[hmellor]

Overview

Bump Transformers to v5.22.0. Follows #64049.

Claims

• requirements/common.txt allows transformers >= 5.16.1, < 5.23.0, and CI pins 5.22.0.
• The only breaking change in 5.22 (🚨 Separate EP logic into its own file huggingface/transformers#49679, EP styles moved out of tensor_parallel.py) does not touch vLLM: nothing imports transformers.distributed or the moved classes.
• 5.22 adds a native RadioModel (q/k/v_proj, o_proj), which vLLM's RADIO loader already maps, so only test_radio.py needed updating. Its weight-loading test also compared uninitialized (possibly NaN) parameters, which is fixed here too.

Validation

• No tests are gated on 5.22 (min_transformers_version/skip markers).
• test_radio.py on B200 with 5.22: 10 passed.
• Qwen2.5-VL PPL on B200: 22.2441 (+0.30%) on both 5.21.1 and 5.22, so the PPL job's failure is not from this bump.

pytest tests/models/multimodal/pooling/test_radio.py
pytest "tests/models/multimodal/generation_ppl_test/test_qwen.py::test_ppl" -k model_info1

Expected failures from /ci run nightly (#98344); none are caused by this PR (main status from ci.vllm.ai/alerts):

Job	Status
(MI355 DPX) / (H200 MIG 18GB) Multimodal Models (Extended Pooling)	Fixed in this PR (RADIO tests, 7e9455d)
(MI300) DeepSeek-V4.1 gfx942 Mono Decode Kernels	Fixed on main by #64532, merged in: failing, passing
(MI355) Kimi K3	Failing on main (stale _StubSpeculativeConfig)
(4xB200) DeepSeek-V4.1 Specialized Decode	Failing on main (_Mxfp8Linear.params_dtype)
(B200) [Nightly] Fusion E2E Config Sweep	Failing on main (act_quant_fusion count)
(B200) [Nightly] LM Eval Humming Activation FP8/INT8	Failing on main (two-CTA UMMA assert)
(H200 MIG 35GB) [Nightly] Multimodal Models (PPL)	Failing on main (same -1.014% for Qwen2.5-VL)
(BMG) V1 test	Failing on main (HiSparse, test_ucx_backend_params); test_register_packed_dsv4_mla_cache_as_single_region broken by #57662 after that nightly (comment)
(B200) [Nightly] MoE Refactor DP Integration Shard 3/3	Flaky on main: same DeepEP LL server start timeout, 10/23 nightly failures in 14 days (ci.vllm.ai)
(L4) Distributed DP Basic Shard 2/2	Failing on main (test_mtp_sharded_sampling_equivalence, same prompts diverge)
(H200 MIG 18GB) [Nightly] Python-only Installation	Flaky on main: same wheel metadata 404
(H100) [Nightly] DeepSeek V2-Lite Prefetch Offload Accuracy	Agent lost on first attempt; retries in this build passed

---

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 #98344 for commit dbd45d5d77a8.

ℹ️ This PR is 1 commit behind upstream main. Merged upstream main into this branch as dbd45d5d77a8 before starting CI. Pull the branch before pushing more changes.

[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]

/ci run

[github-actions]

✅ Triggered Buildkite CI #98410 for commit 44d033587fa7.

[AndreasKaratzas]

/ci retry

[github-actions]

✅ Triggered Buildkite CI #98566 for commit 1c7addde6067, running 5 failed step(s) from Buildkite CI #98410.