Skip to content

MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX: tuned config-only curve on nightly-49cb2bbf / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-49cb2bbf 的纯配置调优曲线 - #4357

Merged
edwingao28 merged 5 commits into
mainfrom
minimaxm3-b300-fp4-vllm-agentic-config
Oct 26, 2026
Merged

edwingao28 merged 5 commits into
mainfrom
minimaxm3-b300-fp4-vllm-agentic-config

Conversation

@hshrivastava-droid

@hshrivastava-droid hshrivastava-droid commented Oct 20, 2026 •

Copy link
Copy Markdown
Collaborator

Description

Config-only update for minimaxm3-fp4-b300-vllm-agentic-mtp.

  • Config: minimaxm3-fp4-b300-vllm-agentic-mtp
  • Image: vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:8c1434bfee7d7626eef78ef70600434ffba4a1b8813cb1d1c8707ad25f47035e

Type of Change

  • Bug fix
  • New feature
  • Configuration change
  • Documentation update
  • Other (please describe)

Checklist

  • I have tested my changes locally
  • I have updated documentation if necessary
  • If I changed a container image or config, I have already updated perf-changelog.yaml
    • New perf-changelog.yaml entries are appended to the end of the file (the file is chronological: oldest at top, newest at bottom)
  • Before merging via reuse, an authorized maintainer (OWNER/MEMBER/COLLABORATOR) has commented /reuse-sweep-run on this PR — do this only once there is a final full sweep that is all green with evals passing, since after this comment the sweep label will no longer automatically kick off new sweeps (remove and re-add the label to force one)

@hshrivastava-droid
hshrivastava-droid requested a review from a team October 20, 2026 20:15
@hshrivastava-droid hshrivastava-droid added the full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures label Oct 20, 2026
…htly-ac68c308 / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-ac68c308 的纯配置调优曲线
@hshrivastava-droid
hshrivastava-droid force-pushed the minimaxm3-b300-fp4-vllm-agentic-config branch from 1a6f2c7 to b7c0601 Compare October 20, 2026 20:16

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I reviewed this PR and didn't find any bugs. Both the recipe and the master config are covered by NVIDIA CODEOWNERS, and the tuning picks are judgment calls, so it's worth having an owner take a look too.

What was reviewed:

  • Recipe override names against the new conc-lists (TP8 c1, TP4 c1/c24, TP4 DRAM c36/c40). Each point has exactly one override, and no overrides are left over after the TP2 and c48/c64 removals.
  • The new perf-changelog entry is appended at the end and leaves existing bytes untouched. It names the config key, and the image tag matches the master config.
  • I looked at max-num-seqs: 7 on the TP4 c24 point, which is far below CONC 24 and well under the ~1.5x ratio the other points use. The hunt ruled it out as a bug, but an owner should confirm it's the intended tuning. max-num-active-seqs (c36/c40) is also used by no other recipe in the repo, so it depends on the new nightly image supporting it.
Extended reasoning...

The change is config-only: it bumps the digest-pinned vLLM nightly image for minimaxm3-fp4-b300-vllm-agentic-mtp, shrinks the search space, retunes the matching srt-slurm AgentX recipe overrides and appends one perf-changelog entry. It touches no security-sensitive surface. I deferred instead of approving because nvidia-master.yaml and the b300-* recipe directories have dedicated CODEOWNERS. The new values (max-num-seqs below CONC at c24, and a max-num-active-seqs arg no other recipe uses) are tuning choices that the owners are better placed to confirm.

This review covers commit 1a6f2c7, which is no longer the latest commit on this pull request; later commits are not covered by it.

@github-actions

github-actions Bot commented Oct 21, 2026 •

Copy link
Copy Markdown
Contributor

@github-actions

github-actions Bot commented Oct 23, 2026 •

Copy link
Copy Markdown
Contributor

@hshrivastava-droid staged run 39490622498.

