vllm: changelog & breaking changes
Auto-tracked from github.com/vllm-project/vllm. Updated Jul 06, 2026.
v0.24.0
2026-06-29
âš Breaking changes
- Device selection change: vLLM no longer sets CUDA_VISIBLE_DEVICES internally; a new device_ids argument is provided instead (). On ROCm, a deprecation window for CUDA_VISIBLE_DEVICES has begun ().
- torch stable ABI: Continued (and completed) migration of kernels to the libtorch stable ABI — MoE 10c/n (), Marlin 11a/n (), Machete 11b/n (), final _C library migration 12/n ().
- Deprecations & Removals
- Removed models: ERNIE (obsolete) (), Xverse (), Dots1 (), Bamba (), Mono-InternVL (), InternLM registry alias ().
- Deprecated: First-generation Qwen and QwenVL models (), Transformers v4 support (), CUDA_VISIBLE_DEVICES on ROCm (); general deprecations for v0.23/v0.24 ().
v0.23.0
2026-06-15
âš Breaking changes
- Unified parser: Reasoning and tool-call parsing are now unified behind a single Parser.parse() interface (), with the Responses parser migrated to it ().
- Removed the stale cuDNN frontend upper bound (); Docker fixes for flashinfer-jit-cache (), FlashInfer CuTe DSL JIT libcublas-dev (), and CUTLASS DSL cu13 install order ().
- Deprecations
- Deprecate JAISLMHeadModel ().
- Begin the deprecation cycle for the NixlConnector kv_both role ().
- Remove functions previously scheduled for deprecation in v0.21.0 ().
v0.22.1
2026-06-05
✓ No breaking changes detected — looks safe to upgrade.
v0.22.0
2026-05-29
âš Breaking changes
- Kernel ABI: continued migration to libtorch stable ABI — 5/n (), 6/n (), 7/n ().
- Experimental: breakable CUDA graph ().
- EPLB: change default EPLB communicator (), VLM-wrapper init fix (), remove dead torch.accelerator.synchronize() ().
- Deprecations & Removals
- Removed old locations of get_tokenizer and resolve_hf_chat_template ().
- Removed deprecated MLA prefill arguments ().
v0.21.0
2026-05-15
âš Breaking changes
- Transformers v4 deprecated: This release formally deprecates transformers v4 support (). Users should migrate to transformers v5.
- C++20 build requirement: vLLM now requires a C++20-compatible compiler for compatibility with PyTorch (). This is a breaking build change.
- Breaking: C++20 required for PyTorch compatibility ().
- Breaking: Transformers v4 deprecated ().
v0.20.2
2026-05-10
âš Breaking changes
- Qwen3-VL: Removed an invalid deepstack boundary check that could fail under heavy load ().
v0.20.1
2026-05-04
✓ No breaking changes detected — looks safe to upgrade.
v0.20.0
2026-04-27
âš Breaking changes
- PyTorch 2.11 upgrade (): vLLM ships on torch 2.11 for CUDA, and XPU is now also on torch 2.11 () — XPU is no longer pinned to 2.10. This is a breaking change for environment dependency.
- Deprecations: Petit NVFP4 removed ().
- PyTorch 2.11 for CUDA () and XPU () — XPU no longer pinned to 2.10.
- Removed / moved: resampy dependency dropped (), librosa direct dependency dropped (), pyav and soundfile moved to common requirements ().
- Breaking Changes
- 3. Metrics rework: vllm:prompt_tokens_recomputed removed (); num_cached_tokens / num_external_computed_tokens replaced with PrefillStats ().
⚡ Want an alert the moment your stack breaks — not a weekly digest? ChangeRadar Pro adds instant alerts + unlimited libraries + AI migration steps.Go Pro — $9/mo
Related breaking-change trackers