Repository navigation
Conversation
When return_mm_kwargs is false, use exact processor metadata so families whose prompt updates already ignore processed kwargs skip HF tensor construction, and fail closed to full apply otherwise. Co-authored-by: Cursor <cursoragent@cursor.com> Signed-off-by: Guy Girmonsky <guygir@gmail.com>
Qwen prompt updates read image_grid_thw from processed tensors, so the generic /render path cannot skip HF construction until ProcessingInfo owns the exact image token counts. Co-authored-by: Cursor <cursoragent@cursor.com> Signed-off-by: Guy Girmonsky <guygir@gmail.com>
6 tasks done
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Depends on #12 . Opt Qwen2-VL, Qwen2.5-VL, and Qwen3-VL image processors
into exact routing-metadata so
/rendercan skip pixel-tensor construction.Production code under
vllm/once merged, assuming rebased on #12 : +162 / −1 (approx).Claims
vllm launch render+ Qwen2.5-VL-7B-AWQ, 4×1280×720 + ~1000 texttokens (6111 prompt tokens), unique media,
return_mm_kwargs=false:This is
/render/ token-producer latency, not 300-token generate E2E.Validation
Those tests compare full
apply()vsapply_routing_metadata()and requireidentical token IDs, hashes, and placeholders, plus empty routing kwargs.
CI:
.buildkite/test_areas/models_multimodal.yamlprocessor shard.Reproduce the cluster
/rendernumbers withmm-routing-audit/cluster_render_saturation_curve.pyagainst stockorigin/mainand this branch.Details
Qwen prompt updates currently read
image_grid_thwfrom processed tensors,so PR1's generic path cannot skip HF construction. The adapters compute the
same smart-resize token counts from image dimensions only.
Exact processor and
ProcessingInfoclass opt-in is required so subclassessuch as Jina-VL do not inherit an incorrect fast path.
AI assistance was used. A human submitter reviewed the diff.
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.