[ci][llm] Upgrade to vLLM 0.29.0

[jeffreywang88]

Description

• Transformers pin: Constrain Transformers to <5.17 because vLLM 0.29 imports a Pixtral API removed in 5.17. fix: Pixtral compat with transformers 5.17 vllm-project/vllm#56484
• EOS patch: Apply a ray-llm image patch so model-config EOS tokens reach xgrammar when auto generation config omits them.
• Error unwrapping: Remove Ray’s obsolete vLLM error-unwrapping workaround; vLLM 0.29 now handles these errors correctly.
• NIXL: Stop Ray from adding the data-parallel rank to the NIXL port; vLLM applies that offset itself.
• Qwen2.5-VL OOM: Set video sampling to 0.5 FPS in Ray’s prepare stage and vLLM to bound multimodal encoder tokens.
• Reward-model plugin: Update the Qwen3 reward-model example for vLLM’s updated AutoWeightsLoader API.

Related issues

Link related issues: "Fixes #1234", "Closes #1234", or "Related to #1234".

Additional information

Optional: Add implementation details, API changes, usage examples, screenshots, etc.

[gemini-code-assist]

Code Review

This pull request upgrades vLLM to version 0.28.0 and updates related dependencies (such as nixl, fastsafetensors, and huggingface-hub) across configuration files, setup scripts, and lockfiles. It also removes deprecated exception-handling workarounds in the vLLM engine integration that are no longer needed with the new vLLM version. To resolve dependency conflicts, a new script is introduced to strip constraints, and an overrides file is added. Feedback on the pull request suggests improving the constraint-stripping script to dynamically parse overridden packages from the overrides file rather than hardcoding them, along with a corresponding update to the documentation comments.

[gemini-code-assist]

Maintenance Hazard: Hardcoded Package Stripping

Currently, the package click is hardcoded to be stripped from the constraints file. If new overrides are added to python/requirements/llm/base_pin_overrides.txt in the future, developers must remember to manually update this script to strip those pins as well.

We can eliminate this maintenance hazard by dynamically parsing base_pin_overrides.txt and stripping any overridden packages automatically using pure bash pattern matching.

Suggested change

	sed \
	-e '/^--extra-index-url /d' \
	-e '/^--find-links /d' \
	-e '/^click==/d' \
	"$SOURCE_FILE" > "$OUTPUT_FILE"
	SED_EXPRS=(-e '/^--extra-index-url /d' -e '/^--find-links /d')
	OVERRIDES_FILE="python/requirements/llm/base_pin_overrides.txt"
	if [[ -f "$OVERRIDES_FILE" ]]; then
	while read -r line; do
	[[ "$line" =~ ^[[:space:]]*# ]] && continue
	[[ "$line" =~ ^[[:space:]]*$ ]] && continue
	trimmed="${line##[[:space:]]}"
	pkg="${trimmed%%[^a-zA-Z0-9_-]*}"
	if [[ -n "$pkg" ]]; then
	SED_EXPRS+=(-e "/^${pkg}==/d")
	fi
	done < "$OVERRIDES_FILE"
	fi
	sed "${SED_EXPRS[@]}" "$SOURCE_FILE" > "$OUTPUT_FILE"

[gemini-code-assist]

Documentation Sync

With the dynamic stripping implemented in strip-llm-constraints.sh, we should update this comment to reflect that the stripping is now automatic.

# Each entry here is automatically stripped from the compiled constraints by
# ci/raydepsets/pre_hooks/strip-llm-constraints.sh. A constraint beats an
# override, and dropping the pin alone just leaves uv on the version already in
# the lock.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Premerge ID enables extra Serve tests

Medium Severity

The extra Python 3.12 serve_python_tests step now also matches the premerge pipeline ID. Sibling extra-version jobs in core, data, and ml only run on the continuous-build label, postmerge, or microcheck. Including premerge makes that label check ineffective and schedules the extra Serve suite on every regular PR.

 

^{Reviewed by Cursor Bugbot for commit bf0f114. Configure here.}

[jeffreywang88]

Still working on CI. Will ensure everything passes before asking for reviews.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Dropped tied embedding skip logic

Medium Severity

The load_weights rewrite filters score. weights but no longer skips lm_head. when tie_word_embeddings is set. This class replaces the parent loader entirely, so tied Qwen3 checkpoints can still present lm_head tensors that AutoWeightsLoader will try to apply.

 

^{Reviewed by Cursor Bugbot for commit 1757167. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

There are 3 total unresolved issues (including 2 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 482632e. Configure here.}