Skip to content

[MLA] Support 96-head 128-dim reduction #4625

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
LiuYinfeng01 merged 2 commits into from
Aug 10, 2026
+9 −2
Merged

[MLA] Support 96-head 128-dim reduction #4625

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a70e891
[MLA] Support 96-head 128-dim reduction
LiuYinfeng01 Aug 7, 2026
69d40da
[CI] Restore Black formatting in MLA reduce test
LiuYinfeng01 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
1 change: 1 addition & 0 deletions csrc/kernels/mla/reduce.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -935,6 +935,7 @@ __launch_bounds__(Traits::kNumThreads, Traits::kOccupancy) __global__
MLA_REDUCE_CASE_EF(NUM_HEAD, 64, HEAD_DIM, 64, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 64, HEAD_DIM, 128, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 64, HEAD_DIM, 512, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 96, HEAD_DIM, 128, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 128, HEAD_DIM, 128, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 128, HEAD_DIM, 512, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
MLA_REDUCE_CASE_EF(NUM_HEAD, 8, HEAD_DIM, 512, NUM_WG_PER_BH, NAME, __VA_ARGS__) \
Expand Down
10 changes: 8 additions & 2 deletions op_tests/test_mla_reduce.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,7 +273,12 @@ def run_case(splits_per_tile, num_heads, head_dim, out_dtype, qo_len=1):
atol=2e-2,
msg=f"[{tag}] lse ",
)
ok = (err_out == 0) and (err_lse == 0)
finite = (
torch.isfinite(gpu_out).all().item() and torch.isfinite(gpu_lse).all().item()
)
if not finite:
print(f"FAIL [{tag}] non-finite kernel output")
ok = finite and (err_out == 0) and (err_lse == 0)
print(f"{'PASS' if ok else 'FAIL'} {tag} {us:>8.2f} us")

# Bytes moved by the reduce: read all partial outputs + partial lses +
Expand Down Expand Up @@ -331,7 +336,8 @@ def main():

# (head_dim -> supported head counts from MLA_REDUCE_ROUTER in reduce.cu)
heads_for_dim = {
128: [1, 16, 128],
# Kimi K3 PP8 uses the 96-head, 128-dimension reducer.
128: [1, 16, 96, 128],
512: [8, 16, 128],
}

Expand Down
Loading