[hmellor]

Overview

Bump Transformers to v5.21.1. Follows #61983.

Claims

• requirements/common.txt allows transformers >= 5.16.1, < 5.22.0.
• CI pins 5.21.1, which adds Fix rope_scaling setter for configs that assign it during init huggingface/transformers#50080 on top of 5.21.0. That fixes the rope_scaling setter regression from Deprecate rope_scaling assignment and normalize legacy RoPE keys huggingface/transformers#49515, which broke remote-code configs (Phi-3/3.5, Phi-4-multimodal, InternVL2, InternLM2.5, Step-3.7), Step5Config and the EAGLE3 speculators config (rope_theta=None).
• MiniCPM-V 4.6 passes downsample_mode to find_best_resize/get_refine_size on 5.21, which made it required and changed alignment in "4x" mode.
• Keye-VL-1.5's remote processor breaks on 5.21 (convert_to_rgb rejects tensors); Keye was removed in [Model] Remove Keye-VL and Keye-VL-1.5 #64287.
• The 5.21 TP plan changes (🚨 TP: give embeddings their own style, shard input embeddings by default, and shard widths that don't divide by the TP size huggingface/transformers#49570) do not affect the Transformers modeling backend, which detects vocab embeddings itself.

Validation

• First /ci run nightly on 5.21.0 (build 97620): 37 of 41 failures came from the bump. The other 4 also fail on main.
• test_common.py -k MiniCPM-V-4_6: 3 failed on Transformers main without the fix, 3 passed with it, and passes on 5.20.
• Configs vLLM resolves for 384 checkpoints (test registry plus speculator drafts), 5.20 vs Transformers main vs main + [Bugfix][Spec Decode] Scope MTP completeness checks for Kimi K3 #50080: only Step-5-Preview and nm-testing/Llama3_2_1B_speculator.eagle3 change on main, and both match 5.20 with the fix.
• The package mirror lacks 5.21.1, so only the pin lines were edited (pip-compile skipped locally); CI pre-commit verifies the lockfiles.
• Second /ci run nightly on v5.21-release (build 97995): 7 failures, all also failing on main (HiSparse/MLA attention, Humming Activation, BMG V1).
• /ci run on 5.21.1 (identical code to v5.21-release): pending.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (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.

🤖 Generated with Claude Code

[hmellor]

/ci run nightly

[github-actions]

✅ Triggered Buildkite CI #97620 for commit 2ad45f5da7f3.

[hmellor]

/ci run nightly

[github-actions]

✅ Triggered Buildkite CI #97989 for commit ee1b3eaafcef.

[hmellor]

/ci run nightly

[github-actions]

✅ Triggered Buildkite CI #97995 for commit 1f98e9df25ca.

[hmellor]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98263 for commit 4ccd63e0fff8.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[hmellor]

cc @AndreasKaratzas the penultimate run used nightly CI and the only failures were present on ci.vllm.ai/alerts. I'm only using regular CI for the final merge to save time/cost

[hmellor]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98274 for commit 8905a060c531.