[DO NOT MERGE] Run-only: gb300 dsr1 measured power+temp validation#1686
[DO NOT MERGE] Run-only: gb300 dsr1 measured power+temp validation#1686arygupt wants to merge 5 commits into
Conversation
Standalone branch off main to RUN the gb300 dsr1 perfmon validation without waiting for #1574 to merge. Carries #1574's self-contained consumer code (utils/aggregate_power.py, utils/process_result.py + tests) so the Process-result step patches the agg JSON, plus the gb300-nv launcher perfmon wiring and a single 1-job changelog entry (dsr1-fp4-gb300-dynamo-sglang-powercheck). Power AND temperature ride the same pipeline: the srt-slurm fork perfmon captures power.draw + temperature.gpu + utilization.gpu + memory.used; aggregate_power.py emits avg_power_w, per-stage prefill/decode power, and per-worker avg_temp_c/peak_temp_c/avg_util_pct/avg_mem_used_mb. DO NOT MERGE ahead of #1574 — this duplicates #1574's aggregate_power.py. It exists to produce the measured dsr1-gb300 data now; close after the data lands. Production landing happens via the clean post-#1574 path. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook If it is not, please create a PR first before we can merge your single node PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers. If additional help is needed, PR authors can reach out to core maintainers over Slack. |
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
|
|
||
|
|
||
| # Back-compat shim — some external callers may have imported _parse_power. | ||
| _parse_power = _parse_numeric_cell |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
| perf_csv_count=$(ls "$LOGS_DIR"/perf_samples_*.csv 2>/dev/null | wc -l | tr -d ' ') | ||
| if [ "$perf_csv_count" -gt 0 ]; then | ||
| mkdir -p "$GITHUB_WORKSPACE/perf_samples" | ||
| cp "$LOGS_DIR"/perf_samples_*.csv "$GITHUB_WORKSPACE/perf_samples/" |
There was a problem hiding this comment.
Stale perf CSVs not cleared
Medium Severity
Measured-power staging copies into $GITHUB_WORKSPACE/perf_samples/ without clearing that directory first. process_result.py globs every perf_samples_*.csv there, so files left from an earlier job on the same runner can be aggregated with the current run and skew avg_power_w, worker counts, and joules fields.
Reviewed by Cursor Bugbot for commit 9839a0c. Configure here.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27159252858 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27162265261 |
…gb300-cw)
Retry on gb300-cw_2 failed instantly: launch_gb300-cw.sh hard-rejects dsr1 ('Unsupported model prefix/precision combination on gb300-cw: dsr1/fp4'). The bare 'gb300' runs-on label is shared across the NVIDIA (gb300-nv_*) and CoreWeave (gb300-cw_*) pools, and only launch_gb300-nv.sh handles dsr1+perfmon. Pinning to the gb300-nv label keeps it on the NVIDIA pool.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
| echo "[perfmon] staged $perf_csv_count per-node perf_samples_*.csv to \$GITHUB_WORKSPACE/perf_samples/" | ||
| else | ||
| echo "[perfmon] WARNING: monitoring enabled but no perf_samples_*.csv found in $LOGS_DIR — measured power aggregation will be skipped" >&2 | ||
| fi |
There was a problem hiding this comment.
Perfmon failure omits glob guard
Medium Severity
When PERFMON_ENABLED is set but no perf_samples_*.csv files are found, the launcher logs a warning and does not set GPU_METRICS_CSV_GLOB. process_result.py then uses the single-node gpu_metrics.csv fallback, which can patch multinode agg JSON with stale single-node power instead of omitting telemetry.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 5b3eb12. Configure here.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27162908884 |
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27163380777 |
…eave) The gb300-nv runner fleet wedged on its pre-run cleanup (stale NFS), so pivot the validation to the healthy CoreWeave GB300 pool. Adds a dsr1/fp4 branch to launch_gb300-cw.sh that clones the perfmon fork (reusing its recipes/gb300-fp4 dsr1 recipes), maps model.path 'dsr1' -> /mnt/vast/models/dsr1-fp4 (weights pre-staged on CW), injects monitoring: into the recipe, and stages perf_samples_*.csv + GPU_METRICS_CSV_GLOB for Process-result. Config runner gb300-nv -> gb300-cw. CoreWeave-segment data (proxy for the NVIDIA campaign numbers) — validates the full perfmon -> aggregate_power -> power+temp pipeline end-to-end today while the gb300-nv fleet is fixed separately. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
There are 3 total unresolved issues (including 2 from previous reviews).
