Skip to content

[Bugfix] Skinny GEMM in tuned gemm.py: add output conversion to tuned_gemm.mm #665

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
valarLip merged 4 commits into from
Jul 25, 2025
Merged

[Bugfix] Skinny GEMM in tuned gemm.py: add output conversion to tuned_gemm.mm #665

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
057928d
add output conversion to tuned_gemm.mm
vllmellm Jul 16, 2025
3dbe153
fix black lint
vllmellm Jul 21, 2025
1b134fb
Merge branch 'main' into main
vllmellm Jul 21, 2025
c652c26
Merge branch 'main' into main
valarLip Jul 24, 2025
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
2 changes: 2 additions & 0 deletions aiter/tuned_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,6 +300,8 @@ def mm(
)
if batched:
out = out.view(*inp.shape[:-1], weights.shape[0])
if otype is not None:
out = out.to(otype)
return out


Expand Down
17 changes: 15 additions & 2 deletions op_tests/test_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,17 @@ def test_gemm(dtype, m, n, k, bias=False, otype=None, scaleA=None, scaleB=None):
(a, *_), avg_a = run_torch(x, weight, bias, otype, scaleA, scaleB)
(b, *_), avg_b = run_gemm_b(x, weight, bias, otype, scaleA, scaleB)

assert (
a.dtype == b.dtype
), f"Expected a.dtype == b.dtype, but a={a.dtype}, b={b.dtype}, input dtype={dtype}"
if otype is not None:
assert (
a.dtype == otype
), f"a={a.dtype}, expected output dtype={otype}, input dtype={dtype}"
assert (
b.dtype == otype
), f"b={b.dtype}, expected output dtype={otype}, input dtype={dtype}"

msg = f"[perf] dim: {str(dim):<20} dtype: {dtype}, torch avg: {avg_a:<8.2f} us, B avg: {avg_b:<8.2f} us, uplift: {avg_a/avg_b-1:<5.1%}"
checkAllclose(a, b, msg=msg)

Expand Down Expand Up @@ -274,7 +285,8 @@ def test_normal_gemm():
)
test_gemm(dtypes.bf16, 128, 32, 8192)
for dtype in [dtypes.fp16, dtypes.bf16]:
test_gemm(dtype, 128, 32, 8192)
for otype in [None, dtypes.fp16, dtypes.bf16, dtypes.fp32]:
test_gemm(dtype, 128, 32, 8192, otype=otype)
# # qkv_proj
# for (m, n, k) in [(4096, 1280, 8192),
# (128, 1280, 8192),
Expand Down Expand Up @@ -337,7 +349,8 @@ def test_skinny_gemm():
for mnk in test_mnk_list:
m, n, k = mnk
for dtype in [dtypes.fp16, dtypes.bf16]:
test_gemm(dtype, m, n, k)
for otype in [None, dtypes.fp16, dtypes.bf16, dtypes.fp32]:
test_gemm(dtype, m, n, k, otype=otype)


# test_normal_gemm()
Expand Down