Conversation
📝 WalkthroughWalkthroughThis PR updates CI workflow matrices to include vllm in the build extras for specific test configurations and bumps the vllm version constraint in setup.py from 0.11.1 to 0.13.0 for Python 3.11+ environments. Changes
Estimated code review effort🎯 2 (Simple) | ⏱️ ~8 minutes Possibly related PRs
Suggested labels
Suggested reviewers
🚥 Pre-merge checks | ✅ 1 | ❌ 2❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing touches
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (3)
.github/workflows/multi-gpu-e2e.yml.github/workflows/tests.ymlsetup.py
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (7)
- GitHub Check: PyTest from Source Dist (3.11, 2.9.1)
- GitHub Check: PyTest from Source Dist (3.11, 2.8.0)
- GitHub Check: PyTest from Source Dist (3.11, 2.9.0)
- GitHub Check: PyTest (3.11, 2.9.1)
- GitHub Check: PyTest (3.11, 2.9.0)
- GitHub Check: PyTest (3.11, 2.8.0)
- GitHub Check: test-axolotl-multigpu (128, 12.8.1, 3.11, 2.8.0, fbgemm-gpu, 2, true)
🔇 Additional comments (2)
.github/workflows/multi-gpu-e2e.yml (1)
43-43: LGTM! Consistent addition of vllm to multi-GPU test configuration.The addition of vllm to the extras list for the PyTorch 2.9.1 test configuration aligns with the version constraint update in setup.py.
.github/workflows/tests.yml (1)
318-318: LGTM! Consistent addition of vllm to single-GPU test configuration.The addition of vllm to the extras list for the PyTorch 2.9.1 test configuration ensures vllm functionality is tested in the CI pipeline.
| extras_require_map.pop("fbgemm-gpu") | ||
| extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.4.1"] | ||
| extras_require_map["vllm"] = ["vllm==0.11.1"] | ||
| extras_require_map["vllm"] = ["vllm==0.13.0"] |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
# Check vllm version availability on PyPI
echo "=== Checking vllm versions on PyPI ==="
curl -s https://pypi.org/pypi/vllm/json | jq -r '.releases | keys[]' | grep -E '^0\.(11|12|13)\.' | sort -VRepository: axolotl-ai-cloud/axolotl
Length of output: 141
🌐 Web query:
vllm 0.13.0 release notes breaking changes compatibility
💡 Result:
Here are the key breaking changes and compatibility notes for vLLM v0.13.0 (summary from the project release/changelog):
Breaking changes / API/config changes
- PyTorch / environment dependency changes: vLLM requires newer runtime versions (PyTorch 2.8.0 / updated CUDA toolchain noted as a breaking dependency in the project changelog). [2]
- Removed/deprecated V0 features and backends (examples in the changelog: V0 Neuron backend, V0 pooling model support, V0 FlashInfer attention backend; several API deprecations such as prompt_token_ids fallback removal). [2]
- Quantization changes: some older/legacy quantization formats/hacks were removed (users must migrate to supported quantization paths). [2]
- Config/key rename and removals (vLLM Ascend notes): chunked_prefill_for_mla in additional_config was dropped; dump_config renamed to dump_config_path and changed from dict → string. [1]
- vLLM Ascend: vLLM version upgraded to 0.13.0 and support for v0.12.0 was dropped (Ascend-specific). [1]
Compatibility / dependency requirements
- vLLM v0.13.0 packages published on PyPI (0.13.0 release artifacts available). [3]
- vLLM Ascend compatibility matrix (for the Ascend build of vLLM v0.13.0rc1): Python >=3.10 and <3.12; tied to vLLM v0.13.0; specific Stable CANN / torch_npu versions listed for Ascend deployments. (See the matrix for exact platform bindings). [1]
- General project notes: environment updates (CUDA / driver / PyTorch) are required for the breaking dependency changes introduced in recent 0.13.x line — check your CUDA/PyTorch/tooling versions before upgrading. [2]
If you want, I can:
- fetch the full v0.13.0 changelog (raw GitHub release/compare) and extract the top 10 breaking/API changes with line-by-line source links; or
- check specific compatibility for your platform (CUDA, Python, PyTorch) — tell me your OS, Python and PyTorch/CUDA versions.
Sources:
[1] vLLM Ascend release notes & compatibility matrix (v0.13.0rc1).
[2] vLLM project GitHub release changelog (breaking changes summary).
[3] vLLM 0.13.0 artifacts on PyPI.
🌐 Web query:
vllm 0.13.0 PyTorch 2.9 compatibility
💡 Result:
Short answer: vLLM 0.13.0 is not officially compatible with PyTorch 2.9. vLLM 0.13.0 packages and third‑party plugins reference torch 2.7.x (e.g., 2.7.1), and multiple GitHub issues/threads report breakages and binary/CUDA incompatibilities when using PyTorch 2.9. If you need PyTorch 2.9 you must expect to build vLLM from source and debug CUDA/toolchain mismatches — it’s not a drop‑in supported configuration. [1][2][3]
Sources:
- vllm‑musa plugin compatibility matrix (lists vLLM 0.13.0 with torch 2.7.1). [1]
- vLLM installation notes (binary builds tied to specific CUDA/PyTorch; recommends building from source for nonstandard PyTorch/CUDA). [2]
- vLLM GitHub issues reporting problems and requests around PyTorch 2.9 (users report bugs/incompatibilities). [3]
vllm 0.13.0 is incompatible with PyTorch >= 2.9 — do not merge without resolution.
vllm 0.13.0 officially supports PyTorch 2.7.x, not 2.9+. PyPI builds are compiled against torch 2.7.1, and multiple reports confirm breakages when used with PyTorch 2.9. Building from source is required for PyTorch 2.9 compatibility, which is not a supported or tested configuration.
Additionally, vllm 0.13.0 introduces breaking changes from 0.11.1:
- Removed V0 backends and features
- Deprecated quantization formats
- Config/API changes (e.g.,
dump_configrenamed todump_config_path) - Minimum PyTorch requirement is 2.8.0
Either pin vllm to 0.12.0 (or stay at 0.11.1), reduce PyTorch to 2.7.x, or document that vllm 0.13.0 requires custom source builds for PyTorch 2.9.
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
winglian
force-pushed
the
vllm-2026
branch
2 times, most recently
from
09d46ad to
0b54d75
Compare
|
📖 Documentation Preview: https://6971554939be4b2227e10f30--resonant-treacle-0fd729.netlify.app Deployed on Netlify from commit e5fe9b8 |
2 participants
Description
Motivation and Context
How has this been tested?
Screenshots (if appropriate)
Types of changes
Social Handles (Optional)
Summary by CodeRabbit
✏️ Tip: You can customize this high-level summary in your review settings.