fix: bump hf_kl_threshold for customizer_llama_3_2_1b_full_sft_chat#1938
Closed
fix: bump hf_kl_threshold for customizer_llama_3_2_1b_full_sft_chat#1938
Conversation
After the transformers 5.3 -> 5.5 upgrade (#1734) the vanilla HF Llama 3.2 1B forward diverges slightly from the FSDP2 + kernel-patched training-time forward at Phase 4 of the checkpoint robustness test. Phase 3 max KL is still exactly 0 (save/reload is bit-exact), but Phase 4 max KL climbs to ~6.9e-3, overshooting the pre-v5.5 5e-3 threshold. Bumps ci.checkpoint_robustness.hf_kl_threshold from 5e-3 to 2.5e-2 (~1.5x margin over the observed 6.93e-3), matching the pattern already applied to gemma_3_270m_squad (#1932) and qwen2_5_7b_squad (#1937). Evidence on cw-dfw 8xH100 with CI launcher overrides, transformers 5.5.0: [Phase 3] max KL = 0.000000e+00 [Phase 4] max KL = 6.926899e-03 (threshold 2.5e-2) [Phase 6] Step 5 / 6 / 7 diff = 0.000000e+00 (3 steps compared) 1 passed, 24 warnings in 61.42s Signed-off-by: Adil Asif <adasif@nvidia.com> Signed-off-by: adil-a <adil.asif2000@hotmail.com>
adil-a
requested review from
HuiyingLi,
ZhiyuLi-Nvidia,
akoumpa,
hemildesai and
pthombre
as code owners
This was referenced Apr 21, 2026
Collaborator
Author
|
Closing as redundant. Empirically verified on fresh CI sqsh (
The threshold bump to 2.5e-2 is not needed — with #1921 baked into the current sqsh, the test passes cleanly under the original 5e-3 threshold. |
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
customizer_llama_3_2_1b_full_sft_chatcheckpoint robustness overshoots the pre-v5.55e-3KL threshold by reaching~6.93e-3. Phase 3 is bit-exact (max KL = 0), so this is a forward-pass drift in vanilla-HF Llama 3.2, not a save/reload correctness bug.ci.checkpoint_robustness.hf_kl_thresholdfrom5e-3to2.5e-2(~1.5× observed margin, matching the pattern from fix: gemma_3_270m_squad HF KL regression in ckpt robustness #1932 / fix: qwen2_5_7b_squad ckpt robustness thresholds for transformers v5.5 #1937 for other post-v5.5 SFT robustness jobs).Note: the underlying CI job also hit a separate chat-dataset
tool_calls.idbug (already fixed in main by #1921 /fc46ae53). With that fix in place, the remaining failure on cw-dfw reproduced exactly as a Phase 4 threshold overshoot, handled here.Test plan
transformers==5.5.0using CI launcher overrides —[Phase 4] max KL = 6.926899e-03 > threshold 5e-3.[Phase 3] max KL = 0.000000e+00,[Phase 4] max KL = 6.926899e-03 (threshold 2.500000e-02),[Phase 6] Step 5 / 6 / 7 diff = 0(3 steps compared). Final:1 passed, 24 warnings in 61.42s.customizer_llama_3_2_1b_full_sft_chatjob.