[PowerX] use profiling-free counters for H200 Kimi-K3 / [PowerX] H200 Kimi-K3 使用无 profiling 计数器

inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/kimik3/vllm/h200-fp4/agentx/agg-tp16dp2ep32-latency.yaml

@@ -106,5 +106,9 @@ benchmark:
     IS_MULTINODE: "true"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    # Long responses admitted near the end of the measurement window can take
+    # minutes to finish; bound their drain without extending admission.
+    # Default 30s grace cancels mid-response and fails ProfileMetricCoverage (~94%).
+    AIPERF_BENCHMARK_GRACE_PERIOD: "1800"
     HF_HUB_CACHE: "/hf_hub_cache"
     WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"

inferencex-e2e/configs/nvidia-master.yaml

@@ -1407,6 +1407,7 @@ kimik3-fp4-h200-vllm-agentic-latency:
     - dram-utilization: 0.80
       search-space:
       # Latency-oriented: TP16 x DP2 with EP32 across four H200 nodes.
+      # Power telemetry omits DCGM profiling counters to avoid watch-repair stalls.
       - spec-decoding: mtp
         kv-offloading: none
         conc-list: [1, 2, 3, 4, 5, 6, 7, 8, 10, 12]

inferencex-e2e/perf-changelog.yaml

@@ -9168,3 +9168,25 @@
     - "Relevant ATOM changes in the range: V4 decode reuses the sparse prefill ASM (ROCm/ATOM#2271), greedy sampler picks via aiter.topk_select (ROCm/ATOM#2244), and FP8 block scales declared scale_fmt ue8m0 are now stored as E8M0 on gfx950 by default (ROCm/ATOM#2419; previously FP32 unless ATOM_FP8_BLOCKSCALE_USE_E8M0_SCALE=1). The upstream DeepSeek-V4 recipes are unchanged across the range, and every recipe flag and choice (all2all-backend rccl, dp-load-balance least_tokens, moe-backend standard) remains valid."
     - "No data-type or precision change to the DeepSeek-V4-Pro-0813 DSpark draft: no online quantization is configured, so it keeps its checkpoint precision; the E8M0 scale storage represents the checkpoint's power-of-two block scales exactly. kv-cache-dtype and index-cache-dtype touch cache storage only."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3605
+
+- config-keys:
+    - kimik3-fp4-h200-vllm-agentic-latency
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use profiling-free DCGM counters for Kimi-K3 H200 latency power telemetry to avoid synchronous profiling-watch recovery stalls. Retain power and GPU utilization; omit profiling metrics."
+    - "Keep the stock exporter image, 100 ms watch interval, 1 s scrape interval, 2 s timeout, and strict power validation unchanged."
+    - "Kimi-K3 H200 latency 功耗遥测改用无 profiling 的 DCGM counters，避免同步恢复 profiling watch 导致采样卡顿；保留功耗和 GPU 利用率，省略 profiling 指标。"
+    - "保持原 exporter 镜像、100 ms watch 间隔、1 s 抓取间隔、2 s 超时与严格功耗校验不变。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3550
+
+- config-keys:
+    - kimik3-fp4-h200-vllm-agentic-latency
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Raise AgentX benchmark grace period to 1800s on the H200 Kimi-K3 latency recipe so in-flight long responses admitted near the 3600s cutoff can finish and extend TTFT/ITL coverage past the 95% ProfileMetricCoverage gate without changing admission duration, concurrency, or topology."
+    - "Matches the stock srt_agentic.sh AIPERF_BENCHMARK_GRACE_PERIOD wiring and the MI325X GLM-5.2 agentic precedent; leaves the one-hour profile window and noprof power path unchanged."
+    - "将 H200 Kimi-K3 latency AgentX 的 benchmark grace 提到 1800s，使接近 3600s 截止时已准入的长响应可完成并延续 TTFT/ITL 覆盖至 ≥95%，不改变 admission 时长、并发或拓扑。"
+    - "对齐 srt_agentic.sh 的 AIPERF_BENCHMARK_GRACE_PERIOD 接线与 MI325X GLM-5.2 agentic 先例；保持一小时 profile 窗口与 noprof 功耗路径不变。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3550