null
+    " || exit 1
+    test -r "$DCGM_EXPORTER_SQSH" || { echo "Error: DCGM exporter squash not readable: $DCGM_EXPORTER_SQSH" >&2; exit 1; }
+    sha256sum "$DCGM_EXPORTER_SQSH" > "$GITHUB_WORKSPACE/exporter-image.sha256"
+fi
+
 export ISL="$ISL"
 export OSL="$OSL"
 export EVAL_ONLY="${EVAL_ONLY:-false}"
@@ -173,6 +231,11 @@ default_mounts:
   "/opt/ucx-no-ud": "/usr/local/ucx"
 EOF
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    sed -i "/^  nginx-sqsh:/a\\  dcgm-exporter: ${DCGM_EXPORTER_SQSH}" srtslurm.yaml
+    grep -q "^  dcgm-exporter: " srtslurm.yaml || { echo "Error: dcgm-exporter injection failed: nginx-sqsh anchor not found in srtslurm.yaml" >&2; exit 1; }
+fi
+
 echo "Generated srtslurm.yaml:"
 cat srtslurm.yaml
 
@@ -288,6 +351,12 @@ fi
 
 echo "Found logs directory: $LOGS_DIR"
 
+if [[ "$USES_DCGM_POWER" == "1" ]]; then
+    mkdir -p "$LOGS_DIR/power"
+    cp "$GITHUB_WORKSPACE/exporter-image.sha256" "$LOGS_DIR/power/exporter-image.sha256"
+    cp "$GITHUB_WORKSPACE/power-producer-sha.txt" "$LOGS_DIR/power/power-producer-sha.txt"
+fi
+
 cp -r "$LOGS_DIR" "$GITHUB_WORKSPACE/LOGS"
 tar czf "$GITHUB_WORKSPACE/multinode_server_logs.tar.gz" -C "$LOGS_DIR" .
 
diff --git a/utils/test_b200_b300_power_official_contract.py b/utils/test_b200_b300_power_official_contract.py
new file mode 100644
index 0000000000..e083409df0
--- /dev/null
+++ b/utils/test_b200_b300_power_official_contract.py
@@ -0,0 +1,251 @@
+"""Contract for required B200/B300 multinode dcgm-power lanes."""
+
+from __future__ import annotations
+
+import re
+import subprocess
+from collections import Counter
+from pathlib import Path
+
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+RECIPE_ROOT = REPO_ROOT / "benchmarks/multi_node/srt-slurm-recipes"
+MASTER = REPO_ROOT / "configs/nvidia-master.yaml"
+WORKFLOW = REPO_ROOT / ".github/workflows/test-process-result.yml"
+LAUNCHERS = {
+    "cluster:b200-dgxc": REPO_ROOT / "runners/launch_b200-dgxc.sh",
+    "b200-nscale": REPO_ROOT / "runners/launch_b200-nscale-slurm.sh",
+    "b200-new": REPO_ROOT / "runners/launch_b200-nscale-slurm.sh",
+    "b300": REPO_ROOT / "runners/launch_b300-nv.sh",
+}
+
+PRODUCER_URL = "https://github.com/edwingao28/srt-slurm.git"
+PRODUCER_SHA = "e5c837f06a362dc888dfea2ee588e9f19c298270"
+EXPORTER_IMAGE = "nvcr.io/nvidia/k8s/dcgm-exporter:4.6.0-4.8.3-distroless"
+PROCESS_RESULT_RECIPE_GLOBS = (
+    "benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/**/*.yaml",
+    "benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/**/*.yaml",
+    "benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k2.6/b200-fp4/**/*.yaml",
+)
+
+
+def _config_files(value):
+    if isinstance(value, dict):
+        for child in value.values():
+            yield from _config_files(child)
+    elif isinstance(value, list):
+        for child in value:
+            yield from _config_files(child)
+    elif isinstance(value, str) and value.startswith("CONFIG_FILE="):
+        yield value.removeprefix("CONFIG_FILE=").split(":", 1)[0]
+
+
+def _power_recipes():
+    master = yaml.safe_load(MASTER.read_text(encoding="utf-8"))
+    recipes = {}
+    for config in master.values():
+        if not isinstance(config, dict):
+            continue
+        runner = config.get("runner")
+        if (
+            runner not in LAUNCHERS
+            or config.get("model-prefix") not in {"dsv4", "kimik2.6"}
+            or not config.get("disagg")
+        ):
+            continue
