Build Fix: Update abi3 tag to cp310 and minimum python version to 3.10 by aw920h · Pull Request #2532 · Dao-AILab/flash-attention

aw920h · 2026-05-02T16:43:20Z

Currently, the setup.py script hardcodes a cp39-abi3 tag on the output wheel. Since PyTorch 2.9+ dropped support for Python 3.9, this creates a misleading wheel tag. Installing this wheel in a Python 3.9 environment results in an ImportError: undefined symbol crash due to PyTorch C++ ABI mismatches.

Since Python 3.9 reached EOL in October 2025 and the ecosystem has moved on, this PR simplifies the build process and aligns it with PyTorch's support matrix:
Bumped Python Requirement: Globally updated the wheel tag to cp310 and set python_requires=">=3.10". This officially drops Python 3.9 support and prevents the ABI crash.

Closes Issue #2530

…nsion

Fix: Remove misleading py_limited_api=cp39 wheel tag for PyTorch extension

Add dynamic ABI tag based on PyTorch version for correct and improved naming of the wheel.

Implement dynamic ABI tagging for PyTorch versions

Copilot

Pull request overview

This PR updates the flash_attn_3 build/packaging logic in hopper/setup.py to avoid producing misleading cp39-abi3 wheel tags when building against newer PyTorch versions (e.g., 2.11+) that no longer support Python 3.9, preventing downstream ABI/symbol mismatch crashes.

Changes:

Adds PyTorch version detection logic in hopper/setup.py.
Dynamically sets bdist_wheel.py_limited_api to cp310 for PyTorch >= 2.11, otherwise cp39.
Replaces the previously hardcoded options={"bdist_wheel": {"py_limited_api": "cp39"}} with a computed dynamic_options.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

aw920h · 2026-05-02T18:24:31Z

+# Determine the minimum Python version supported by this PyTorch
+if torch_major_minor >= (2, 11):
+    # PyTorch 2.11 and newer require Python 3.10+
+    dynamic_abi_tag = "cp310"
+else:
+    # Older PyTorch versions support Python 3.9
+    dynamic_abi_tag = "cp39"


I applied the first suggestion to fix the parsing logic. For this second suggestion, I'd prefer to keep it simple with this if/else check for now, as it already fixes the immediate PyTorch 2.11 crash without adding complex metadata logic. Let me know if it is required.

Update: Applied the second suggestion in latest commit (cae80f3) and verified its workability.

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

aw920h · 2026-05-07T05:39:52Z

Hi @drisspg, could you please review the changes made for correct tagging based on pytorch version?

drisspg · 2026-05-07T13:34:04Z

@janeyx99

janeyx99 · 2026-05-08T04:40:15Z

This change makes sense and is an improvement from the existing process, but technically, we should target the minimum python supported by the torch specified in TORCH_TARGET_VERSION right?

Refactor dynamic ABI tag and Python version requirements based on installed PyTorch version and streamline CUDA extension arguments.

Restored some accidentally removed content

aw920h · 2026-05-08T13:45:40Z

Hi @janeyx99 , You are completely right!
I've updated the PR with the following changes (418670f):

The script now first checks os.environ.get("TORCH_TARGET_VERSION") and parses the hex string (e.g., 0x0208... -> 2.8). If it's not set, it safely falls back to parsing torch.__version__. If the resolved target is >= 2.9, it correctly tags the wheel as cp310 and sets python_requires=">=3.10". Otherwise, it falls back to cp39.
While testing this, I noticed the logic around line 580 was hardcoded -DTORCH_TARGET_VERSION=0x0209000000000000 ignoring the environment variable if the installed torch was >= 2.9, I updated this.
Also removed -DPy_LIMITED_API flag from the compiler arguments that was causing a redefinition warning. The build now correctly and cleanly uses the flag injected by py_limited_api=True.

I've tested this locally with both PyTorch 2.8 and 2.10, and the compiler flags and final wheel tags are generating perfectly now. Let me know if this looks good to you.

janeyx99 · 2026-05-08T13:47:20Z

The hardcoded 2.9 is intentional! We want people to be able to run with any torch 2.9+!

aw920h · 2026-05-08T14:04:18Z

@janeyx99 , Thank you for the clarification. I've reverted that specific change to restore the hardcoded -DTORCH_TARGET_VERSION=0x0209...
Also brought back -DPy_LIMITED_API.
The other fixe in the PR remain as mentioned. I've pushed the updated commit. Let me know if this looks correct now!

minor fix: cleaned up the comments

janeyx99

If we have to update the conditional logic for future versions of torch anyway, is the dynamic setting really that useful? basically, could this PR just bump the python abi3 version to 3.10 and call it a day?

janeyx99 · 2026-05-14T15:12:46Z

+        major, minor = map(int, torch.__version__.split('.')[:2])
+
+    # Since PyTorch 2.9+ dropped Python 3.9 support
+    if major >= 2 and minor >= 9:


Note that this will require adding more conditionals here for future versions that drop more python support

janeyx99 · 2026-05-14T15:13:30Z

+    if major >= 2 and minor >= 9:
        flash_api_source = "flash_api_stable.cpp"
-        stable_args = ["-DTORCH_TARGET_VERSION=0x0209000000000000"]  # Targets minimum runtime version torch 2.9.0
+        stable_args = ["-DTORCH_TARGET_VERSION=0x0209000000000000"]


why this change?

