Repository navigation
[Bugfix][Model] Fix EXAONE-4.5 image inputs: ViT received un-normalised pixels - #61104
Merged
Merged
Conversation
Signed-off-by: godmook <cmoh4135@naver.com>
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
|
✅ @Godmook, CI is now available for this PR.
|
Contributor
|
/ci run |
|
✅ Triggered Buildkite CI #94285 for commit |
Collaborator
CI selector (shadow): 6 test steps (29 jobs) instead of 60 (92 jobs)Shadow mode: this changes nothing about what CI runs. It shows what the evidence-based selector would pick for this PR, next to today's rules. How it works. Feedback welcome: reply here if it would skip a step this change needs, or runs something unrelated.
Selector would run (6)
Would skip (today's rules run them) (57)
Would add (today's rules do not run them) (3)
AMD mirrors: would skip (56)
AMD mirrors: would add (0)none 2 changed files · base |
Contributor
Author
|
I think CI is green. @cjackal Could you kindly merge it? Thanks! |
AndreasKaratzas
pushed a commit
to AndreasKaratzas/vllm
that referenced
this pull request
Oct 16, 2026
…ed pixels (vllm-project#61104) Signed-off-by: godmook <cmoh4135@naver.com> Co-authored-by: cjackal <44624812+cjackal@users.noreply.github.com> Signed-off-by: Andreas Karatzas <AndreasKaratzas@users.noreply.github.com>
michael-gpu
pushed a commit
to practical-inference/vllm-ampere-extended
that referenced
this pull request
Oct 20, 2026
…ed pixels (vllm-project#61104) Signed-off-by: godmook <cmoh4135@naver.com> Co-authored-by: cjackal <44624812+cjackal@users.noreply.github.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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
EXAONE-4.5 (
Exaone4_5_ForConditionalGeneration) produces garbage for any request that contains an image on every release since #50411 (v0.29.0, v0.30.0, v0.31.0, main). Text-only requests are unaffected.Root cause: the class inherits
supports_mm_device_do_normalize = TruefromQwen2_5_VLForConditionalGeneration, somm_device_do_normalizedefaults to on and the HF processor is tolddo_rescale=False, do_normalize=False. ButEXAONE4_5_VisionTransformeris constructed withoutinput_norm, so the inheritedQwen2_5_VisionTransformerusesIdentityInputNormand the ViT receives raw 0–255 pixels.Two smaller issues in the same constructor are fixed alongside, since they are the same missing wiring relative to the parent class:
self.model_configandself.video_pruning_rateare never set, but the inheritedembed_multimodal/ encoder-CUDA-graph paths read them (AttributeErrorwith--compilation-config '{"cudagraph_mm_encoder": true}').Exaone4_5MultiTokenPredictor) builds itsExaone4DecoderLayerwithoutcache_config, so the drafter's attention ignores--kv-cache-dtype.Claims
build_mm_input_norm(model_config)), exactly asQwen2_5_VLForConditionalGenerationdoes.--no-mm-device-do-normalize.Validation
LGAI-EXAONE/EXAONE-4.5-33B-FP8, TP=2, 2× A100-SXM4-40GB, greedy, thinking off, image = the model card'sassets/exaone45_input1.jpg(a bowl of kimchi):The\n</think>\n</think>one\n</think>…/or/or/or/or/or…A person in a light blue shirt is serving freshly made kimchi from a large black bowl into a smaller black dish, …Second prompt ("What objects are in this image? List them.") shows the same pattern:
/or/or/or…before, a correct itemised list (kimchi, dark ceramic bowl, …) after.pre-commitpasses (ruff, mypy, typos).tests/models/multimodal/processing/test_common.py) already covers it and still passes.Details
exaone4_5.py:EXAONE4_5_VisionTransformer.__init__gainsinput_normand forwards it toQwen2_5_VisionTransformer;Exaone4_5_ForConditionalGeneration.__init__passesinput_norm=build_mm_input_norm(self.model_config)and setsself.model_config/self.video_pruning_ratelike the parent.exaone4_5_mtp.py: passcache_config=vllm_config.cache_configto the MTP decoder layer.Not in this PR (follow-ups):
Exaone4ModellacksEagleModelMixinalthough the VLM advertisesSupportsEagle3(asserts at load);EXAONE4_5_VisionTransformerbuilds and discards the full Qwen2.5 MHA block stack insuper().__init__before replacing it with GQA blocks.Duplicate-work check: no open PR or issue mentions EXAONE-4.5 image output,
mm_device_do_normalizefor EXAONE, orexaone4_5(searchedexaone4_5,EXAONE-4.5,mm_device_do_normalize input_norm; the closed #45071 / #42281 are different bugs).Pull Request Checklist
I used vLLM's
/pr-checklistskill. (Mandatory for agents, optional for humans).AI assistance was used during the creation of this PR.
Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.
Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.
Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.
Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.
AI assistance was used while preparing this PR; I reviewed every changed line and ran all of the validation above myself.