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

@Ankur-singh

Ankur-singh commented Oct 23, 2026 •

Copy link
Copy Markdown
Collaborator

As a PR reviewer and CODEOWNER, I have reviewed this and have:

  • Verified that as of the moment of typing this, this is the latest version of PR_REVIEW_CHECKLIST.md
  • Verified that the general code quality meets the InferenceX standard and does not make the code quality any worse.
  • Verified that this PR has passed PR validation. Please link to GitHub Action workflow that shows this.
  • Verified that this PR passes evals. Please link to GitHub Action workflow that shows this.
  • Verified that speculative decoding PRs uses chat templates to align the AL distribution to real world
  • Verified that every draft model and draft head is served as it ships: the draft that ships with the served checkpoint, at its stored precision, through the pinned upstream image's default handling, with the shipped and effective draft precision recorded in the additional detail section. No submission-side quantization, dtype override, checkpoint substitution, or patch may lower draft precision below that default, regardless of eval results or AL. Explicitly verified that SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is not enabled in the effective recipe, including inherited settings; enabling it is prohibited going forward, and historical runs do not grant an exception. See Draft-model precision for what counts as the default and the MLPerf comparison.
  • For agentic workloads: verified that speculative-decoding configs (EAGLE / MTP / draft models) run with simulated synthetic acceptance, with the acceptance-length value taken from the committed golden AL curve in infx/golden_al_distribution/ for that model, thinking mode, and draft length. A submission may choose any supported draft length, but it may not substitute a different acceptance target.
  • Verified against the current MODELS.md that this PR does not submit a deprecated model, scenario, or model-scenario combination.
  • Verified that the model architecture isn't changed with benchmark hacks like using --hf-overrides to skipping indexer for every x layers on models that don't natively support this. As a general rule, we won't accept optimizations that reduces the number of model architecture FLOPs. Anything that makes that same computation run faster is fair game; target/verifier FLOPs at lower precisions is fine, given that the config passes private evals, but this does not permit lowering draft-model or draft-head precision below what ships. As an general north star princple, we should only use optimizations which is used in production by customers that care about accuracy
  • For affected Dynamo TRT-LLM benchmarks, verified that iteration/KV-cache metrics are collected and published under the pinned Dynamo implementation and effective engine settings. Resolve the installed Dynamo version separately from the container tag; use supported metrics controls without requiring KV events. Record the version and configuration evidence in the additional detail section. See verifier Check 17. N/A for unaffected serving paths or frameworks without these Dynamo settings.
  • If an company claims that they support vLLM/SGLang as first class LLM inference engines on their hardware, I have verified that the respective vLLM submission made using upstream https://hub.docker.com/u/vllm docker repo, upstream SGLang https://hub.docker.com/u/lmsysorg docker repo. The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet as supported by vLLM/SGLang community maintainers
  • If an company claims that they support vLLM/SGLang as first class upstream in-tree LLM inference engines on their hardware, I have have verified that the respective vLLM/SGLang submission has been made before additional frameworks (TRT-LLM, ATOM, etc.). The only exceptions are for new hardware, such as MI455X UALoE72, Vera Rubin NVL72, Rubin NVL8, etc., and for new model architectures where there is an actual reason why vLLM/SGLang does not fundamentally support them yet.
  • Verified that every affected vLLM/SGLang production serving configuration in this PR, including Dynamo deployments, is documented in the official vLLM recipes and/or the SGLang cookbook, for both single-node and multi-node aggregated and disaggregated serving. The documentation covers production-useful flags, environment variables, and launch commands (including prefill, decode, and router/frontend commands for disaggregation); InferenceX-only benchmark/harness settings are excluded with reasoning in the additional detail section:
    • I linked the corresponding upstream PR in the vLLM recipe repo or SGLang cookbook source repo and verified that it is MERGED before this InferenceX PR merges. An opened, draft, or closed-without-merge upstream PR does not satisfy this requirement. If the matching recipe was already published, I linked the published recipe/cookbook page in the additional detail section below.
  • Verified that this PR does not patch the inference engine or serving stack — the pinned image must run as shipped. This covers .patch files / git apply / patch, inline patches embedded in benchmark scripts (e.g. a python3/sed heredoc that rewrites installed engine sources before serving), in-place edits of site-packages, monkey-patching, overwriting container files, and installing forked/rebuilt engine wheels on top of the pinned image. The only exception is a patch covered by a filled-out waiver at docs/waiver/<PR_NUMBER>.md — named after the PR that introduces the patch and filed in that same PR, stating what is patched, why the unmodified upstream image cannot run this benchmark, the upstream PR/issue link, and the removal plan — which I have linked below in the additional detail section.
  • If this PR uses append-only: true, verified that it only adds generated points or recipe variants inside a selected existing config/scenario and existing same-image visual curve: every previously generated point remains present with the same recipe, no prior point is removed or rerun, and every benchmark-affecting change in the complete diff can affect only the corresponding newly appended points (never an existing point), regardless of which file contains it.
  • If any of the above criteria cannot reasonably be satisfied, I have provided additional reasoning below.
  • Reported measured throughput/E2EL Pareto counts and evidence per affected curve (≥5 points strongly recommended). Below 5 or unverifiable: tag a core maintainer for review; recorded admin bypass required before merge. N/A if no curves are affected. Details.

