Skip to content
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.

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

Bump flash-attn to v2.0.4 #816

Merged
merged 4 commits into from
Aug 11, 2023
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
4 changes: 2 additions & 2 deletions setup.py
Original file line number Diff line number Diff line change
Expand Up @@ -143,13 +143,13 @@ def get_flash_attention_extensions(cuda_version: int, extra_compile_args):
nvcc_windows_flags = ["-Xcompiler", "/permissive-"]

flash_root = os.path.join(this_dir, "third_party", "flash-attention")
if not os.path.exists(flash_root):
cutlass_inc = os.path.join(flash_root, "csrc", "cutlass", "include")
if not os.path.exists(flash_root) or not os.path.exists(cutlass_inc):
raise RuntimeError(
"flashattention submodule not found. Did you forget "
"to run `git submodule update --init --recursive` ?"
)

flash_root = os.path.join(this_dir, "third_party", "flash-attention")
sources = ["csrc/flash_attn/flash_api.cpp"]
for f in glob.glob(os.path.join(flash_root, "csrc", "flash_attn", "src", "*.cu")):
sources.append(str(Path(f).relative_to(flash_root)))
Expand Down
16 changes: 3 additions & 13 deletions xformers/ops/fmha/flash.py
Original file line number Diff line number Diff line change
Expand Up @@ -67,6 +67,7 @@ def _flash_fwd(
out_padded,
softmax_lse,
p,
rng_state,
) = _C_flashattention.varlen_fwd(
query,
key,
Expand All @@ -83,7 +84,7 @@ def _flash_fwd(
return_softmax,
None,
)
return out, softmax_lse, None
return out, softmax_lse, rng_state

def _flash_bwd(
grad,
Expand Down Expand Up @@ -123,6 +124,7 @@ def _flash_bwd(
False, # zero_tensors
causal,
None,
rng_state,
)
return dq

Expand Down Expand Up @@ -302,18 +304,6 @@ class BwOp(AttentionBwOpBase):

MAX_HEADDIM_SM8x = 192

@classmethod
def shape_not_supported_reasons(
cls, Mq: int, Mkv: int, K: int, Kv: int
) -> List[str]:
reasons = super().shape_not_supported_reasons(Mq, Mkv, K, Kv)
if (Mq % 128) or (Mkv % 128):
reasons.append(
"flashv2 beta: BW is incorrect when seqlen is not aligned on 128 "
"(https://github.com/Dao-AILab/flash-attention/issues/334)"
)
return reasons

@classmethod
def not_supported_reasons(cls, d: Inputs) -> List[str]:
reasons = super(BwOp, cls).not_supported_reasons(d)
Expand Down