mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-09-12 12:37:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

SDPA backend priority (#9299)

Browse Source
This commit is contained in:
contentis
2025-08-13 20:53:27 +02:00
committed by GitHub
parent afa0a45206
commit 3da5a07510
4 changed files with 17 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
comfy/ops.py
View File

@@ -23,9 +23,18 @@ from comfy.cli_args import args, PerformanceFeature
import comfy.float
import comfy.rmsnorm
import contextlib
from torch.nn.attention import SDPBackend, sdpa_kernel
cast_to = comfy.model_management.cast_to #TODO: remove once no more references
SDPA_BACKEND_PRIORITY = [
SDPBackend.FLASH_ATTENTION,
SDPBackend.EFFICIENT_ATTENTION,
SDPBackend.MATH,
]
if torch.cuda.is_available():
SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
def cast_to_input(weight, input, non_blocking=False, copy=True):
return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
@@ -249,6 +258,10 @@ class disable_weight_init:
else:
raise ValueError(f"unsupported dimensions: {dims}")
@staticmethod
@sdpa_kernel(backends=SDPA_BACKEND_PRIORITY, set_priority=True)
def scaled_dot_product_attention(q, k, v, *args, **kwargs):
return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
class manual_cast(disable_weight_init):
class Linear(disable_weight_init.Linear):