Skip to content

fix(mlu): make Sage attention compile safe #1435

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

Open
Watebear wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(mlu): make Sage attention compile safe #1435

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
54 changes: 42 additions & 12 deletions lightx2v_platform/ops/attn/cambricon_mlu/sage_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,47 @@
tmo = None


@torch.library.custom_op(
"lightx2v::mlu_sage_attention",
mutates_args=(),
device_types="mlu",
)
def _mlu_sage_attention(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
max_seqlen_q: int,
max_seqlen_kv: int,
softmax_scale: float,
causal: bool,
) -> torch.Tensor:
return tmo.sage_attn(
q=q,
k=k,
v=v,
cu_seq_lens_q=None,
cu_seq_lens_kv=None,
max_seq_len_q=max_seqlen_q,
max_seq_len_kv=max_seqlen_kv,
softmax_scale=softmax_scale,
is_causal=causal,
compute_dtype=torch.bfloat16,
)


@_mlu_sage_attention.register_fake
def _mlu_sage_attention_fake(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
max_seqlen_q: int,
max_seqlen_kv: int,
softmax_scale: float,
causal: bool,
) -> torch.Tensor:
return v.new_empty((*q.shape[:-1], v.shape[-1]))


@PLATFORM_ATTN_WEIGHT_REGISTER("mlu_sage_attn")
class MluSageAttnWeight(AttnWeightTemplate):
def __init__(self):
Expand All @@ -27,17 +68,6 @@ def apply(self, q, k, v, cu_seqlens_q=None, cu_seqlens_kv=None, max_seqlen_q=Non
if softmax_scale is None:
softmax_scale = 1 / math.sqrt(q.shape[-1])
causal = kwds.get("causal", False)
x = tmo.sage_attn(
q=q,
k=k,
v=v,
cu_seq_lens_q=None,
cu_seq_lens_kv=None,
max_seq_len_kv=max_seqlen_kv,
max_seq_len_q=max_seqlen_q,
is_causal=causal,
compute_dtype=torch.bfloat16,
softmax_scale=softmax_scale,
)
x = _mlu_sage_attention(q, k, v, max_seqlen_q, max_seqlen_kv, softmax_scale, causal)
x = x.reshape(bs * max_seqlen_q, -1)
return x
Loading