Skip to content

Avoid using block_m=64 when having mxfp8 downcast in epilogue #8830

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

Draft
qnie-oai wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Avoid using block_m=64 when having mxfp8 downcast in epilogue #8830

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
5 changes: 5 additions & 0 deletions python/triton_kernels/tests/test_matmul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,6 +173,11 @@ def _build_test_op_cases():
Case(*even_shape, "ragged", "bfloat16", "bfloat16", epilogue_subtile=val, swiglu_opts=(1.1, 1.4))
for val in (1, 2, 4)
])
# swiglu together with mxfp8 downcastepilogue
test_cases.extend([
Case(*shape, mode, "mxfloat8_e4m3fn", "mxfloat4_e2m1", hbm_swizzling=True, split_k=split_k, swiglu_opts=(1.1, 7))
for shape in [odd_shape2, even_shape] for mode in ["ragged", "batched"] for split_k in [1, 5]
])

return test_cases

Expand Down
3 changes: 3 additions & 0 deletions python/triton_kernels/triton_kernels/matmul_details/opt_flags.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,6 +210,9 @@ def make_default_opt_flags_nvidia(
block_m = max(16, min(triton.next_power_of_2(8 * slice_size), 128))
else:
block_m = max(16, min(triton.next_power_of_2(2 * slice_size), 64))
if block_m == 64 and precision_config.c_mx_scale is not None and rhs_dtype == FP4 and torch.cuda.get_device_capability()[0] >= 10:
# when having both fused_activation and mxfp8 downcast in epilogue, block_m=64 causing shared memory overflow
block_m = 128
else:
block_m = max(16, min(triton.next_power_of_2(slice_size), 128))
# block n
Expand Down