Additional detail section:

  • Assessed head: 8c8d778989c172fbdd16590a0b80b1493fd260f1 (policy checked at InferenceX main @ 9f500761). This replaces my earlier assessment of b7c06012: the PR now carries the re-tuned curve. Config minimaxm3-fp4-b300-vllm-agentic-mtp, vLLM single-node aggregated on B300, nvidia/MiniMax-M3-NVFP4 with the Inferact/MiniMax-M3-EAGLE3-GQA draft (3 tokens). Image vllm/vllm-openai:nightly-49cb2bbf71cfd36ee5aa33d1e184b1cddff55294@sha256:8c1434bfee7d7626eef78ef70600434ffba4a1b8813cb1d1c8707ad25f47035e (upstream; the digest's ai.vllm.build.commit label and the run logs v0.32.0rc1.dev281+g49cb2bbf7 agree). Points: TP8 c1, TP4 c20 (--max-num-active-seqs 6, --max-num-seqs 30), TP4 c24 (--max-num-seqs 7, VLLM_FLASHINFER_ALLREDUCE_BACKEND=mnnvl), TP2 c24 and c28 with 749 GB lazy SimpleCPUOffloadConnector, --gpu-memory-utilization 0.92 and minimax_m3_msa_decode_backend: cutlass.
  • PR validation + evals: run 40022391822 (attempt 1, success, head_sha = assessed head): all 5 benchmark jobs and all 6 eval jobs succeeded. Evals (eval_results_all, infrastructure_success: true on every row): minimax_m3_full em_strict 0.99 on all 5 points (n=102); gsm8k em_strict 0.9659 at TP2 c28 (n=1319). The TP2 points' minimax_m3_msa_decode_backend: cutlass ran in the evals: the TP2 c24 minimax_m3_full, TP2 c28 minimax_m3_full and TP2 c28 lm-eval eval server logs (eval_server_logs_* artifacts of this run) each print [sparse_attention_msa.py:232] MiniMax M3 MSA sparse decode selected CUTLASS (line 65 of the *_agg_w0.out worker log), after MiniMax M3 sparse attention selected MSA (kv_cache_dtype=fp8, topk_blocks=16). /use 40022391822 was posted by @hshrivastava-droid (COLLABORATOR) and staged; it supersedes the earlier /use 39490622498, which named the previous head's run.
  • Golden AL: minimaxm3_eagle3_gqa.yaml, thinking_on, 3 draft tokens = 2.78. All 5 points' server configs show rejection_sample_method: synthetic with synthetic_acceptance_length: 2.78 and num_speculative_tokens: 3.
  • Chat template: AgentX replays chat-message traces against /v1/chat/completions.
  • Draft precision: the EAGLE3 draft is Inferact/MiniMax-M3-EAGLE3-GQA (a separate repo; the served checkpoint ships no draft or MTP head), served at its stored BF16 precision. Shipped: BF16, no quantization_config (as verified for the same snapshot on [NVIDIA][AgentX] MiniMax-M3 B200 vLLM digest pin, TP4 DRAM-offload tuning and capped-active c10 entries / [NVIDIA][AgentX] MiniMax-M3 B200 vLLM 按摘要固定镜像、调优 TP4 DRAM 卸载并新增限制活跃序列的 c10 配置 #4382). Effective at vLLM 49cb2bbf: vllm/config/speculative.py copies the target quantization config only in the native-MTP / same-model branches (e.g. MiniMax-M3 MTP), not for a separate EAGLE3 draft; Eagle3LlamaForCausalLM takes its quant_config from get_draft_quant_config (the draft's own config, none); draft_lm_head_quantization is not set. Runtime: all 5 server logs show Resolved architecture: LlamaForCausalLMEagle3 with a single 5.73 GiB checkpoint (the BF16 size) beside the 232.93 GiB target. Source trace plus logs, not a per-module dtype dump. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is SGLang-only and absent.
