Skip to content

[Test] Use UE8M0 weight scales in block-FP8 DeepGEMM matmul test when the E8M0 oracle is active #51372

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
BabyDrangoner wants to merge 2 commits into from
Closed

[Test] Use UE8M0 weight scales in block-FP8 DeepGEMM matmul test when the E8M0 oracle is active #51372

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cf60853
[Test] Use UE8M0 weight scales in block-FP8 DeepGEMM matmul test when…
BabyDrangoner Aug 7, 2026
dfee40a
Merge branch 'main' into test/block-fp8-deepgemm-ue8m0-weights
BabyDrangoner Aug 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion tests/kernels/quantization/test_block_fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
from vllm.utils.deep_gemm import (
fp8_gemm_nt,
get_tma_aligned_size,
is_deep_gemm_e8m0_used,
per_block_cast_to_fp8,
should_use_deepgemm_for_fp8_linear,
)
Expand Down Expand Up @@ -287,7 +288,13 @@ def test_w8a8_block_fp8_deep_gemm_matmul(M, N, K, block_size, out_dtype, seed):
A_fp8, As_fp8 = per_token_group_quant_fp8(
A_fp32, block_size[1], column_major_scales=True, tma_aligned_scales=True
)
B_fp8, Bs_fp8 = per_block_cast_to_fp8(B_fp32, block_size=block_size)
# When the E8M0 oracle is active (default on Hopper/Blackwell), production
# requantizes weights to power-of-two scales before calling DeepGEMM
# (requant_weight_ue8m0_inplace). SM100/SM120 mxfp8 kernels require this;
# arbitrary float32 weight scales silently produce NaN on SM120.
B_fp8, Bs_fp8 = per_block_cast_to_fp8(
B_fp32, block_size=block_size, use_ue8m0=is_deep_gemm_e8m0_used()
)

As = As_fp8.to(torch.float32)
Bs = Bs_fp8.to(torch.float32)
Expand Down
Loading