use python version agnostic binding for mxfp8 cuda kernels#3471
Merged
danielvegamyhre merged 3 commits intomainfrom Dec 16, 2025
Merged
use python version agnostic binding for mxfp8 cuda kernels#3471danielvegamyhre merged 3 commits intomainfrom
danielvegamyhre merged 3 commits intomainfrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3471
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit 39db553 with merge base 08e5e20 ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
danielvegamyhre
added
topic: bug fix
Contributor
|
spoke offline, we need to be using torchlib and not use pybind |
meta-cla
bot
added
the
CLA Signed
danielvegamyhre
changed the title
danielvegamyhre
changed the title
danielvegamyhre
changed the title
danielvegamyhre
force-pushed
the
pyag
branch
from
4e59508 to
78a8b79
Compare
danielvegamyhre
commented
Dec 10, 2025
| not is_cuda_version_at_least(12, 8), | ||
| reason="CUDA version >= 12.8 required for MXFP8 CUDA kernels", | ||
| ) | ||
| def test_cuda_mx_dim1_invalid_block_size(): |
Contributor
Author
There was a problem hiding this comment.
deleting this test since block_size of 32 is hard coded in the python wrapper for the kernel now, since we always use this for mxfp8
danielvegamyhre
force-pushed
the
pyag
branch
from
78a8b79 to
e576e41
Compare
Contributor
|
you should be able to use nm -D to also investigate the symbols and ensure there are none from python |
Contributor
Author
|
current CI failures will be resolved once this rollback in upstream pytorch is included in the next torch nightly: pytorch/pytorch#169985 |
Contributor
Author
|
other upstream pytorch issue causing CI issues has now been resolved: pytorch/pytorch#170184 |
drisspg
approved these changes
Dec 16, 2025
Contributor
Author
|
mac-os test failure is unrelated |
danielvegamyhre
added a commit
that referenced
this pull request
Dec 16, 2025
* use py agnostic c++ extension for mxfp8_cuda * refactor mxfp8 cuda from pybind to torch_library api * put schema def inside guard
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
torchao/ops.pyfor other CUDA C++ extensionstorchao._C_mxfp8so file (lands in build/ dir) instead of separate torchao.prototype.mxpf8_cuda`extension (landed in torchao/prototype)Context
While doing the 0.15.0 torchao release and testing the test build for cuda 12.8, and i found the "torchao.prototype.mxfp8_cuda" c++ extension cannot be found (import error, module not found). we only build the extension for cuda 12.8+, so i checked the logs and i see logs indicating it was built: https://github.com/pytorch/ao/actions/runs/20046209265/job/57498462190
so then i checked the local installation itself, and i do see a .so file for the extension in the torchao/prototype dir, so it is definitely being built.
i tried asking claude about this and it says the build for python3.10 must match the python version in the conda env due to ABI incompatibility (i'm using python 3.12). as a test, i tried a fresh conda env with python 3.10, and instead of module not found, i get an undefined symbol error, so that does seem to indicate some python ABI issue.
asking @drisspg he said we should be building with a py agnostic flag, so i looked into this and we are doing this for other c++ extensions but not mxfp8_cuda, so I am fairly certain this is the root cause and this PR will fix the issue.