NVIDIA-NeMo / NVIDIA-NeMo/RL

Remove Transformer Engine Triton 3.5 source patch

Open
#3,307 1 comment 0 reactions 1 assignee Claimed by @terrykong View on GitHub
bug
Dominant language
Python
Stars
2k
Forks
561
Avg merge
4d 5h
Merged PRs (30d)
145

Description

**Describe the bug**

NeMo RL still carries `apply_transformer_engine_patch()` in `nemo_rl/models/policy/workers/patches.py`. The patch rewrites `transformer_engine/pytorch/triton/permutation.py` to work around NVIDIA/TransformerEngine#2286.

That TE issue appears fixed upstream, and TE release notes say v2.10 fixed the Triton 3.5 permute-kernel crash. Current NeMo RL upstream also pins newer TE versions, e.g. `release_v2.15` so the workaround may no longer be needed.

Could we remove this local source-rewriting patch, or at least gate it to only run for vulnerable TE versions?

**Steps/Code to reproduce bug**
N/A

**Expected behavior**

Remove `apply_transformer_engine_patch()` and its call sites, or add a TE version guard.

**Additional context**

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.