Skip to content

Fix SM120 forward pass crash: parent __init__ overwrites arch, enabling unsupported TMA path #2416

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
moghon92 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix SM120 forward pass crash: parent __init__ overwrites arch, enabling unsupported TMA path #2416

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
10 changes: 10 additions & 0 deletions flash_attn/cute/flash_fwd_sm120.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@

import cutlass
import cutlass.utils as utils_basic
from cutlass.base_dsl.arch import Arch

from flash_attn.cute.flash_fwd import FlashAttentionForwardSm80

Expand All @@ -16,6 +17,15 @@ class FlashAttentionForwardSm120(FlashAttentionForwardSm80):
# The compilation target is determined by the GPU at compile time, not this field.
arch = 80

def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# FlashAttentionForwardSm80.__init__ sets self.arch from
# BaseDSL.get_arch_enum() which returns the real GPU arch (sm_120),
# overwriting the class-level arch = 80. This enables TMA code paths
# (use_tma_O) that SM120 doesn't support, since tma_atom_O is never
# created for this subclass. Reset to sm_80 to stay on CpAsync paths.
self.arch = Arch.sm_80

@staticmethod
def can_implement(
dtype,
Expand Down