Drop glm4.7-flash enforce_eager WAR after bumping vLLM to >=0.21.0
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
PR #2895 set `enforce_eager: true` on `grpo-glm47-flash-4n8g-automodel` to dodge a vLLM 0.20.0 colocated-refit MoE bug (`replace_parameter` reallocates `w13`/`w2` → captured decode CUDA graphs read stale weights → garbage generations). Fixed upstream in vLLM 0.21.0 (vllm-project/vllm#40390).
Once vLLM is bumped to >=0.21.0:
1. Drop `enforce_eager: true` (+ comment) from `examples/configs/recipes/llm/grpo-glm47-flash-4n8g-automodel.yaml:63-68`.
2. In `tests/test_suites/llm/grpo-glm47-flash-4n8g-automodel.sh`: `STEPS_PER_RUN: 15 → 30`, drop the `checkpointing.checkpoint_must_save_by=00:03:45:00` line.
3. Re-run the release test to confirm reward / `gen_kl_error` match the pre-0.20 baseline.
Scope: only the WAR from #2895. Other recipes with `enforce_eager: true` use it for unrelated reasons — don't sweep.
## Links
- PR #2895 (the WAR)
- Upstream fix: vllm-project/vllm#40390
- WAR validation wandb: https://wandb.ai/nvidia/nemorl-pr2895
Contributor guide
Assessment
This issue has not been assessed yet.