fix(qwen2_5_omni): put the thinker prefix inside the peft prefix on lora saves - #3672
Merged
yuhezhang-ai merged 2 commits intoSep 18, 2026
Merged
Conversation
…ora saves Fixes NVIDIA-NeMo#3655. Same bug NVIDIA-NeMo#3630 fixes for qwen3_omni_moe: on peft saves the adapter prepended "thinker." to every key, including ones already carrying the "base_model.model." outer prefix, so the exported adapter had keys HF PEFT can't attach. It also broke imports of correctly named external adapters, which passed through from_hf unchanged. The thinker namespace now goes inside the peft prefix on save, from_hf strips it from either position and drops talker/token2wav keys in either position too, and old malformed saves still resume through the leading-strip fallback. The adapter also overrides map_peft_target_module_to_hf so adapter_config.json target_modules carry the thinker namespace, keeping peft's suffix matching off the talker's structurally identical submodules. Signed-off-by: stanley1208 <stanley.mei08@gmail.com>
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
Contributor
|
/ok to test 0ec219f |
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
yuhezhang-ai
approved these changes
Sep 18, 2026
yuhezhang-ai
left a comment
yuhezhang-ai
left a comment
Contributor
There was a problem hiding this comment.
Reviewed the prefix conversion, metadata target mapping, and legacy loading paths; no blocking findings.
Validation: all 17 focused adapter tests passed. An additional tiny-model CPU check using Transformers 5.15.1 and PEFT 0.19.1 exercised native Checkpointer.save_model and real PeftModel.from_pretrained: all eight adapter tensors loaded exactly, bulk and single-tensor exports agreed, and no adapters were injected into talker/token2wav. Reloaded text logits matched exactly; merged logits differed by at most 5.96e-8. Checkpointer.load_model restored both current and legacy malformed checkpoint names exactly. CI is green on this revision.
This is a suitable incremental fix toward #3867. Please carry the real Qwen2.5 Omni consumer regression into that shared suite as a follow-up.
svcnvidia-nemo-ci
removed
the
waiting-on-maintainers
stanley1208
added a commit
to stanley1208/Automodel
that referenced
this pull request
Sep 29, 2026
…family Adding qwen2_5_omni to the reload suite turned up two problems in its map_peft_target_module_to_hf, both already fixed on its qwen3_omni_moe sibling. The signature was missing v4_compatible. addons.py:566 always passes it, so every PEFT save raised TypeError while writing adapter_config.json. The base class has taken that argument since NVIDIA-NeMo#3866; the override landed a week later in NVIDIA-NeMo#3672 without it, and nothing noticed because this family had no coverage. It also added the thinker. namespace unconditionally, while to_hf and convert_single_tensor_to_hf both check _uses_thinker_prefix first. On a thinker-only base that leaves adapter_config.json pointing at modules the receiving model does not have. Same defect as the one found in review on NVIDIA-NeMo#3945, which this family never had applied to it. The tiny fixture needed three rounds of shrinking: the talker and token2wav sub-configs are not coerced from dicts, and the talker carries its own sizes as well as a nested text_config. Left alone the reload target built 6.6B parameters in five minutes; it is now 1.4M in two seconds. Verified it bites: restoring the old hook fails four of the seven cases. Signed-off-by: stanley1208 <stanley.mei08@gmail.com>
This branch was successfully deployed
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
4 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.
Fixes #3655. same bug #3630 fixes for qwen3_omni_moe, just in this adapter.
when you train a lora and save it, every weight gets a name. this adapter added
thinker.to the start of every name, but peft names must start withbase_model.model.. withthinker.in front, hf peft doesn't recognize any of the names, so the saved adapter can't be used. loading adapters made outside automodel failed for the same reason.what changed:
thinker.now goes in the right spot, afterbase_model.model.instead of before it. loading understands both the new and the old naming, so existing saves still workadded cpu tests for all of the above.