Skip to content

fix fwd perf calc error #1305

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
minmengdie merged 2 commits into from
Oct 31, 2025
Merged

fix fwd perf calc error #1305

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
a73457a
fix fwd perf calc error
minmengdie Oct 31, 2025
872cc55
black aiter/ops/triton/_triton_kernels/gemm_afp4wfp4.py
minmengdie Oct 31, 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
8 changes: 6 additions & 2 deletions aiter/ops/triton/_triton_kernels/gemm_afp4wfp4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,7 +141,9 @@ def _gemm_afp4_wfp4_kernel(
cache_modifier=cache_modifier,
)

accumulator = tl.dot_scaled(a, a_scales, "e2m1", b, b_scales, "e2m1", accumulator)
accumulator = tl.dot_scaled(
a, a_scales, "e2m1", b, b_scales, "e2m1", accumulator
)

# Advance the ptrs to the next K block.
a_ptrs += (BLOCK_SIZE_K // 2) * stride_ak
Expand Down Expand Up @@ -340,7 +342,9 @@ def _gemm_afp4_wfp4_kernel_preshuffled_scales(
b_ptrs, mask=offs_k[:, None] < K - k * (BLOCK_SIZE_K // 2), other=0
)

accumulator = tl.dot_scaled(a, a_scales, "e2m1", b, b_scales, "e2m1", accumulator)
accumulator = tl.dot_scaled(
a, a_scales, "e2m1", b, b_scales, "e2m1", accumulator
)

# Advance the ptrs to the next K block.
a_ptrs += (BLOCK_SIZE_K // 2) * stride_ak
Expand Down
18 changes: 13 additions & 5 deletions op_tests/test_mha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -365,18 +365,26 @@ def test_flash_attn_output(
dbias_tol = max(10 * (dbias_pt - dbias_ref).abs().max().item(), 0.01)
assert (dbias - dbias_ref).abs().max().item() <= dbias_tol

fwd_flop = nheads * (seqlen_q * seqlen_k * d * 2 + seqlen_q * seqlen_k * d_v * 2)
fwd_flop = (
batch_size
* nheads
* (seqlen_q * seqlen_k * d * 2 + seqlen_q * seqlen_k * d_v * 2)
)
dtype_bytes = torch.finfo(dtype).bits // 8
fwd_num_bytes = (
nheads
batch_size
* nheads
* dtype_bytes
* (seqlen_q * d + seqlen_k * d + seqlen_k * d_v + seqlen_q * d_v)
)
bwd_flop = nheads * (
seqlen_q * seqlen_k * d * 2 * 3 + seqlen_q * seqlen_k * d_v * 2 * 2
bwd_flop = (
batch_size
* nheads
* (seqlen_q * seqlen_k * d * 2 * 3 + seqlen_q * seqlen_k * d_v * 2 * 2)
)
bwd_num_bytes = (
2 * fwd_num_bytes + nheads * (torch.finfo(torch.float).bits // 8) * seqlen_q
2 * fwd_num_bytes
+ batch_size * nheads * (torch.finfo(torch.float).bits // 8) * seqlen_q
)
ret = {}
ret["fwd_us"] = us_fwd
Expand Down