fix(client): update vLLM 0.26 token-serving import

[S1ro1]

vLLM 0.26 moved the token-in/token-out serialization helpers from the serve.disagg package to scale_out.token_in_token_out. Update the lazy multimodal encoder import and its documentation so Qwen-VL feature serialization continues to work with the new release.

Validation:

• uvx --from ruff==0.15.12 ruff format --isolated --check .
• uvx --from ruff==0.15.12 ruff check --isolated renderers/client.py

---

Note

Low Risk
Import-path and comment-only compatibility fix for vLLM 0.26; multimodal encoding logic is unchanged.

Overview
Aligns Qwen-VL multimodal feature encoding with vLLM 0.26, which moved encode_mm_kwargs_item from vllm.entrypoints.serve.disagg.mm_serde to vllm.entrypoints.scale_out.token_in_token_out.mm_serde. The lazy import in _build_qwen_vl_features and the engine-pluggability note in _build_mm_features are updated accordingly.

Docs and comments no longer pin the client to “vLLM 0.20”; they describe the path generically as vLLM’s /inference/v1/generate integration.

A few MalformedGenerateResponseError and OverlongPromptError messages are tightened to single-line f-strings with no behavior change.

^{Reviewed by Cursor Bugbot for commit b37d40e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix vLLM token-serving import path for encode_mm_kwargs_item in _build_qwen_vl_features

Updates the import path for encode_mm_kwargs_item in client.py from vllm.entrypoints.serve.disagg.mm_serde to vllm.entrypoints.scale_out.token_in_token_out.mm_serde to match the module layout in vLLM 0.26. Also removes a version-specific reference in the module docstring and consolidates error message f-strings.

^{Macroscope summarized b37d40e.}

[macroscopeapp]

Approvability

Verdict: Approved b37d40e

Mechanical dependency update to fix vLLM 0.26 import path compatibility. Changes are limited to updating an import path, removing version-specific comments, and consolidating multi-line strings. No logic or behavioral changes.

^{You can customize Macroscope's approvability policy. Learn more.}