Repository navigation
[Frontend] Replace return_mm_kwargs with an mm_payload enum on render requests - #61427
Merged
Merged
Conversation
return_mm_kwargs from vllm-project#60728 is a bool, so it has no room for the "source bytes" option discussed in vllm-project#46722. Switch to mm_payload: "tensors" | "none" (default "tensors") on the chat, completions, Messages and Responses request models. return_mm_kwargs still works as a deprecated alias. false means mm_payload="none" and mm_payload overrides if both are set. Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
hickeyma
marked this pull request as ready for review
October 13, 2026 14:21
hickeyma
requested review from
DarkLight1337,
aarnphm,
aoshen02,
chaunceyjiang,
mgoin,
noooop and
sfeng33
as code owners
October 13, 2026 14:21
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Contributor
|
Documentation preview: https://vllm--61427.org.readthedocs.build/en/61427/ |
mergify
Bot
added
documentation
1 task done
Member
|
/ci run |
DarkLight1337
enabled auto-merge (squash)
October 15, 2026 11:14
github-actions
Bot
added
the
ready
|
✅ Triggered Buildkite CI #95118 for commit |
4 tasks done
tessapham
pushed a commit
to tessapham/llm-d-router
that referenced
this pull request
Oct 15, 2026
vLLM replaced the return_mm_kwargs render option with mm_payload in vllm-project/vllm#61427 and kept return_mm_kwargs as an alias that is ignored when mm_payload is set, so a client-sent mm_payload overrides omitMMKwargs. Sending both keeps v0.32, which knows only return_mm_kwargs, and lets the router's values win on later versions. Signed-off-by: Maroon Ayoub <mayoub@redhat.com> Signed-off-by: Maroon Ayoub <Maroonay@gmail.com>
khluu
pushed a commit
that referenced
this pull request
Oct 17, 2026
… requests (#61427) Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
michael-gpu
pushed a commit
to practical-inference/vllm-ampere-extended
that referenced
this pull request
Oct 20, 2026
… requests (vllm-project#61427) Signed-off-by: Martin Hickey <martin.hickey@ie.ibm.com>
This was referenced Oct 29, 2026
Open
1 task done
zbrad
added a commit
to zbrad/vllm
that referenced
this pull request
Nov 2, 2026
Conflicts (17 files). Took upstream's side where the fork only carried upstream PRs that v0.33.0 contains at the same shas (vllm-project#61718, vllm-project#61858, vllm-project#61427, vllm-project#61409, vllm-project#58297, vllm-project#61556, vllm-project#61167, vllm-project#59815, vllm-project#59604). Kept the fork's patches where they are fork-only: - cmake/external_projects/vllm_flash_attn.cmake: GB10 prebuilt-package block kept; upstream's FetchContent GIT_TAG bump (b273a0262) applied. - vllm/env_override.py: tuned-flags block kept; upstream removed the Triton force-first-config block, so it is dropped. - vllm/model_executor/warmup/kernel_warmup.py: torchvision-missing guard kept; upstream's step5_dsa_warmup import added. - requirements/cuda.txt: upstream's pins (humming 0.1.19). Upstream's new model-microbenchmark skill moved under rules_advisory/. Assisted-by: Claude
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
Replaces the
return_mm_kwargsbool from #60728 with anmm_payloadenum on render requests. This enables the "source bytes" option from #46722 has somewhere to go later.Claims
mm_payload: "tensors" | "none"field (default"tensors") on the chat, completions, Messages and Responses request modelsreturn_mm_kwargsstill works as a deprecated alias.falsemeansmm_payload="none"return_mm_kwargs: falseValidation
pytest tests/entrypoints/scale_out/token_in_token_out/test_mm_serde.py -v: passespre-commit runpasses, including mypyNew test
test_return_mm_kwargs_is_alias_for_mm_payloadcovers the alias on all four request models. No model output change, so no evals.Details
return_mm_kwargsis inv0.32.0rc0but not in a final release. This means it is the cheap time to change its shape. @vMaroon suggested one three valued field (tensors, source bytes, none) in #46722. This PR lands the first two values and leaves"source"to #53169.If both fields are set,
mm_payloadoverrides.extract_mm_featuresnow takesmm_payloadinstead ofinclude_mm_kwargs.Not a duplicate, no open PR changes this flag. #58660 speeds up render serialization, #61418 only uses the flag and #53169 adds raw image bytes which would plug into this enum.
@vMaroon @DarkLight1337
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.