Repository navigation
[Bugfix] Qwen4Exp PLE embedding accepts INC (AutoRound) checkpoints (cherry-pick vllm#59990) - #91
Merged
Merged
Conversation
…ints (vllm-project#59990) Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com> Co-authored-by: Claude <noreply@anthropic.com> (cherry picked from commit 4f52fa3)
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Cherry-pick of upstream vllm-project/vllm#59990 (
4f52fa35efc7, author Vadim Gimpelson, commit and trailers kept as upstream).v0.31.0 predates it, so
rdna_extrasis affected. An AutoRound/INC checkpoint, e.g.Intel/Qwen3.8-Flash-Next-W4A16-AutoRound, fails at model construction withNotImplementedError: Qwen4Exp PLE embedding does not support quantization config INCConfig. The fix asksINCConfigwhether the PLE layer is quantized; that checkpoint marks.*ple.*as 16-bit float, so the PLE table loads as the unquantized BF16 method. If INC says the PLE layer is quantized, the old error is still raised.It applies cleanly next to our branches: the INT4 PLE table, and our compressed-tensors branch, which differs from upstream's. +6 lines in
ngram_embedding.py, +19 lines of test intests/models/qwen4_exp/test_ple.py.tests/models/qwen4_exp/test_ple.pyon gfx1030: 74 passed (V620, venv-7.14.0_0.31.0-final)Not covered: serving the Intel AutoRound checkpoint on RDNA2. This only fixes PLE loading; whether its INC-quantized linear/MoE layers take our RDNA2 kernels is a separate question.