+        for config_file in _config_files(config.get("scenarios", {})):
+            relative = config_file.removeprefix("recipes/")
+            path = RECIPE_ROOT / relative
+            if path.exists() and "/agentic/" not in path.as_posix():
+                recipes[path.relative_to(REPO_ROOT).as_posix()] = {
+                    "runner": runner,
+                    "model": config["model-prefix"],
+                    "framework": config["framework"],
+                }
+    return recipes
+
+
+POWER_RECIPES = _power_recipes()
+
+
+def test_exactly_37_supported_recipes_are_selected_from_the_master_config():
+    assert len(POWER_RECIPES) == 37
+    assert Counter(item["runner"] for item in POWER_RECIPES.values()) == {
+        "cluster:b200-dgxc": 5,
+        "b200-nscale": 10,
+        "b200-new": 6,
+        "b300": 16,
+    }
+    assert Counter(item["framework"] for item in POWER_RECIPES.values()) == {
+        "dynamo-sglang": 18,
+        "dynamo-vllm": 19,
+    }
+    assert all(
+        "kimi-k3" not in path and "/agentic/" not in path for path in POWER_RECIPES
+    )
+    dedicated = sum(
+        yaml.safe_load((REPO_ROOT / path).read_text(encoding="utf-8"))
+        .get("infra", {})
+        .get("etcd_nats_dedicated_node", False)
+        for path in POWER_RECIPES
+    )
+    assert dedicated == 15
+
+
+def test_every_selected_recipe_uses_the_required_contract():
+    for path, item in POWER_RECIPES.items():
+        recipe = yaml.safe_load((REPO_ROOT / path).read_text(encoding="utf-8"))
+        assert recipe["benchmark"]["type"] == "sa-bench", path
+        assert recipe["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": 19401 if item["runner"] == "b300" else 9401,
+            },
+        }, path
+
+
+def test_launchers_use_recipe_gating_and_the_exact_fork_commit():
+    for launcher in set(LAUNCHERS.values()):
+        text = launcher.read_text(encoding="utf-8")
+        assert re.findall(r"^\s*USES_DCGM_POWER=(\S+)$", text, re.MULTILINE) == [
+            "0",
+            "1",
+        ], launcher
+        assert re.search(r'_RECIPE_REL="\$\{(?:_POWER_)?CONFIG_FILE%%:\*\}"', text), (
+            launcher
+        )
+        assert "/^telemetry:/ { t = 1; next }" in text
+        assert "t && /^  provider: dcgm-power$/ { p = 1 }" in text
+        assert "t && /^  enabled: true$/        { e = 1 }" in text
+        assert f'POWER_SRT_SLURM_URL="{PRODUCER_URL}"' in text
+        assert f'POWER_SRT_SLURM_PIN="{PRODUCER_SHA}"' in text
+        assert 'git clone "$POWER_SRT_SLURM_URL" "$SRT_REPO_DIR"' in text
+        assert 'git checkout "$POWER_SRT_SLURM_PIN" || exit 1' in text
+        assert 'test "$(git rev-parse HEAD)" = "$POWER_SRT_SLURM_PIN"' in text
+        assert 'git rev-parse HEAD > "$GITHUB_WORKSPACE/power-producer-sha.txt"' in text
+
+
+def test_launcher_awk_gate_accepts_all_selected_recipes_and_rejects_non_power():
+    program = """
+        /^telemetry:/ { t = 1; next }
+        t && /^[^ ]/  { t = 0 }
+        t && /^  provider: dcgm-power$/ { p = 1 }
+        t && /^  enabled: true$/        { e = 1 }
+        END { exit !(p && e) }
+    """
+    for path in POWER_RECIPES:
+        result = subprocess.run(
+            ["awk", program, str(REPO_ROOT / path)],
+            capture_output=True,
+            text=True,
+            check=False,
+        )
+        assert result.returncode == 0, (path, result.stderr)
+
+    non_power = RECIPE_ROOT / "vllm/kimi-k3/agentic/agg-b200-dep8-mtp-agentic.yaml"
+    result = subprocess.run(