  • Upstream recipe (Check 3): the B300 lane at this head is documented on vLLM main by merged vllm-project/vllm#63407 (moved B300 to this curve) and #63534 (the b300_agentx preset): image 49cb2bbf, all five points with their --max-num-seqs, active-sequence caps, memory utilization, all-reduce backend, capture sizes, the 749 GB lazy offload and CUTLASS MSA decode on TP2. A 135-row comparison against the PR-head recipe and the run's logged launch commands found the preset omits only --enable-chunked-prefill, which is the default here: at vLLM 49cb2bbf, EngineArgs._set_default_chunked_prefill_and_prefix_caching_args (vllm/engine/arg_utils.py) sets it from ModelConfig.is_chunked_prefill_supported, which is True for this generative decoder-only model, and the run logs show enable_chunked_prefill=True. vllm-project/vllm#63810 (open) makes the flag explicit in the preset and corrects a stale autotune-flag note. InferenceX-only settings excluded: concurrency lists, KV_OFFLOADING/TOTAL_CPU_DRAM_GB labels, synthetic acceptance, AIPerf, --served-model-name, PYTHONNOUSERSITE, health checks and the TP2 c28 Slurm time limit.
  • No engine patch: the diff changes the recipe overrides, one master-config entry and the changelog; no setup script, patch, sed or pip step.
  • Not deprecated (MiniMax-M3 agentic coding is active in MODELS.md); no architecture hacks; append-only N/A; Dynamo TRT-LLM metrics item N/A (vLLM engine).
  • Pareto (run 40022391822, raw bmk_agentic_*, request_metrics.throughput.per_gpu.total_tput_tps vs latency.e2el.p90, upper-right frontier): 4 of 5 points on the frontier: TP4 c20 (34.8k tput/GPU, 12.25 s), TP4 c24 (42.1k, 13.99 s), TP2 c24 (73.0k, 21.51 s), TP2 c28 (74.3k, 29.22 s). TP8 c1 (3.1k, 12.50 s) is dominated by TP4 c20. Below the 5-point recommendation, so a core maintainer should review and an admin bypass is needed before merge. Local computation over this run only.
  • TTFT (p50 / p90): TP8 c1 0.90 / 1.49 s; TP4 c20 0.67 / 3.15 s; TP4 c24 0.89 / 4.91 s; TP2 c24 0.57 / 1.70 s; TP2 c28 0.72 / 4.71 s. TP4 c20 runs with active sequences capped at 6 below concurrency 20 and TP4 c24 with --max-num-seqs 7, so part of their TTFT is queueing by design.
  • Not assessed in this edit: Check 18(g) (whether any point removed relative to the published curve was on the frontier and is not dominated by a head point). The PR keeps 5 points; the curve on main has more.
  • Warnings: the PR is CONFLICTING with main in perf-changelog.yaml only; the reuse merge script resolves that file. Check 16 (NVIDIA overview): MiniMax-M3 is listed with vLLM and links recipes.vllm.ai/MiniMaxAI/MiniMax-M3.

