Repository navigation
config: update B300 MiniMax-M3 TRT-LLM AgentX to 1.4.0rc0 with 15 points / 配置：将 B300 MiniMax-M3 TRT-LLM AgentX 更新至 1.4.0rc0（15 个点位） - #3835
Conversation
|
Thanks for the contribution!
中文感谢你的贡献！
|
hshrivastava-droid
added
the
full-sweep-enabled
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Beyond the inline finding on the pull/TBD placeholder, I also checked two other things and ruled them out: the health-check timeout comment removal (max_attempts/interval unchanged, so no regression even though the capture-time guidance was reworded) and the override_tp4_c15 point missing a max_draft_len override (it inherits the base speculative_config.max_draft_len: 3, same as most other points, rather than being left unset).
Extended reasoning...
The diff is a config-only change (benchmark recipe, master config, changelog) with a confirmed CI-blocking issue: the new changelog entry uses pr-link pull/TBD, which validate_perf_changelog.py's placeholder allowlist (only XXX or the literal .../pull/XXX) will reject. That alone warrants not approving; I additionally checked two other candidate concerns raised during the hunt (health-check timeout margin, tp4_c15 draft length) and found both unaffected by the change.
| description: | ||
| - "Update the B300 TRT-LLM MiniMax-M3 AgentX recipe to the 13 scheduled TP2/TP4/TP8 profiles with NVFP4 KV cache, EAGLE3 speculative decoding, and native DRAM KV offload." | ||
| - "Pin the recipe and master config image to nvcr.io#nvidia/tensorrt-llm/release:1.4.0rc0.dev202610070001." | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD |
There was a problem hiding this comment.
🔴 The new changelog entry's pr-link placeholder 'pull/TBD' will fail CI validation and block merge of this PR. validate_perf_changelog.py's PR_LINK_PLACEHOLDERS only accepts the literal 'XXX' or 'https://github.com/SemiAnalysisAI/InferenceX/pull/XXX' for new entries (validate_perf_changelog.py:22-24); 'pull/TBD' matches neither that set nor the canonical pull/<pr_number> form, so validate_added_pr_link raises ChangelogValidationError. Fix: use the 'XXX' placeholder (or the real PR number) instead of 'TBD' for new entries' pr-link, matching the validator's accepted forms.
Why this was flagged
The new perf-changelog.yaml entry appended at the tail (around line 9421) sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD. validate_added_pr_link in inferencex-e2e/infx/workflows/validate_perf_changelog.py:139-150 checks a new entry's pr-link against PR_LINK_PLACEHOLDERS = {'XXX', 'https://github.com/SemiAnalysisAI/InferenceX/pull/XXX'} (line 22-24) or the exact expected pull/<pr_number> link; 'pull/TBD' is neither, so it raises ChangelogValidationError(f"new PR entry must use {expected!r} or an XXX placeholder; found {link!r}"). This validator runs from .github/workflows/run-sweep.yml:106 (python -m infx.workflows.validate_perf_changelog), so CI fails and blocks merge until TBD is replaced with XXX or the real PR number.
Verification: The appended changelog entry (perf-changelog.yaml, final line) sets pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD. Being new, compare_entries runs validate_added_pr_link on it (validate_perf_changelog.py:192-193). The link must be in PR_LINK_PLACEHOLDERS ("XXX" or the pull/XXX URL, lines 22-25) or equal the expected pull/{pr_number}; pull/TBD is neither, so ChangelogValidationError is raised (lines 148-150).
