avoid propagation of NaN #3723
Closed
avoid propagation of NaN #3723
Conversation
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
|
This pull request was exported from Phabricator. Differential Revision: D69522001 |
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify site configuration. |
Aya-ZIbra
added a commit
to Aya-ZIbra/FBGEMM
that referenced
this pull request
Feb 22, 2025
Summary: X-link: facebookresearch/FBGEMM#806 as title Introduce padding in dequantization kernel to avoid passing of NaNs to the output of FA3 in prefill stage. Reviewed By: jianyuh Differential Revision: D69522001
Aya-ZIbra
force-pushed
the
export-D69522001
branch
from
4afe692 to
6984077
Compare
|
This pull request was exported from Phabricator. Differential Revision: D69522001 |
Aya-ZIbra
force-pushed
the
export-D69522001
branch
from
6984077 to
481519e
Compare
Aya-ZIbra
added a commit
to Aya-ZIbra/FBGEMM
that referenced
this pull request
Feb 22, 2025
Summary: X-link: facebookresearch/FBGEMM#806 as title Introduce padding in dequantization kernel to avoid passing of NaNs to the output of FA3 in prefill stage. Reviewed By: jianyuh Differential Revision: D69522001
|
This pull request was exported from Phabricator. Differential Revision: D69522001 |
Summary: X-link: facebookresearch/FBGEMM#806 as title Introduce padding in dequantization kernel to avoid passing of NaNs to the output of FA3 in prefill stage. Reviewed By: jianyuh Differential Revision: D69522001
Aya-ZIbra
force-pushed
the
export-D69522001
branch
from
481519e to
9e31670
Compare
|
This pull request was exported from Phabricator. Differential Revision: D69522001 |
sijiac
approved these changes
Feb 23, 2025
| @@ -1919,23 +1928,27 @@ std::tuple<at::Tensor, at::Tensor> dequantize_fp8_cache( | |||
| block_tables_b_stride = block_tables.value().stride(0); | |||
| } | |||
|
|
|||
| constexpr int32_t kMaxBlocks = 256; | |||
| constexpr int32_t kMaxBlocks = 512; | |||
There was a problem hiding this comment.
is this change for better performance (increased parallelism)?
| // each thread writes 4 elements of type bf16 | ||
| *reinterpret_cast<uint2*>(&row_k_dq[4 * threadIdx.x]) = | ||
| *reinterpret_cast<uint2*>(&kv_dq.vals[0]); | ||
| *reinterpret_cast<uint2*>(&row_v_dq[4 * threadIdx.x]) = | ||
| *reinterpret_cast<uint2*>(&kv_dq.vals[2]); | ||
| } | ||
|
|
There was a problem hiding this comment.
nit, maybe add some comments to explain why we do the padding of the last tile?
| or not HAS_XFORMERS, | ||
| "Skip when H100 is not available or MI300 is not available", | ||
| "Skip when H100 is not available", | ||
| ) | ||
| def test_fp8_kv_cache(self, MAX_T: int, N_KVH_L: int) -> None: |
There was a problem hiding this comment.
nit, update the test to check the padding logic?
|
This pull request has been merged in e97b388. |
q10
pushed a commit
to q10/FBGEMM
that referenced
this pull request
Apr 10, 2025
Summary: X-link: pytorch#3723 Pull Request resolved: facebookresearch/FBGEMM#806 as title Introduce padding in dequantization kernel to avoid passing of NaNs to the output of FA3 in prefill stage. Reviewed By: jianyuh Differential Revision: D69522001 fbshipit-source-id: 9ce8c1840be75c78727e952feb1fbb962c57543a
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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:
X-link: https://github.com/facebookresearch/FBGEMM/pull/806
as title
Introduce padding in dequantization kernel to avoid passing of NaNs to the output of FA3 in prefill stage.
Reviewed By: jianyuh
Differential Revision: D69522001