Repository navigation
Conversation
Qwen3.5 ships its MTP draft layer inside the target checkpoint. All four target loaders in `models/qwen3_5.py` skip `mtp` weights on load, and the draft worker picks them up through the `Qwen3_5ForCausalLMMTP` rewrite in `ModelConfig._config_draft_model`. `_handle_eagle_family` does not know this, so none of the four appear in the list that defaults `speculative_draft_model_path` to `model_path`. Omitting `--speculative-draft-model-path` therefore leaves it `None`, and it is passed straight to `ModelConfig.from_server_args` as the draft worker's `model_path` with no fallback. Unlike DFLASH and UNO, the EAGLE path has no "requires setting" validation either, so the result is an obscure load failure rather than a clear message. Add the four architectures, next to `Qwen4ExpForConditionalGeneration`, which was added for exactly the same reason. This can only turn a failure into a working default: callers that already pass the flag keep using what they passed. Also make the accompanying warning name the architecture it fired for. It read "DeepSeek MTP does not require setting speculative_draft_model_path", which was already inaccurate for the GLM, Bailing, HunYuan and Qwen4-Exp entries on the list. Co-authored-by: Cursor <cursoragent@cursor.com>
omirosh
force-pushed
the
fix/qwen3_5-bundled-mtp-draft-path
branch
from
September 11, 2026 09:09
5a8d8a5 to
0ca32f3
Compare
omirosh
marked this pull request as ready for review
September 15, 2026 14:16
4 of 5 tasks
Contributor
Author
|
@hnyls2002 @Qiaolin-Yu could you review this? It is argument resolution only: four Qwen3.5 architectures now default |
5 tasks done
This branch has not been 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
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.
Motivation
Qwen3.5 already bundles its MTP draft in the target checkpoint, so --speculative-draft-model-path is redundant, but omitting it currently fails the load. This PR adds the four Qwen3.5 architectures to the existing list of models that default the draft path to --model-path.
Modifications
Add the four Qwen3.5 architectures next to
Qwen4ExpForConditionalGeneration, which is listed for the same reason, and make the warning name the architecture it fired for. It read "DeepSeek MTP does not require setting speculative_draft_model_path", already wrong for the GLM, Bailing, HunYuan and Qwen4-Exp entries.This cannot regress an existing setup. The new entries only take effect when
speculative_draft_model_path is None, which previously could not work for these architectures. Callers that already pass the flag keep the path they passed.InternS2PreviewForConditionalGeneration,InternS2MobiusForConditionalGenerationandQwen3MoeForCausalLMhave MTP remaps too and are absent for the same reason, one line each if their owners confirm the published checkpoints bundle the draft. Left out because I cannot test them.Accuracy Tests
None needed: this resolves a launch argument to the same path users pass by hand today. Checked on
amd/Qwen3.8-2.4T-A95B-Quark-MXFP4(Qwen3_5MoeForCausalLM, NEXTN 3/1/4, TP8).Speed Tests and Profiling
None; argument resolution only, no runtime path changes.
Related: #39064, a separate fix for mixed Quark MXFP4 Qwen3.5 MTP checkpoints.
Checklist
_handle_eagle_familyneeds a resolvedModelConfig, so exercising it wants a real checkpoint config rather than a CPU unit test, and the change is four string literals. Happy to add one if you can point me at the preferred fixture.Review and Merge Process
/tag-and-rerun-ci,/tag-run-ci-label,/rerun-failed-ciCI States
Latest PR Test (Base): ❌ Run #37437644012
Latest PR Test (Extra): ❌ Run #37437643739
Latest PR Test (AMD ROCm 10): ❌ Run #37437644040