Signed: ankur-singh

@github-actions

github-actions Bot commented Oct 23, 2026 •

Copy link
Copy Markdown
Contributor

❌❌❌ REJECTED ❌❌❌

⚠️ Pareto coverage needs additional review: @functionstackx @cquil11 @Oseltamivir @adibarra. At least 5 points per affected throughput-versus-E2EL frontier are highly recommended. Below 5, or when coverage cannot be verified, merge only with an explicit, recorded admin bypass for the assessed commit; this advisory comment does not grant or enforce a bypass.

@Ankur-singh Blocking: the PR still drops six points that were on the published frontier, and no head point dominates them (18g, which the sign-off left unassessed). Separately, this verifier could not read the runtime log that would show the TP2 CUTLASS MSA decode path ran in the evals (2c).

❌ Check 2 (Evals pass): FAIL — The TP2 CUTLASS MSA decode path could not be independently verified as having run in the evals. Accuracy is fine: in run 40022391822 at 8c8d778, minimax_m3_full scores 0.99 on all 5 points and gsm8k 0.966 at TP2 c28. Every row is ok against its siblings and the exact-topology cohort (run 38481970789, run 29984401390). The TP2 c24 eval job log (job 120650372648) shows only the configured minimax_m3_msa_decode_backend: cutlass flag. The cited runtime line [sparse_attention_msa.py:232] MiniMax M3 MSA sparse decode selected CUTLASS is inside the eval_server_logs_* tarballs, which this verifier was not permitted to extract. The cited evidence is specific; rerun verification in an environment that can read the tarball, or have a core maintainer confirm the line.

❌ Check 18 (Recipe/image/sweep consistency): FAIL — (g) Six removed points were on the published frontier, and no head point dominates them in tput_per_gpu and mean_intvty. The removed points are TP4 c1 (5,526 / 398.3), c5 (9,128 / 353.6), c10 (17,161 / 333.8), c15 (23,127 / 277.3), c30-dram (48,186 / 164.5) and c40-dram (56,831 / 124.2). The measurements come from raw bmk_agentic_* results in run 38481970789 (image nightly-5555a2d8, source 6dbbfe26). The closest head points are TP8 c1 (3,104 / 497.4), TP4 c20 (34,833 / 261.4) and TP2 c24 (72,965 / 122.2). The head's TP2 c24 dominates TP4 c48/c64 and TP2 c32/c40. Keep or re-measure the low-concurrency TP4 and TP4 DRAM points, or show head points that dominate them. (a)-(f) pass: there are no generation errors, master and recipe overrides agree, the digest resolves, no orchestration pin or cross-repository dispatch is touched, and the changelog entry is accurate. (h) Against the old image, TP8 c1 gains +7.4% throughput and +14.6% interactivity, TP4 c20 -1.5% and +14.0%, and TP2 c24 +7.0% and +20.8%.

⚠️ Check 14 (Pareto coverage): WARN — The curve minimaxm3 / agentic-coding / B300 vLLM / fp4 / p90 E2EL has 4/5 frontier points in run 40022391822 (bmk_agentic_*, total_tput_tps vs e2el.p90): TP4 c20 (12.25 s, 34.8k), TP4 c24 (13.99 s, 42.1k), TP2 c24 (21.51 s, 73.0k) and TP2 c28 (29.22 s, 74.3k). TP8 c1 (12.50 s, 3.1k) is dominated. Admin bypass not verified; an explicit admin bypass for 8c8d778 and this curve is required.

Passed and not applicable checks

✅ Check 0 (CODEOWNER): PASS — @Ankur-singh is a listed owner of inferencex-e2e/configs/nvidia-master.yaml and of /inferencex-e2e/benchmarks/*_node/srt-slurm-recipes/**/b300-*/. perf-changelog.yaml falls only under the catch-all.

✅ Check 1 (In-PR sweep + evals): PASS — Pinned head 8c8d778 is still the PR tip and has successful agentic / (5) and agentic eval / (5) check-runs from run 40022391822. There are no later commits to assess.

✅ Check 3 (Upstream recipe): PASS — Merged vllm#63407 and vllm#63534 publish the b300_agentx preset and the B300 nvfp4 block. They match the image, all 5 points, per-point max-num-seqs and active caps, all-reduce backends, capture sizes, the 749 GB lazy offload and TP2 CUTLASS decode. The omitted --enable-chunked-prefill is the default at 49cb2bbf, and the sign-off records that evidence.

✅ Check 4 (Reuse command): PASS — hshrivastava-droid (COLLABORATOR) posted /use 40022391822 on 2026-10-26; it names the Check 1 run.

✅ Check 5 (Checklist template): PASS — Every item in the current template is present and checked.

✅ Check 6 (Upstream image/ordering): PASS — framework: vllm uses upstream vllm/vllm-openai:nightly-49cb2bbf...@sha256:8c1434bf.... This is an open-source engine entry, so the ordering rule does not apply.

✅ Check 7 (Deprecated models): PASS — MiniMax-M3 (minimaxm3) agentic coding is active in MODELS.md on 2026-10-26.

