[qgallouedec]

Checked the v0.27.0 release notes against the vLLM surface TRL actually uses: No breaking changes for TRL. (v0.27.1 is a patch release with a single change, irrelevant to TRL.)

Relevant

• RL: stateful trainer-send IPC ([rl] Stateful Trainer Send: IPC [2/N] vllm-project/vllm#48981) and weight version tagging for RL rollouts ([Core][Frontend] Add weight version tagging for RL rollouts vllm-project/vllm#49040): touch vllm.distributed.weight_transfer, which the experimental async GRPO weight transfer imports — the [2/N] step of the migration tracked in Migrate async GRPO weight transfer to vLLM's stateful engine API #6732. Nothing breaks: 0.27.1 ships the new stateful factory.py alongside the static functions TRL uses, which still resolve unchanged. The deletion step ([4/N]) is unwritten upstream, so this bump doesn't force the migration.
• vllm config set during weight reload ([Bugfix][RL] Set vLLM config during weight reload vllm-project/vllm#45989): internal to the reload path, no API change.
• PyTorch 2.13.0 / torchvision 0.28.0 / Triton 3.7.1 ([Core] Update PyTorch to 2.13.0, torchvision to 0.28.0, triton to 3.7.1 vllm-project/vllm#48155): a breaking environment change on vLLM's side, but TRL doesn't pin torch, so nothing to do here beyond noting that installing trl[vllm] at 0.27.x pulls torch 2.13.
• Removed max_num_partial_prefills / max_long_partial_prefills ([Misc] Remove old now unsupported max_num_partial_prefills and max_long_partial_prefills vllm-project/vllm#49244): TRL never sets either;trl vllm-serve doesn't expose them.
• 400 instead of 500 for non-numeric logprobs (fix(openai): reject non-numeric logprobs with 400 instead of 500 vllm-project/vllm#49144): OpenAI-server-side only; TRL's /generate endpoint and extract_logprobs are unaffected.

also:

• Transformers 5.14.1 bump (Bump Transformers version to 5.14.1 vllm-project/vllm#49223): TRL is already v5-aligned (transformers>=4.56.2).
• Every other vLLM symbol TRL imports still resolves at v0.27.1. All LLM(...) kwargs trl vllm-serve passes are still valid engine args.

Support window

vLLM 0.27.0 was released on 10 Aug 2026 and 0.27.1 on 11 Aug 2026, so on the rolling 5-month window they age out around 10 and 11 Jan 2027. Next scheduled drop is 0.18.0 (released 20 Mar 2026 → out on 20 Aug 2026).

Test results

Run against vLLM 0.27.1 (torch 2.13.0, transformers 5.15.0.dev0), 4×H100:

$ pytest -v tests/test_vllm_client_server.py
============================= test session starts ==============================
platform linux -- Python 3.13.13, pytest-9.1.1, pluggy-1.6.0 -- /fsx/qgallouedec/miniconda3/envs/trl/bin/python
cachedir: .pytest_cache
rootdir: /fsx/qgallouedec/trl
configfile: pyproject.toml
plugins: anyio-4.14.2, xdist-3.8.0
collecting ... collected 42 items

tests/test_vllm_client_server.py::TestChunkList::test_even_split PASSED  [  2%]
tests/test_vllm_client_server.py::TestChunkList::test_uneven_split PASSED [  4%]
tests/test_vllm_client_server.py::TestChunkList::test_more_chunks_than_elements PASSED [  7%]
tests/test_vllm_client_server.py::TestChunkList::test_n_equals_len PASSED [  9%]
tests/test_vllm_client_server.py::TestChunkList::test_n_is_1 PASSED      [ 11%]
tests/test_vllm_client_server.py::TestChunkList::test_single_element_list PASSED [ 14%]
tests/test_vllm_client_server.py::TestChunkList::test_any_dtype PASSED   [ 16%]
tests/test_vllm_client_server.py::TestExtractLogprobs::test_extract_logprobs_sorts_by_rank_and_replaces_nan PASSED [ 19%]
tests/test_vllm_client_server.py::TestExtractLogprobs::test_extract_logprobs_returns_none_token_ids_when_logprobs_missing PASSED [ 21%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_generate PASSED [ 23%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_generate_with_logprobs_none PASSED [ 26%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_chat PASSED [ 28%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_chat_with_logprobs_none PASSED [ 30%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_chat_with_tools PASSED [ 33%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_generate_with_token_ids PASSED [ 35%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_generate_with_params PASSED [ 38%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_update_model_params PASSED [ 40%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_reset_prefix_cache PASSED [ 42%]
tests/test_vllm_client_server.py::TestVLLMClientServer::test_logprobs_match_with_non_default_sampling XFAIL [ 45%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_generate PASSED [ 47%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_generate_with_logprobs_none PASSED [ 50%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_chat PASSED [ 52%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_chat_with_logprobs_none PASSED [ 54%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_chat_with_tools PASSED [ 57%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_generate_with_token_ids PASSED [ 59%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_generate_with_params PASSED [ 61%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_update_model_params PASSED [ 64%]
tests/test_vllm_client_server.py::TestVLLMClientServerBaseURL::test_reset_prefix_cache PASSED [ 66%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_generate PASSED [ 69%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_generate_with_logprobs_none PASSED [ 71%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_chat PASSED [ 73%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_chat_with_logprobs_none PASSED [ 76%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_chat_with_tools PASSED [ 78%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_generate_with_token_ids PASSED [ 80%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_generate_with_params PASSED [ 83%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_update_model_params PASSED [ 85%]
tests/test_vllm_client_server.py::TestVLLMClientServerTP::test_reset_prefix_cache PASSED [ 88%]
tests/test_vllm_client_server.py::TestVLLMClientServerDeviceParameter::test_init_communicator_with_device_int PASSED [ 90%]
tests/test_vllm_client_server.py::TestVLLMClientServerDeviceParameter::test_init_communicator_with_device_string PASSED [ 92%]
tests/test_vllm_client_server.py::TestVLLMClientServerDeviceParameter::test_init_communicator_with_torch_device PASSED [ 95%]
tests/test_vllm_client_server.py::TestVLLMClientServerVLM::test_generate_with_token_ids_and_image PASSED [ 97%]
tests/test_vllm_client_server.py::TestVLLMClientServerVLM::test_generate_with_token_ids_mixed_images PASSED [100%]

================== 41 passed, 1 xfailed in 1018.52s (0:16:58) ==================

No warnings: the version-range check passes on 0.27.1 with this change applied.

---

Note

Low Risk
Version-cap and documentation-only changes; no logic changes to vLLM training or serving beyond allowing 0.27.x installs.

Overview
Extends the supported vLLM range from 0.26.0 to 0.27.1 so TRL users can install and run recent vLLM releases without compatibility warnings.

The upper bound is updated consistently in docs/source/vllm_integration.md, the trl[vllm] extra in pyproject.toml (vllm>=0.18.0,<=0.27.1), and is_vllm_available() in trl/import_utils.py, including the warning text when an installed version falls outside the window. No TRL integration code paths change beyond these version gates.

^{Reviewed by Cursor Bugbot for commit 0b9a0c9. Bugbot is set up for automated code reviews on this repo. Configure here.}

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.