Support fp8 block gemm with fp8_e8m0 scale#398
Open
Yejing-Lai wants to merge 1 commit into
Open
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
Note
Copilot was unable to run its full agentic suite in this review.
This PR adds optional UE8M0 (float8 e8m0) scale support to the FP8 quantization reference utilities and expands the oneDNN FP8 GEMM test matrix to cover that mode.
Changes:
- Parameterize FP8 GEMM per-block tests to run with/without UE8M0 scales.
- Extend
fp8_block_quant_2dto optionally round scales to powers-of-two and return UE8M0 scale tensors; make dequant paths robust by casting scales to FP32. - Update oneDNN FP8 matmul scale handling and primitive caching key to account for UE8M0 usage.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.
| File | Description |
|---|---|
| tests/test_fp8_gemm_onednn.py | Adds use_ue8m0 parametrization and threads it into quantization helpers. |
| tests/ops/fp8_quant_op.py | Adds UE8M0 scale option for block quantization and ensures dequant casts scales to FP32. |
| csrc/xpu/onednn/fp8_gemm_w8a8.h | Adjusts scale-mode detection, scale attribute setup, and matmul primitive cache key. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
166
to
+171
| int m1_sc_group_size = m1_sc.numel(); | ||
| int m2_sc_group_size = m2_sc.numel(); | ||
| int sc_group_size = (m1_sc_group_size << 8) | m2_sc_group_size; | ||
| if (m1_sc_dtype == at::ScalarType::Float8_e8m0fnu) { | ||
| sc_group_size |= (1 << 30); | ||
| } |
Comment on lines
+28
to
+31
| // MXFP8 weight scale is [k/32, n] | ||
| bool is_block_quant = (m1_sc.dim() == 2) && (m2_sc.dim() == 2) && | ||
| (m1_sc.size(1) != 1) && (m2_sc.size(0) != 1) && | ||
| (m2_sc.size(1) != n); |
| block_m: block rows | ||
| block_n: block cols | ||
| fp8_dtype: torch.float8_e4m3fn | ||
| use_ue8m0: return scales as torch.float8_e8m0fnu |
jikunshang
approved these changes
Jun 8, 2026
jikunshang
left a comment
jikunshang
left a comment
Member
There was a problem hiding this comment.
LGTM. thanks for fixing.
cc @wuxun-zhang
wuxun-zhang
approved these changes
Jun 8, 2026
Signed-off-by: Lai, Yejing <yejing.lai@intel.com>
Yejing-Lai
force-pushed
the
lyj/update_fp8_gemm
branch
from
ce6d92e to
9c69b50
Compare
Contributor
Author
|
@zufangzhu Please review~ |
zufangzhu
approved these changes
Jun 8, 2026
zufangzhu
left a comment
zufangzhu
left a comment
Collaborator
There was a problem hiding this comment.
LGTM. BTW, do we need to update w8a16 as well?
Contributor
Author
w8a16 not include MXFP8 path. It can be kept as it is. |
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
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.
Support FP8 block GEMM with fp8_e8m0 scale and add related UT.
Clean the code that if-else path is too long.