Skip to content

Revert "[WideEP] Integrate DeepEP v2" (#41183)#45008

Closed
vllm-agent wants to merge 1 commit into
vllm-project:mainfrom
vllm-agent:auto-revert/pr-41183
Closed

Revert "[WideEP] Integrate DeepEP v2" (#41183)#45008
vllm-agent wants to merge 1 commit into
vllm-project:mainfrom
vllm-agent:auto-revert/pr-41183

Conversation

@vllm-agent

Copy link
Copy Markdown
Contributor

Revert of PR #41183

This reverts the merge commit e2f993d from #41183.

Reason

Nightly CI build #70731 detected 1 new failure linked to this PR:

  • MoE Refactor Integration Test (B200 - TEMPORARY): Llama-4-Scout-Fp8-ModelOpt-fi-trtllm server crashed with KeyError: 'layers.0.feed_forward.experts.w2_input_scale' in llama4.py weight loading; Nemotron-Nano-30B GSM8K accuracy dropped catastrophically to 0.12 (threshold 0.21).

The PR changed trtllm_fp8_moe.py, routed_experts.py, fused_moe/config.py, and other MoE-related files that are directly exercised by the failing test.

Auto-generated

Auto-generated by CI failure analyzer. Please review before merging.

@tlrmchlsmth

Copy link
Copy Markdown
Member

nemotron gsm8k fixed in #45037 and looks like the llama weight loading was broken by #41184

@tlrmchlsmth

Copy link
Copy Markdown
Member

llama4 bug fixed here: #45047 (#41183 is innocent!)

@tlrmchlsmth tlrmchlsmth closed this Jun 9, 2026
@github-project-automation github-project-automation Bot moved this to Done in NVIDIA Jun 9, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

2 participants