Skip to content

[TEST] Modify mxfp tests to allow faster runtime #5602

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

Merged
etiotto merged 1 commit into from
Dec 3, 2025
Merged

[TEST] Modify mxfp tests to allow faster runtime #5602

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 3 additions & 3 deletions python/test/unit/intel/test_mxfp_matmul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,9 +36,9 @@ def mxfp_matmul( #
a_scale, b_scale, #
M, N, K, #
stride_scale, #
stride_am, stride_ak, #
stride_bk, stride_bn, #
stride_cm, stride_cn, #
stride_am: tl.constexpr, stride_ak: tl.constexpr, #
Copy link
Contributor

@anmyachev anmyachev Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Doesn't this increase compilation time? So, each new value requires a recompile, right?

Copy link
Contributor Author

@whitneywhtsang whitneywhtsang Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, but without constexpr, 2d block io cannot be used, and runtime is slow.

stride_bk: tl.constexpr, stride_bn: tl.constexpr, #
stride_cm: tl.constexpr, stride_cn: tl.constexpr, #
DTYPE_A: tl.constexpr, #
DTYPE_B: tl.constexpr, #
BLOCK_M: tl.constexpr, #
Expand Down
4 changes: 2 additions & 2 deletions python/test/unit/language/test_matmul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1252,9 +1252,9 @@ def create_operand(dtype: str, size0: int, size1: int, k_dim: int, transpose: bo
kernel_kwargs = {}
if is_hip():
kernel_kwargs["matrix_instr_nonkdim"] = nonKDim
if is_xpu() and (128, 256, 256) == (BLOCK_M, BLOCK_N, BLOCK_K) and not CONST_SCALE and not PACK_B_ALONG_K:
kernel_kwargs["num_warps"] = 8
if is_xpu():
# since the block size are big we use num_warps = 32 to avoid pressure problems.
kernel_kwargs["num_warps"] = 32
Copy link
Contributor

@anmyachev anmyachev Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm just curious how you chose it. Why not 16, for example?

Copy link
Contributor Author

@whitneywhtsang whitneywhtsang Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

32 is the max with double GRF.

kernel_kwargs["grf_mode"] = "256"
out = mxfp8_mxfp4_matmul[grid](a, b, output, a_scale, b_scale, M, N, K, stride_scale, a.stride(0), a.stride(1),
b.stride(0), b.stride(1), output.stride(0), output.stride(1), not CONST_SCALE,
Expand Down
Loading