Skip to content

Tune MHA config & small-head pipeline pathology #4414

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
brunomazzottiamd merged 1 commit into from
Aug 5, 2026
Merged

Tune MHA config & small-head pipeline pathology #4414

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
7 changes: 7 additions & 0 deletions aiter/ops/triton/_triton_kernels/attention/mha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -935,6 +935,7 @@ def _get_config(
enable_dropout: bool,
dtype: torch.dtype,
has_pe: bool = False,
head_dim_v: int | None = None,
):
if not hasattr(_get_config, "_config_dict"):
dev = arch_info.get_arch()
Expand All @@ -952,5 +953,11 @@ def _get_config(
return fwd_cfg["pe"]
elif enable_dropout or dtype == torch.float32:
return fwd_cfg["dropout_or_fp32"]
elif head_dim_v is not None and 16 < head_dim_v <= 64 and "small_head" in fwd_cfg:
# Mid-small V head dims (16 < d <= 64) hit a num_stages=1 software-pipelining
# pathology on this backend (e.g. ~3x slower at d64). Using num_stages=3
# recovers performance and is numerically verified for these dims, but
# regresses d128 and miscompiles d<=16, so only 16 < d <= 64 uses this path.
return fwd_cfg["small_head"]
else:
return fwd_cfg["default"]
4 changes: 3 additions & 1 deletion aiter/ops/triton/attention/mha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -243,7 +243,9 @@ def _flash_attn_forward(
), f"dao_ai softmax_lse shape {softmax_lse.shape} != expected (batch={batch}, nheads={num_q_heads}, ...)"
else:
if config is None:
config = _get_config(enable_dropout, q.dtype, has_pe=pe_head_dim > 0)
config = _get_config(
enable_dropout, q.dtype, has_pe=pe_head_dim > 0, head_dim_v=v_head_dim
)

grid = lambda META: (
batch * num_q_heads * triton.cdiv(seqlen_q, META["BLOCK_M"]),
Expand Down
11 changes: 10 additions & 1 deletion aiter/ops/triton/configs/gfx950-MHA-DEFAULT.json
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,14 +18,23 @@
"num_ctas": 1,
"num_stages": 1
},
"small_head": {
"BLOCK_M": 128,
"BLOCK_N": 64,
"PRELOAD_V": true,
"waves_per_eu": 2,
"num_warps": 4,
"num_ctas": 1,
"num_stages": 3
},
"pe": {
"BLOCK_M": 128,
"BLOCK_N": 32,
"PRELOAD_V": true,
"waves_per_eu": 2,
"num_warps": 4,
"num_ctas": 1,
"num_stages": 1
"num_stages": 3
},
"pe_dropout_or_fp32": {
"BLOCK_M": 256,
Expand Down
Loading