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40384083329 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40384083329 |
hshrivastava-droid
changed the title
|
/use 37814368606 |
|
@hshrivastava-droid staged run 37814368606. This run remains available across future |
|
/use 37814368606 |
|
As a PR reviewer and CODEOWNER, I have reviewed this and have:
Additional detail section:
Signed: [by Claude Code] |
|
@Ankur-singh staged run 37814368606. This run remains available across future |
❌❌❌ REJECTED ❌❌❌@Ankur-singh Blocking: this PR removes the published TP4 c15 point, which was on the frontier, and no head point dominates it. Restore TP4 c15, or add a point that dominates it, then rerun the sweep before merge. ❌ Check 18 (Recipe, image, and sweep consistency): FAIL — Removed point is on the frontier and not dominated:
Passed and not applicable checks✅ Check 0 (CODEOWNER): PASS — ✅ Check 1 (Passing sweep on in-PR commit): PASS — pinned head ✅ Check 2 (Evals): PASS — the same run used image ➖ Check 3 (Upstream recipe): N/A — only a TensorRT-LLM ✅ Check 4 (Reuse command): PASS — ✅ Check 5 (Latest checklist template): PASS — all current template items are present and checked. ✅ Check 6 (Upstream images and engine-first ordering): PASS — the ✅ Check 7 (Deprecated models/scenarios): PASS — ✅ Check 8 (No architecture hacks): PASS — there are no ✅ Check 9 (Spec-decode chat template): PASS — the AgentX replay uses the chat endpoint with ✅ Check 10 (No engine patches): PASS — ✅ Check 11 (Agentic golden AL): PASS — the connector injected ➖ Check 12 (Append-only): N/A — no ✅ Check 14 (Pareto coverage): PASS — one curve (minimaxm3 / agentic / B300 trt / fp4 / run 40384083329 / P90 E2EL vs tok/s/GPU) has 8 frontier points: TP4 c10, TP4 c20, TP4 c25, TP2 c20, c25, c30, c40, c45 (run 40384083329 ✅ Check 15 (PR description): PASS — the image, 15-point list, dropped TP4 c15/c30, draft lengths, NVFP4 KV, host-pool sizes, tool parser, and patch removal all match the pinned code. ➖ Check 16 (NVIDIA overview): N/A — the only affected NVIDIA config uses TensorRT-LLM, so there is no vLLM/SGLang pair. ➖ Check 17 (Metrics publication): N/A — Assessed commit: |
Ankur-singh
added
full-sweep-enabled
|
/use 40011778426 |
|
@hshrivastava-droid staged run 40011778426. This run remains available across future |
…00 TRT-LLM AgentX
…tllm-agentic # Conflicts: # inferencex-e2e/perf-changelog.yaml
Ankur-singh
changed the title
Ankur-singh
removed
the
full-sweep-enabled
Ankur-singh
added
the
full-sweep-enabled
Ankur-singh
changed the title
Ankur-singh
added
full-sweep-enabled
|
/use 40384083329 |
|
@hshrivastava-droid staged run 40384083329. This run remains available across future |
Summary
Update the B300 MiniMax-M3 TensorRT-LLM AgentX config (
minimaxm3-fp4-b300-trtllm-agentic-mtp), its recipe, and the performance changelog.minimaxm3-fp4-b300-trtllm-agentic-mtp(aggregatedtrtllm-serve, EAGLE3 draftInferact/MiniMax-M3-EAGLE3-GQA)nvcr.io#nvidia/tensorrt-llm/release:1.4.0rc0.dev202610070001(was1.3.0rc23.post1)Points (15)
minimaxm3_eagle3_gqagolden curve (thinking on).Recipe changes
block_reuse_config.policy: per_conversation, iteration/perf stats every 10 iterations, tokenization cache,MAXSMS=32.capture_num_tokensladder; TP2 keeps the short ladder.host_cache_size374,500,000,000 bytes per rank at TP2, 374,750,000,000 at TP4 and 309,237,645,312 at TP8 (TOTAL_CPU_DRAM_GB749 / 1499 / 2999,dram-utilization1.0).trtllm-serve --tool_parser minimax_m3.storepatch (setup_script: minimaxm3-trtllm-agentx.sh) is removed.chat_template_kwargs(AIPERF_EXTRA_INPUTS), matching the golden acceptance curve.AI model disclosure
PR checklist and change type
Type of Change
Checklist
inferencex-e2e/perf-changelog.yamland have not edited historical entriesfull-sweep-fail-fast(recommended),full-sweep-enabled, ornon-canary-full-sweep-enabled. Optional modifiersall-evals,evals-only, andagentx-fastrequire a primary label; the last two block reuse while applied.OWNER/MEMBER/COLLABORATOR) has commented/use <run_id>(or the legacy/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 primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.中文
概述
更新 B300 MiniMax-M3 TensorRT-LLM AgentX 配置（
minimaxm3-fp4-b300-trtllm-agentic-mtp）、对应配方及性能变更记录。minimaxm3-fp4-b300-trtllm-agentic-mtp（聚合式trtllm-serve，EAGLE3 草稿模型Inferact/MiniMax-M3-EAGLE3-GQA）nvcr.io#nvidia/tensorrt-llm/release:1.4.0rc0.dev202610070001（原为1.3.0rc23.post1）minimaxm3_eagle3_gqa黄金曲线（开启思考）。MAXSMS=32；主机 KV 池遵循 fix(launch): request explicit Slurm memory for single-node jobs and fit host pools to measured DRAM #3804 的 DRAM 策略；使用--tool_parser minimax_m3解析工具调用；移除 TensorRT-LLM chatstore补丁；通过chat_template_kwargs为回放固定开启思考模式。AI 模型披露
改动类型
配置变更