Repository navigation
[Klaud Cold] Update MiniMax-M3 FP4 B300 TensorRT-LLM AgentX MTP image to 1.4.0rc1 / [Klaud Cold] 将 MiniMax-M3 FP4 B300 TensorRT-LLM AgentX MTP 镜像更新至 1.4.0rc1 - #4607
Conversation
Pin nvcr.io/nvidia/tensorrt-llm/release:1.4.0rc1 by digest, rename kv_cache_config.block_reuse_policy to block_reuse_config.policy as the 1.4 KvCacheConfig requires, and drop the openai_protocol.py store=false patch hook because engine patches are not allowed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks for the contribution!
中文感谢你的贡献！
|
|
Initial attempt · Passed · Run 40656033792 / attempt 1 · 2026-10-28 10:10 UTC
Note: AgentX c20 TP2 EP1 66f346: 224 request errors.
Note: minimax_m3_full/em_strict · c20 TP2, minimax_m3_full/em_strict · c5 TP4, minimax_m3_full/em_strict · c1 TP8: Δ N/A: no matched eval baseline. Next: All 3 smoke benchmarks and 4 evals passed (gsm8k c30 95.1% vs 95.6%); minimax_m3_full is 0.15 diagnostic (threshold 0) because the recipe sets no tool_parser on old or new image, so no tool calls are parsed; append the perf-changelog entry and dispatch the final sweep. 中文初次尝试 · 已通过 · Run 40656033792 / attempt 1 · 2026-10-28 10:10 UTC |
|
parked · Session 1 is waiting on 40656033792. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 1 个会话正在等待 40656033792。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。 |
github-actions
Bot
added
the
klaud-parked
|
resumed · The session ended before finishing; session 2 continues it in a new run of its planner. 中文resumed · 会话未完成即已结束；第 2 个会话将在其规划运行的新运行中继续。 |
github-actions
Bot
added
klaud-in-progress
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Klaud-Cold
added
the
full-sweep-fail-fast
|
Final full sweep · Passed · Run 40670198443 / attempt 1 · 2026-10-28 13:17 UTC
Note: AgentX c1 TP8 EP1 ef14b4: 12 request errors.
Note: minimax_m3_full/em_strict · c1, minimax_m3_full/em_strict · c5, minimax_m3_full/em_strict · c5, minimax_m3_full/em_strict · c10, minimax_m3_full/em_strict · c15, minimax_m3_full/em_strict · c20, minimax_m3_full/em_strict · c20, minimax_m3_full/em_strict · c25, minimax_m3_full/em_strict · c25, minimax_m3_full/em_strict · c30, minimax_m3_full/em_strict · c30: Δ N/A: no matched eval baseline. Next: Mark ready for maintainer review. 中文最终完整 sweep · 已通过 · Run 40670198443 / attempt 1 · 2026-10-28 13:17 UTC |
|
parked · Session 2 is waiting on 40670198443. Recovery reruns preempted runs and starts the next session once they end. 中文parked · 第 2 个会话正在等待 40670198443。恢复流程会重跑被抢占的运行，并在运行结束后启动下一个会话。 |
github-actions
Bot
added
klaud-parked
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40670198443 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40670198443 |
|
validated · Repairs: unknown · Runs: 40656033792, 40670138588, 40670198443 中文validated · 修复次数：未知 · 运行：40656033792, 40670138588, 40670198443 |
github-actions
Bot
added
klaud-finished
|
Claude finished @Klaud-Cold's task in 44s —— View job ReviewLGTM: no blocking issues found.
|
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Apart from the inline finding, I checked two other parts of the diff and found no problems with them. The nvcr.io/...@ sha256 image form matches the existing 1.4.0rc0/rc1 entries in nvidia-master.yaml. The nested block_reuse_config: {policy: ...} key also appears in the other TRT-LLM 1.4 recipes (qwen3.5, musespark, dsv4 on gb300).
Extended reasoning...
The change bumps the B300 MiniMax-M3 TRT-LLM AgentX image to 1.4.0rc1, removes the recipe's setup_script, switches to the nested KV block-reuse config, and appends a perf-changelog entry. It touches no security-sensitive code. The image reference format and the nested block_reuse_config key were checked against existing 1.4 entries and recipes, and both are consistent.
| @@ -14,7 +14,6 @@ base: | |||
| engine: | |||
| type: trtllm | |||
| served_model_name: nvidia/MiniMax-M3-NVFP4 | |||
There was a problem hiding this comment.
🔴 If TRT-LLM 1.4.0rc1 still lacks a store chat field, BFCL eval runs on this B300 key will get request-validation errors that the base image's patch prevented. Removing setup_script: minimaxm3-trtllm-agentx.sh also removes the only step that adds store: Optional[Literal[False]] to ChatCompletionRequest. Nothing in this PR shows the new image accepts store=false natively. Fix: before dropping the hook, confirm 1.4.0rc1's serve/openai_protocol.py accepts store. If it does not, keep the hook, as the b200-fp4-mtp recipe still does.
Why this was flagged
The removed hook is minimaxm3-trtllm-agentx.sh (multi_node/srt-slurm-recipes/configs/minimaxm3-trtllm-agentx.sh:8). It runs runners/patch_trtllm_chat_store.py, which inserts store: Optional[Literal[False]] = False into TensorRT-LLM's ChatCompletionRequest. The trigger is an explicit BFCL suite (bfcl_smoke or bfcl_vllm_minimax_m3 in infx/bench/eval/vendor.py:348-352) run against minimaxm3-fp4-b300-trtllm-agentic-mtp. BFCL's OpenAI client sends store=false. On the base image the hook adds the field, so those requests are accepted. With this diff there is no hook. If 1.4.0rc1's protocol model still rejects unknown fields and has no store field, every BFCL chat request is rejected and the eval fails or scores 0. The perf-changelog entry describes only the image bump and does not mention the dropped hook.
Verification: The diff deletes setup_script: minimaxm3-trtllm-agentx.sh from b300-fp4-mtp/agentic.yaml, at old line 17. runners/patch_trtllm_chat_store.py:13 inserts store: Optional[Literal[False]] = False into ChatCompletionRequest in tensorrt_llm/serve/openai_protocol.py. If the new image still has no store field, every BFCL tool-use request to this key goes back to failing with HTTP 400.
Goal: Update TensorRT-LLM image from nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc23.post1 to nvcr.io/nvidia/tensorrt-llm/release:1.4.0rc1.
Baseline: 2026-09-01 ·
nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc23.post1Mean latency · Sources: API 1, API 2, API 3, API 4
Stop: Close this PR at any time to stop this candidate now; Klaud cancels its runs and deletes its branch.
中文
**目标：**将 TensorRT-LLM 镜像从 nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc23.post1 更新到 nvcr.io/nvidia/tensorrt-llm/release:1.4.0rc1。
**基线：**2026-09-01 ·
nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc23.post1平均延迟 · 来源： API 1, API 2, API 3, API 4；数值及异常说明见上表。
**停止：**随时关闭此 PR 即可立即停止该候选：Klaud 会取消其运行并删除其分支。