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit ce2c4ba. Configure here.
| echo "[perfmon] WARNING: zero recipe YAMLs found under recipes/ — power data will be MISSING from this run." >&2 | ||
| else | ||
| echo "[perfmon] injected monitoring: into $INJECTED_COUNT of $FOUND_COUNT recipes." | ||
| fi |
There was a problem hiding this comment.
dsr1 branch ignores framework
High Severity
The new dsr1 clone path keys only on MODEL_PREFIX, so any GB300-NV job with model-prefix: dsr1—including dynamo-trt configs—checks out the perfmon srt-slurm fork, enables PERFMON_ENABLED, and rewrites every recipe YAML. TRT (and other non–dynamo-sglang) dsr1 runs previously used the default sa-submission-q2-2026 checkout and can fail or run the wrong stack.
Reviewed by Cursor Bugbot for commit ce2c4ba. Configure here.
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27165970072 |
…hfs OOM) First Path B run (job 6867) failed: pyxis killed unsquashfs (exit 137) extracting the container on the prefill node -> etcd never started -> decode workers crashed with 'Could not connect to etcd'. Root cause: the fork's gb300-fp4 recipe omits sbatch_directives.mem, so SLURM capped node memory and the cgroup OOM-killed the ~15-30GB squashfs extraction. Every working CW recipe (dsv4/glm5) sets mem: "0"; inject it (idempotent) alongside monitoring:. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27167819530 |
… > 30min) Run 6870 reached full server bringup but timed out: 'Model did not get healthy in 1800 seconds' while decode was still capturing CUDA graphs (25%, bs=184). dsr1 warmup = load 671B weights + FlashInfer autotune + capture 36 graph sizes (cuda-graph-max-bs 256) ~= 35min, over the default health_check (max_attempts 180 x interval 10 = 1800s). On timeout the orchestrator killed etcd -> lease errors -> teardown (not a crash). Inject health_check max_attempts=540 interval=10 (5400s/90min). It's a ceiling; the sweep proceeds as soon as servers are healthy. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
|
see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=27170607165 |
1 participant
Purpose: produce the first measured power + temperature data for
dsr1disagg on GB300 now, without waiting for #1574 to merge.This branch carries #1574's self-contained consumer code (
utils/aggregate_power.py,utils/process_result.py+ their tests) on top ofmain, plus thegb300-nvlauncher perfmon wiring and a single 1-job changelog entry (dsr1-fp4-gb300-dynamo-sglang-powercheck). The changelog diff expands to exactly one gb300 job (1k/1k, conc 8, 1×prefill TP4 + 2×decode TP4) — verified withprocess_changelog.py --trim-conc.Power AND temperature ride the same pipeline: the srt-slurm fork perfmon captures
power.draw+temperature.gpu+utilization.gpu+memory.used;aggregate_power.pyemitsavg_power_w, per-stageprefill_avg_power_w/decode_avg_power_w, and per-workeravg_temp_c/peak_temp_c/avg_util_pct/avg_mem_used_mb.Why not merge: it duplicates #1574's
aggregate_power.py. The production landing of thegb300-nvlauncher happens via the clean post-#1574 path (cherry-pick onto mergedmain). This PR is closed once the measured data lands.Success criteria: job green +
perf_samples_*.csvstaged + agg JSON patched with the power/temp fields above.🤖 Generated with Claude Code
Note
Medium Risk
Changes benchmark result JSON schema and multinode launcher behavior (fork checkout, recipe mutation, env-driven CSV selection); mistakes could publish wrong power numbers or break GB300 jobs, though aggregation remains best-effort and tests are broad.
Overview
Adds a GB300 measured-power validation path for DeepSeek-R1 FP4 disagg before the full NVIDIA sweep: a single-job benchmark config (
dsr1-fp4-gb300-dynamo-sglang-powercheck), changelog entry, and runner wiring ongb300-cwandgb300-nvto clone the srt-slurm perfmon fork, injectmonitoring:into recipes, stageperf_samples_*.csvviaGPU_METRICS_CSV_GLOB, and fix CoreWeave recipe gaps (unlimited mem, longer health checks).Telemetry pipeline expands
aggregate_power.pyandprocess_result.pyto merge multi-node CSVs (GPU index namespacing), parse perfmon filenames forworkers[], support disagg per-stage power/joules (prefill_avg_power_w,decode_avg_power_w,joules_per_input_token), add temp/util/memory fields, and resolve bench windows from srt-slurmdatewhen Unix timestamps are absent. Multinode runs must not fall back to stale single-nodegpu_metrics.csvwhen the glob is set.CI reliability: multinode workflow pre/post cleanup now
sudo rm -rf benchmark_logsso root-owned leftovers from cancelled jobs do not break checkout on shared runners.Extensive tests cover multinode aggregation, disagg attribution, and
process_resultglob/disagg wiring.Reviewed by Cursor Bugbot for commit 63cbe19. Bugbot is set up for automated code reviews on this repo. Configure here.