+        ["awk", program, str(non_power)],
+        capture_output=True,
+        text=True,
+        check=False,
+    )
+    assert result.returncode != 0
+
+
+def test_launchers_provision_and_preserve_power_provenance():
+    for launcher in set(LAUNCHERS.values()):
+        text = launcher.read_text(encoding="utf-8")
+        assert f'DCGM_EXPORTER_IMAGE="{EXPORTER_IMAGE}"' in text
+        assert (
+            'DCGM_EXPORTER_ENROOT_REF="${DCGM_EXPORTER_IMAGE/nvcr.io\\//nvcr.io#}"'
+            in text
+        ), launcher
+        assert (
+            'import_squash "$DCGM_EXPORTER_SQSH" "$DCGM_EXPORTER_ENROOT_REF"' in text
+            or 'enroot import -o \\"$DCGM_EXPORTER_SQSH\\" \\"docker://$DCGM_EXPORTER_ENROOT_REF\\"'
+            in text
+        ), launcher
+        assert 'test -r "$DCGM_EXPORTER_SQSH"' in text
+        assert (
+            'sha256sum "$DCGM_EXPORTER_SQSH" > "$GITHUB_WORKSPACE/exporter-image.sha256"'
+            in text
+        )
+        assert 'dcgm-exporter: ${DCGM_EXPORTER_SQSH}" srtslurm.yaml' in text
+        assert 'grep -q "^  dcgm-exporter: " srtslurm.yaml ||' in text
+        assert (
+            'cp "$GITHUB_WORKSPACE/exporter-image.sha256" "$LOGS_DIR/power/exporter-image.sha256"'
+            in text
+        )
+        assert (
+            'cp "$GITHUB_WORKSPACE/power-producer-sha.txt" "$LOGS_DIR/power/power-producer-sha.txt"'
+            in text
+        )
+
+
+def test_non_power_specialized_clone_revisions_remain_available():
+    expected = {
+        "runners/launch_b200-dgxc.sh": {
+            "aflowers/vllm-gb200-v0.20.0",
+            "c180328b98c3793ca84a1e24a030f90545eb7d5d",
+            "df5baa93f4caf5169dea2a4236ad2cc742fe40e7",
+        },
+        "runners/launch_b200-nscale-slurm.sh": {
+            "04e87fcc505d6d851451781a5499ca19a02ec2b4",
+            "aflowers/vllm-gb200-v0.20.0",
+            "c180328b98c3793ca84a1e24a030f90545eb7d5d",
+            "217f9438",
+        },
+        "runners/launch_b300-nv.sh": {
+            "aflowers/vllm-gb200-v0.20.0",
+            "c180328b98c3793ca84a1e24a030f90545eb7d5d",
+            "cam/sa-submission-q2-2026",
+        },
+    }
+    for relative, revisions in expected.items():
+        text = (REPO_ROOT / relative).read_text(encoding="utf-8")
+        assert revisions <= {revision for revision in revisions if revision in text}, (
+            relative
+        )
+
+
+def test_process_result_ci_covers_the_new_contract():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    for path in (
+        "configs/nvidia-master.yaml",
+        "runners/launch_b200-dgxc.sh",
+        "runners/launch_b200-nscale-slurm.sh",
+        "runners/launch_b300-nv.sh",
+        "utils/test_b200_b300_power_official_contract.py",
+    ):
+        assert f"- '{path}'" in workflow
+    assert (
+        "test_b200_b300_power_official_contract.py"
+        in workflow.split("- name: Run pytest", 1)[1]
+    )
+
+
+def test_process_result_ci_globs_cover_every_selected_recipe():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    covered = set()
+
+    for recipe_glob in PROCESS_RESULT_RECIPE_GLOBS:
+        assert f"- '{recipe_glob}'" in workflow
+        result = subprocess.run(
+            ["git", "ls-files", f":(glob){recipe_glob}"],
+            cwd=REPO_ROOT,
+            capture_output=True,
+            text=True,
+            check=True,
+        )
+        covered.update(result.stdout.splitlines())
+
+    assert set(POWER_RECIPES) <= covered