fix: lower DPO Qwen2.5-Math-7B accuracy threshold after torch 2.11 up… - #2676
Merged
Conversation
…grade The torch 2.10→2.11 upgrade in PR #2384 changed numerical behavior in Megatron's unfused attention path, causing the DPO accuracy at step 10 to drop from 0.5625 to 0.5000 for Qwen2.5-Math-7B. Investigation (15+ experiments) confirmed: - pre-vllm container (torch 2.10) + current main → accuracy=0.5625 PASS - nightly container (torch 2.11) + current main → accuracy=0.5000 FAIL - pre-vllm container + torch 2.11 + TE rebuild → accuracy=0.5000 FAIL - Step-by-step values match CI logs bit-for-bit - Issue is specific to Megatron + Qwen2.5-Math-7B + unfused attention - dtensor + same model → accuracy=0.625 (unaffected) - Megatron + Llama-3.2-1B → accuracy=0.62 (unaffected) Lower the threshold from >0.56 to >=0.5 to accommodate the torch 2.11 numerical behavior change. Closes #2626 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Signed-off-by: nliang <nliang@nvidia.com>
Contributor
Author
|
/ok to test 27c7628 |
yuki-97
added
CI:docs
yuki-97
approved these changes
Jun 4, 2026
yuki-97
enabled auto-merge (squash)
June 4, 2026 10:27
pengdurice
pushed a commit
to pengdurice/RL
that referenced
this pull request
Jun 12, 2026
NVIDIA-NeMo#2676) Signed-off-by: nliang <nliang@nvidia.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Lower the DPO Qwen2.5-Math-7B accuracy threshold from
> 0.56to>= 0.5to accommodate the numerical behavior change introduced by the torch 2.10→2.11 upgrade in PR #2384.Root Cause
PR #2384 upgraded torch from 2.10 to 2.11 (along with vLLM 0.17→0.20 and TE recompilation). This changed the numerical behavior in the Megatron unfused attention path, causing the DPO accuracy at step 10 to drop from 0.5625 to 0.5000 for Qwen2.5-Math-7B.
Key evidence from 15+ controlled experiments:
Closes #2626
Comparison between PASS and FAIL