✅ Check 8 (No architecture hacks): PASS — The PR adds no --hf-overrides, JSON model overrides or model-file edits. It adds no new accuracy-trading levers; the FP8 KV cache is unchanged from the base recipe and every point has eval coverage.

✅ Check 9 (Chat templates): PASS — AgentX replays chat-message traces against /v1/chat/completions.

✅ Check 10 (No engine patches): PASS — The diff changes only recipe overrides, one master entry and the changelog. There is no setup script, patch, sed or pip step on the reachable path.

✅ Check 11 (Golden AL): PASS — The benchmark launch passes "rejection_sample_method": "synthetic", "synthetic_acceptance_length": 2.78 with num_speculative_tokens: 3. This matches minimaxm3_eagle3_gqa.yaml thinking_on level 3 (2.78).

✅ Check 13 (Draft precision): PASS — The separate draft Inferact/MiniMax-M3-EAGLE3-GQA ships as BF16 (torch_dtype: bfloat16, no quantization_config). The recipe has no draft quantization or dtype override, and the FP8 KV cache applies to target and draft alike. SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE is SGLang-only and absent.

✅ Check 15 (PR description): PASS — The PR body names config minimaxm3-fp4-b300-vllm-agentic-mtp and image nightly-49cb2bbf...@sha256:8c1434bf..., matching the pinned master entry.

✅ Check 16 (NVIDIA overview): PASS — The NVIDIA Framework Recipes table lists MiniMax-M3 with a vLLM entry linking recipes.vllm.ai/MiniMaxAI/MiniMax-M3.

➖ Check 12 (Append-only): N/A — The new changelog entry has no append-only: true.

➖ Check 17 (Metrics publication): N/A — The engine is vLLM, not Dynamo TRT-LLM.

Assessed commit: 8c8d778989c172fbdd16590a0b80b1493fd260f1.

@hshrivastava-droid hshrivastava-droid added full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures and removed full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures labels Oct 23, 2026
@hshrivastava-droid

Copy link
Copy Markdown
Collaborator Author

/use 40022391822

@github-actions

github-actions Bot commented Oct 26, 2026 •

Copy link
Copy Markdown
Contributor

@hshrivastava-droid staged run 40022391822.

This run remains available across future /stage-results requests. Staging the same run ID again updates its staged data. Staging workflow

@SemiAnalysisAI SemiAnalysisAI deleted a comment from github-actions Bot Oct 26, 2026
@Ankur-singh Ankur-singh changed the title MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX: tuned config-only curve on nightly-ac68c308 / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-ac68c308 的纯配置调优曲线 MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX: tuned config-only curve on nightly-49cb2bbf / MiniMax-M3 NVFP4 B300 fp4 vLLM AgentX:基于 nightly-49cb2bbf 的纯配置调优曲线 Oct 26, 2026
@edwingao28
edwingao28 merged commit 4624c17 into main Oct 26, 2026
27 checks passed
@edwingao28
edwingao28 deleted the minimaxm3-b300-fp4-vllm-agentic-config branch October 26, 2026 21:46
@github-actions

Copy link
Copy Markdown
Contributor

Ingest report — ✅ Results ingested into the production database.

Ingest workflow · Merge Ingest run · Benchmark source run

Category new duplicate skipped unmapped failed
benchmark 5 0 0 0 0
eval 6 6 6 0 0
trace 5 0 0 0 0
power 5 0 0 0 0

benchmarkFiles 5, rawBenchmarkRows 5, coveredByPerJobFiles 0, purgedPoints 0, failedRunRows 0, runStatsNew 0, runStatsDuplicate 0, evalSamplesNew 1319, gpuMetricSamplesNew 101884, powerPublicationPoints 5, changelogEntriesWritten 1

Counts cover only this run. Rows: benchmark_results, eval_results, agentic rows linked to a trace, gpu_metrics series. This comment updates if the ingest is rerun.

中文

Ingest 报告 — ✅ 结果已写入 production 数据库。

上表仅统计本次运行写入的数据,各行单位依次为 benchmark_results 行、eval_results 行、已关联 trace 的 agentic 行和 gpu_metrics 序列。重新运行 ingest 时本评论会原地更新。

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-enabled Full sweep with canary gate; matrix jobs run to completion despite failures

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

3 participants