Updated the wheel tag to cp310 and python_requires=">=3.10".

aw920h · 2026-05-15T13:05:47Z

Hi @janeyx99 , to answer your question on the flash_api_stable.cpp change, I simplified that logic because, realistically, nobody is building against an August 2025 nightly anymore now that stable 2.9 through 2.12 are out. Based on that assumption, a simple major/minor check seemed cleaner and allowed me to easily parse the TORCH_TARGET_VERSION environment variable for cross-compilation. However, I realize that would have broken upon PyTorch 3.0, so it is best if the original parse() logic is kept.

Regarding bumping to cp310 globally, since other libraries have dropped Python 3.9, hardcoding cp310 means we officially drop Python 3.9 support even for users cross-compiling for PyTorch 2.8, Python 3.9 is EOL anyway.

I have simplified the PR to:

Globally bump the wheel tag to cp310 and python_requires=">=3.10".
Revert the API selection back to your original parse() logic.

janeyx99

update the PR title and description, i think this is good to go

aw920h · 2026-05-21T04:09:30Z

Done! Title and description have been updated to reflect the final changes. Thanks, @janeyx99!

…10 (Dao-AILab#2532) * Fix: Remove misleading py_limited_api=cp39 wheel tag for PyTorch extension * Implement dynamic ABI tagging for PyTorch versions Add dynamic ABI tag based on PyTorch version for correct and improved naming of the wheel. * Potential fix for pull request finding Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> * Update Python version requirements based on torch metadata * Refactor setup.py for dynamic ABI and CUDA settings Refactor dynamic ABI tag and Python version requirements based on installed PyTorch version and streamline CUDA extension arguments. * Update CUDAExtension compile arguments Restored some accidentally removed content * Update setup.py * Updated setup.py minor fix: cleaned up the comments * Brought back Py_LIMITED_API flag to CUDA extension compilation * Minor fix * Update setup.py for Python version requirements Updated the wheel tag to cp310 and python_requires=">=3.10". --------- Co-authored-by: aw920h <alien@alien.alien> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

aw920h and others added 4 commits May 2, 2026 03:21

Fix: Remove misleading py_limited_api=cp39 wheel tag for PyTorch exte…

3916812

…nsion

Merge pull request #1 from aw920h/fix-abi3-wheel-tag

9343372

Fix: Remove misleading py_limited_api=cp39 wheel tag for PyTorch extension

Implement dynamic ABI tagging for PyTorch versions

b202d0e

Add dynamic ABI tag based on PyTorch version for correct and improved naming of the wheel.

Merge pull request #2 from aw920h/fix-abi3-wheel-tag

3a736a1

Implement dynamic ABI tagging for PyTorch versions

Johnsonms requested review from Johnsonms, Copilot and tridao May 2, 2026 16:58

Copilot started reviewing on behalf of Johnsonms May 2, 2026 16:58 View session

Copilot AI reviewed May 2, 2026

View reviewed changes

aw920h and others added 2 commits May 2, 2026 23:49

Potential fix for pull request finding

c4c9280

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Update Python version requirements based on torch metadata

cae80f3

aw920h added 2 commits May 8, 2026 02:36

Refactor setup.py for dynamic ABI and CUDA settings

85a4cc4

Refactor dynamic ABI tag and Python version requirements based on installed PyTorch version and streamline CUDA extension arguments.

Update CUDAExtension compile arguments

418670f

Restored some accidentally removed content

Update setup.py

3b3beb6

aw920h added 3 commits May 8, 2026 07:18

Updated setup.py

453b396

minor fix: cleaned up the comments

Brought back Py_LIMITED_API flag to CUDA extension compilation

1664622

Minor fix

834a2cf

janeyx99 reviewed May 14, 2026

View reviewed changes

Update setup.py for Python version requirements

176407a

Updated the wheel tag to cp310 and python_requires=">=3.10".

janeyx99 reviewed May 20, 2026

View reviewed changes

janeyx99 approved these changes May 20, 2026

View reviewed changes

aw920h changed the title ~~Build Fix: Dynamically set abi3 tag based on PyTorch compatibility~~ Build Fix: Update abi3 tag to cp310 and minimum python version to 3.10 May 21, 2026

drisspg approved these changes May 22, 2026

View reviewed changes

drisspg merged commit 3da76cd into Dao-AILab:main May 22, 2026

aw920h mentioned this pull request May 22, 2026

[Packaging] setup.py incorrectly tags wheels as cp39-abi3 when compiled with PyTorch versions that require Python 3.10+ #2530

Closed

Conversation

aw920h commented May 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

aw920h May 2, 2026

Choose a reason for hiding this comment

Uh oh!

aw920h May 3, 2026

Choose a reason for hiding this comment

Uh oh!

aw920h commented May 7, 2026

Uh oh!

drisspg commented May 7, 2026

Uh oh!

janeyx99 commented May 8, 2026

Uh oh!

aw920h commented May 8, 2026

Uh oh!

janeyx99 commented May 8, 2026

Uh oh!

aw920h commented May 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

janeyx99 left a comment

Choose a reason for hiding this comment

Uh oh!

janeyx99 May 14, 2026

Choose a reason for hiding this comment

Uh oh!

janeyx99 May 14, 2026

Choose a reason for hiding this comment

Uh oh!

aw920h commented May 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

janeyx99 left a comment

Choose a reason for hiding this comment

Uh oh!

aw920h commented May 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

aw920h commented May 2, 2026 •

edited

Loading

aw920h commented May 8, 2026 •

edited

Loading

aw920h commented May 15, 2026 •

edited

Loading