fix(vllm): expand Kimi K3 media pads at the worker

[furionw]

Summary

• convert each native frontend media pad into the checkpoint Kimi image-placeholder token sequence at the vLLM adapter boundary
• count image or unified vision-chunk payloads for scalar and list media
• reject pad/media mismatches and invalid K3 metadata instead of silently misaligning inputs
• cache only successful K3 resolution and definite non-K3 results
• retain the sparse splice path for very long prompts

Stack

Base: PR1, qiwa/k3-native-frontend

Validation

• focused pre-commit hooks
• pre-commit run --all-files
• Python bytecode compilation
• unit coverage for single and multiple images, vision_chunk, mismatch, rollout compatibility, failures, caching, and long prompts

Runtime qualification follows this branch using vllm/vllm-openai:kimi-k3.

Summary by CodeRabbit

• New Features

  • Added support for Kimi K3 multimodal prompts by converting structural media placeholders into the format expected by the model.
  • Supports prompts containing multiple images and vision chunking.

• Bug Fixes

  • Added validation for mismatched image placeholders and provided images.
  • Preserves prompts that are already correctly formatted.
  • Safely avoids rewriting requests when required media or model metadata is unavailable.

[devin-ai-integration]

Devin Review found 1 potential issue.

[coderabbitai]

Walkthrough

Changes

Kimi K3 prompt expansion

Layer / File(s)	Summary
Pad resolution and token splicing components/src/dynamo/vllm/multimodal_utils/request_processor.py	Adds cached Kimi K3 metadata resolution, structural-pad validation, and replacement of media pads with native image placeholder token ids.
Prompt builder integration components/src/dynamo/vllm/multimodal_utils/request_processor.py	Routes token ids through Kimi K3 expansion during prompt construction.
Expansion behavior validation components/src/dynamo/vllm/tests/multimodal_utils/test_vllm_request_processor.py	Adds coverage for expansion, mismatches, metadata failures, caching, no-op cases, unified vision chunks, and long prompts.

Estimated code review effort: 4 (Complex) | ~45 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description covers the change and validation, but it misses required template sections like Overview, Details, reviewer start, and Related Issues.	Reformat the PR body to the required template and add the missing sections, especially reviewer guidance and either a linked issue or the no-issue checkbox.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise and accurately summarizes the main change: expanding Kimi K3 media pads in the vLLM worker.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[devin-ai-integration]

Devin Review found 1 new potential issue.