Bump flash-attn from 2.4.2+7e2dd4d to 2.8.3.post1 - #240
Conversation
Dongxu-H
left a comment
There was a problem hiding this comment.
This PR replaces vendor-specific flash-attn builds with generic PyPI version, which may cause compatibility issues.
enflame backend uses torch 2.10.0 but flash-attn 2.8.3.post1 may not have the required GCU-specific optimizations.
metax backend uses torch 2.8.0 with metax-specific flash-attn build (2.6.3+metax3.7.2.0torch2.8). Replacing with generic version loses vendor optimizations.
kunlunxin backend flash-attn was customized (2.4.2+7e2dd4d), replacing with generic version may break compatibility.
Each backend should use flash-attn version that matches its torch version and vendor requirements. Dependabot does not understand vendor-specific torch/flash-attn version constraints.
This PR should be closed.
Dongxu-H
left a comment
There was a problem hiding this comment.
This PR replaces vendor-specific flash-attn builds with generic PyPI version 2.8.3.post1, which breaks all backend compatibility.
enflame backend uses flash-attn==2.7.2+torch.2.9.1.gcu which is tightly coupled with GCU hardware.
kunlunxin backend uses flash_attn==2.4.2+7e2dd4d which includes vendor patches.
metax backend uses flash_attn==2.6.3+metax3.7.2.0torch2.8 which is vendor-specific and not interchangeable with standard flash-attn.
Each backend has a vendor-specific flash-attn version that is tightly coupled with their hardware drivers. Replacing them with standard flash-attn 2.8.3.post1 will cause runtime failures.
Dependabot does not understand vendor-specific flash-attn version constraints. This PR should be closed.
Dongxu-H
left a comment
There was a problem hiding this comment.
flash-attn is a critical package and should use vendor-specific versions, not a unified version. The enflame, kunlunxin, and metax backends have vendor-specific flash-attn builds with platform-specific optimizations. Replacing them with a generic version may break compatibility or lose optimizations. Please keep the original vendor-specific versions.
Dongxu-H
left a comment
There was a problem hiding this comment.
flash-attn 2.8.3.post1 移除了厂商特定的版本后缀(如 +metax3.7.2.0torch2.8、+7e2dd4d)。这些后缀标识了针对特定硬件平台编译的版本。移除后缀可能导致 enflame、kunlunxin、metax 等后端无法正常使用 flash-attn。建议保留厂商特定版本或确认所有后端都支持标准版本。
63fdb97 to
52e0bcd
Compare
Dongxu-H
left a comment
There was a problem hiding this comment.
CI failed: code-style check failed. flash-attn version change from custom builds (2.4.2+7e2dd4d, 2.6.3+metax, 2.7.2+torch) to standard version 2.8.3.post1 may affect backend compatibility for enflame, kunlunxin, and metax. Verify that standard flash-attn works with these backends or if vendor-specific builds are still needed.
Dongxu-H
left a comment
There was a problem hiding this comment.
flash-attn 的厂商特定版本(如 +torch.2.9.1.gcu.3.4.20260323、+7e2dd4d、+metax3.7.2.0torch2.8)可能包含针对特定硬件的优化,直接升级到通用版本可能导致兼容性问题,建议确认各厂商是否支持标准版本。
Bumps [flash-attn](https://github.com/Dao-AILab/flash-attention) from 2.4.2+7e2dd4d to 2.8.3.post1. - [Release notes](https://github.com/Dao-AILab/flash-attention/releases) - [Commits](https://github.com/Dao-AILab/flash-attention/commits/v2.8.3.post1) --- updated-dependencies: - dependency-name: flash-attn dependency-version: 2.8.3.post1 dependency-type: direct:development update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com>
52e0bcd to
c5aa78d
Compare
Dongxu-H
left a comment
There was a problem hiding this comment.
CI 失败:code-style 检查失败,请修复代码风格问题。
|
OK, I won't notify you again about this release, but will get in touch when a new version is available. If you'd rather skip all updates until the next major or minor version, let me know by commenting If you change your mind, just re-open this PR and I'll resolve any conflicts on it. |
Bumps flash-attn from 2.4.2+7e2dd4d to 2.8.3.post1.
Commits