Fix: skip trainer config validation in bench mode#535
Merged
pan-x-c merged 1 commit intoMay 9, 2026
Merged
Conversation
Bench mode runs explorer-only; cluster.trainer_gpu_num is left at 0
because the cluster validator (line 244) skips trainer GPU allocation
for bench/explore/serve. The trainer config validator however still
kept 'bench' in its whitelist, so any local-model bench run hit:
trinity/trainer/verl/verl_config.py:430
if train_batch_size % (world_size // sp_size) != 0:
ZeroDivisionError: integer division or modulo by zero
Drop bench from the whitelist; same fast-path semantics as the existing
external_model.enable check immediately below.
pan-x-c
approved these changes
May 9, 2026
chenyushuo
added a commit
to chenyushuo/Trinity-RFT
that referenced
this pull request
May 13, 2026
commit 159743e Author: chenyushuo <297086016@qq.com> Date: Mon May 11 18:12:00 2026 +0800 1. fix unittest 2. add `reset_running_requests=True` to `reset_prefix_cache` commit 1db7973 Merge: e9f8316 8d69663 Author: chenyushuo <297086016@qq.com> Date: Mon May 11 11:06:34 2026 +0800 Merge branch 'main' of github.com:modelscope/Trinity-RFT into dev/fix_qwen3_5 commit 8d69663 Author: Xuchen Pan <32844285+pan-x-c@users.noreply.github.com> Date: Mon May 11 10:42:37 2026 +0800 Support SGLang Inference Engine (agentscope-ai#533) commit 5b1d8a7 Author: weijie <34210233+shiweijiezero@users.noreply.github.com> Date: Sat May 9 20:34:47 2026 +0800 Fix: skip trainer config validation in bench mode (agentscope-ai#535) commit e9f8316 Author: chenyushuo <297086016@qq.com> Date: Sat May 9 11:39:40 2026 +0800 1. Fix Qwen3.5 sequence-parallel training bugs. 2. Fix Qwen3.5 multimodal training bugs. 3. Fix incorrect Qwen3.5 checkpoint parameter naming when saving with Transformers 5.4.0-5.5.4. 4. Add freeze_vision_tower support. 5. Fix compatibility issues with vLLM 0.20. 6. Fix a bug in Experience serialization. 7. Fix the condition for skipping TrainerConfigValidator checks. 8. Improve explorer robustness by safely handling missing rollout coordinator instead of hard-asserting. 9. Propagate checkpoint_job_dir into workflow/taskset runtime arguments. 10. Improve FSDP worker initialization and logging behavior for better stability and observability. 11. Apply typo and minor message fixes. Co-authored-by: Copilot <copilot@github.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Bug
In
benchmode,cluster.trainer_gpu_numis left at its default0because the cluster validator deliberately skips trainer GPU allocation forbench/explore/serve(these modes don't train) — seeconfig_validator.py:244.But the trainer config validator (
config_validator.py:1168) still keepsbenchin its whitelist. So for any local-model bench run (i.e.external_model.enable=false), the call chain reaches:Repro: any yaml with
mode: bench+ a non-external model +engine_num × tensor_parallel_size = total GPU.Fix
Drop
benchfrom the trainer-config-check whitelist. Bench mode runs explorer-only and never touches the trainer, so its trainer parallelism config doesn't need validation. Same fast-path semantics as the existingexternal_model.enable=truecheck immediately below at line 1170.After the fix, all 6 modes behave correctly:
Test
A local-model bench run (Qwen3.6-27B + frozen_lake_obscure eval, 1 node × 8 GPU,
engine_num=4×TP=2) reproduces theZeroDivisionErroron main and runs cleanly with this patch.config.trainer.trainer_configis not accessed by any module outsideconfig_validator.py(verified via repo-wide grep), so skippingsynchronize_config()in bench mode has no downstream effect.