Skip to content

add how_v3_bf16_cvt control to the Python API #1351

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 3 commits into from
Nov 7, 2025
Merged

add how_v3_bf16_cvt control to the Python API #1351

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
380740e
add how_v3_bf16_cvt in fwd_v3
minmengdie Nov 6, 2025
a870614
fix the fwd compile
minmengdie Nov 7, 2025
dc8cca9
Merge branch 'main' into mmd/dev/roundmode
minmengdie Nov 7, 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
11 changes: 9 additions & 2 deletions aiter/ops/mha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -202,6 +202,7 @@ def gen_fmha_v3_fwd_fake_tensors(
window_size_right: int,
return_softmax_lse: bool,
return_dropout_randval: bool,
how_v3_bf16_cvt: int,
out: Optional[Tensor] = None,
bias: Optional[Tensor] = None,
alibi_slopes: Optional[Tensor] = None,
Expand All @@ -226,6 +227,7 @@ def fmha_v3_fwd(
window_size_right: int,
return_softmax_lse: bool,
return_dropout_randval: bool,
how_v3_bf16_cvt: int,
out: Optional[Tensor] = None,
bias: Optional[Tensor] = None,
alibi_slopes: Optional[Tensor] = None,
Expand Down Expand Up @@ -1176,6 +1178,7 @@ def _flash_attn_forward(
alibi_slopes: Optional[torch.Tensor],
return_lse: bool,
return_softmax: bool,
how_v3_bf16_cvt: Optional[int] = 1,
cu_seqlens_q: Optional[torch.Tensor] = None,
cu_seqlens_kv: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
Expand Down Expand Up @@ -1229,6 +1232,7 @@ def _validate_cu(name: str, x: Optional[torch.Tensor]):
window_size_right,
return_lse,
return_softmax,
how_v3_bf16_cvt,
None,
bias,
alibi_slopes,
Expand Down Expand Up @@ -1626,6 +1630,7 @@ def forward(
alibi_slopes=alibi_slopes,
return_lse=return_lse,
return_softmax=return_softmax and dropout_p > 0,
how_v3_bf16_cvt=how_v3_bf16_cvt,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_kv=cu_seqlens_kv,
)
Expand Down Expand Up @@ -1742,6 +1747,7 @@ def flash_attn_func(
deterministic=True,
return_lse=False,
return_attn_probs=False,
how_v3_bf16_cvt=1,
cu_seqlens_q: Optional[torch.Tensor] = None,
cu_seqlens_kv: Optional[torch.Tensor] = None,
):
Expand Down Expand Up @@ -1811,7 +1817,7 @@ def flash_attn_func(
return_attn_probs,
torch.is_grad_enabled(),
True, # is_v3_atomic_fp32
1, # how_v3_bf16_cvt
how_v3_bf16_cvt,
cu_seqlens_q,
cu_seqlens_kv,
)
Expand Down Expand Up @@ -2352,6 +2358,7 @@ def flash_attn_varlen_func(
deterministic=False,
return_lse=False,
return_attn_probs=False,
how_v3_bf16_cvt=1,
block_table=None,
out=None,
cu_seqlens_q_padded: Optional[torch.Tensor] = None,
Expand Down Expand Up @@ -2445,7 +2452,7 @@ def flash_attn_varlen_func(
cu_seqlens_q_padded,
cu_seqlens_k_padded,
True,
1,
how_v3_bf16_cvt,
)


Expand Down
1 change: 1 addition & 0 deletions csrc/include/rocm_ops.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -614,6 +614,7 @@ namespace py = pybind11;
py::arg("window_size_right"), \
py::arg("return_softmax_lse"), \
py::arg("return_dropout_randval"), \
py::arg("how_v3_bf16_cvt"), \
py::arg("out") = std::nullopt, \
py::arg("bias") = std::nullopt, \
py::arg("alibi_slopes") = std::nullopt, \
Expand Down
1 change: 1 addition & 0 deletions csrc/include/torch/mha_v3_fwd.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@ std::vector<at::Tensor> fmha_v3_fwd(at::Tensor &q, // [b, sq, hq, d]
int window_size_right,
bool return_softmax_lse,
bool return_dropout_randval,
int how_v3_bf16_cvt,
std::optional<at::Tensor> out_, // [b, sq, hq, d_v]
std::optional<const at::Tensor> bias_, // [sq, sk]
std::optional<const at::Tensor> alibi_slopes_, // [hq] or [b, hq]
Expand Down
4 changes: 3 additions & 1 deletion csrc/py_itfs_cu/asm_mha_fwd.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,6 +149,7 @@ std::vector<at::Tensor> fmha_v3_fwd(at::Tensor &q, // [b, sq, hq, d]
int window_size_right,
bool return_softmax_lse,
bool return_dropout_randval,
int how_v3_bf16_cvt,
std::optional<at::Tensor> out_, // [b, sq, hq, d_v]
std::optional<const at::Tensor> bias_, // [sq, sk]
std::optional<const at::Tensor> alibi_slopes_, // [hq] or [b, hq]
Expand Down Expand Up @@ -316,7 +317,8 @@ std::vector<at::Tensor> fmha_v3_fwd(at::Tensor &q, // [b, sq, hq, d]
mask.type,
bias_type,
has_lse,
true);
true,
how_v3_bf16_cvt);
TORCH_CHECK(t >= 0, "invalid argument for fmha_fwd");
}
else {
Expand Down
9 changes: 5 additions & 4 deletions op_tests/test_mha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,10 +104,11 @@ def run_ck(
bias,
alibi_slopes,
deterministic,
return_lse,
return_attn_probs,
cu_seqlens_q,
cu_seqlens_kv,
return_lse=return_lse,
return_attn_probs=return_attn_probs,
how_v3_bf16_cvt=1,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_kv=cu_seqlens_kv,
num_rotate_args=1,
)

Expand Down