[Hotragn]

Purpose

Follow-up to #56576 (issue #56527), which made MultiModalHasher tolerate a
malformed EXIF header. Molmo 2 keeps its own copy of the EXIF normalization and
that copy is not tolerant, so for that model the merged fix moved the
failure one step later instead of removing it.

vllm/multimodal/image.py suppresses the read:

def normalize_image(image: Image.Image) -> Image.Image:
    """Normalize EXIF orientation so the pixel data matches visual display."""
    with contextlib.suppress(Exception):
        image = ImageOps.exif_transpose(image)
    return image

vllm/model_executor/models/molmo2.py does not:

def exif_transpose(images: ImageInput | None) -> ImageInput | None:
    ...
    elif images is not None and isinstance(images, Image):
        images = ImageOps.exif_transpose(images)   # bare
    return images

The fix reuses normalize_image rather than adding a second
contextlib.suppress, so there is only one place that decides how a bad EXIF
block is handled. vllm/multimodal/parse.py already imports it the same way.

Why the second read is the one that raises

This is the non-obvious part, and it is why the bug survived #56576. Measured on
Pillow 12.3.0 with the payload from #56576's test:

sequence	result
Image.open(...) then ImageOps.exif_transpose(im)	no raise
ImageMediaIO.load_bytes(...) then ImageOps.exif_transpose(im)	SyntaxError: not a TIFF file (header b'XXXX\x00\x00\x00\x08' not valid)
the same, second call on the same object	no raise

ImageMediaIO.load_bytes (vllm/multimodal/media/image.py:90-92) runs
normalize_image(image) -> image.load() -> _convert_image_mode(image). The
suppressed transpose plus load() is what primes Pillow's lazy EXIF state, and
.convert() carries that state forward — so the next read raises, and only
the next one (Pillow caches the failure afterwards). That last row is why the
regression test loads a fresh image for every case; reusing one would make the
test pass on main by coincidence.

Reachability

POST /v1/chat/completions against a Molmo 2 model with an image_url whose
JPEG carries a malformed APP1/Exif segment:

1. ImageMediaIO.load_bytes primes the EXIF state (above).
2. Hashing runs next and is now tolerant, thanks to [Bugfix] tolerate malformed EXIF metadata during hashing (#56527) #56576.
3. Molmo2MultiModalProcessor._get_prompt_updates ->
   get_image_replacement_molmo2 (molmo2.py:2218) -> the bare
   exif_transpose -> SyntaxError.

SyntaxError is neither VLLMClientError nor GracefulHTTPError, so
AsyncLLM.generate's generic handler wraps it as
raise EngineGenerateError() from e (vllm/v1/engine/async_llm.py:764) — a 500
with an empty message, matching what #56527 reported for the hasher.

Is the class fully fixed?

Yes. grep -rn "exif_transpose\|getexif" --include=*.py vllm/ on current main
returns four sites:

site	state
vllm/multimodal/image.py:24	suppressed
vllm/multimodal/hasher.py:74	tolerant since #56576
vllm/model_executor/models/molmo2.py:1302	fixed here
vllm/model_executor/models/molmo2.py:1299	recursive call into the same function

Molmo 2 was the only remaining bare read, so no follow-up is needed.

Test

No test ships with this PR — removed at the reviewer's request
(review of 2026-09-17: "Thanks but let's remove the tests"). The diff is now
the two-line source change only.

For the record, the regression test that had been in this PR was verified on
fork CI run
35144842256, with
BEFORE run against main (30b847c1fa):

BEFORE — 1 failed, and the failure is the defect itself rather than an
incidental assertion:

tests/models/multimodal/processing/test_molmo2.py:135: in test_exif_transpose_tolerates_malformed_exif
    assert exif_transpose(_served_image()) is not None
vllm/model_executor/models/molmo2.py:1302: in exif_transpose
    images = ImageOps.exif_transpose(images)
PIL/ImageOps.py:703: in exif_transpose
    image_exif = image.getexif()
...
E  SyntaxError: not a TIFF file (header b'XXXX' not valid)

AFTER — the whole file: 6 passed. Whole-suite
pytest tests/models/multimodal/processing on both sides: the same two
pre-existing collection errors (test_llava_next.py, test_llava_onevision.py),
identical on main and on the branch.

Local lint on the remaining change, all clean: ruff check / ruff format
(0.14.0), typos (1.43.5), check_spdx_header, check_forbidden_imports,
check_torch_cuda, check_init_lazy_imports.

pytest cannot run on my local machine (Windows-on-ARM), hence the fork runner;
the Pillow measurements in the table above were made locally with Pillow alone.

Notes

AI assistance was used to research and draft this change. I have reviewed every
changed line and verified the change as described above.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[DarkLight1337]

Thanks but let's remove the tests

[DarkLight1337]

/ci retry

[github-actions]

✅ No earlier CI build exists for this PR. Use /ci run first.

[DarkLight1337]

Sorry you still haven't removed the tests

[Hotragn]

Apologies for the delay — I missed that the request was in the approval body. Done in 5a8c4e8c45.

The test additions are reverted and tests/models/multimodal/processing/test_molmo2.py is byte-identical to main again (note it is a pre-existing file, so this reverts my 47 added lines rather than deleting the file). The PR is now one file, +2/-2: molmo2.py's bare ImageOps.exif_transpose becomes normalize_image.

I kept the fork-CI before/after in the PR body as a record of how the fix was verified, clearly marked as not shipping.

[DarkLight1337]

/ci run --allow-stale

[github-actions]

✅ Triggered Buildkite CI #90023 for commit 5a8c4e8c4579.

⚠️ This PR is 25 commits behind upstream main. Running CI at your own risk because --allow-stale was requested; outdated CI configuration may cause failures. Before merging, merge or rebase onto the latest main, then rerun /ci run on the latest PR commit.