mirrors/ComfyUI
1
0
Fork 0
mirror of https://github.com/comfyanonymous/ComfyUI.git synced 2025-08-16 01:44:22 +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
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 17 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
comfy/ldm/hunyuan3d/vae.py
View File

@ -178,7 +178,7 @@ class FourierEmbedder(nn.Module):
class CrossAttentionProcessor:
def __call__(self, attn, q, k, v):
out = F.scaled_dot_product_attention(q, k, v)
out = ops.scaled_dot_product_attention(q, k, v)
return out

4
comfy/ldm/modules/attention.py
View File

@ -448,7 +448,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
mask = mask.unsqueeze(1)
if SDP_BATCH_LIMIT >= b:
out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
out = ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
if not skip_output_reshape:
out = (
out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@ -461,7 +461,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
if mask.shape[0] > 1:
m = mask[i : i + SDP_BATCH_LIMIT]
out[i : i + SDP_BATCH_LIMIT] = torch.nn.functional.scaled_dot_product_attention(
out[i : i + SDP_BATCH_LIMIT] = ops.scaled_dot_product_attention(
q[i : i + SDP_BATCH_LIMIT],
k[i : i + SDP_BATCH_LIMIT],
v[i : i + SDP_BATCH_LIMIT],

2
comfy/ldm/modules/diffusionmodules/model.py
View File

@ -285,7 +285,7 @@ def pytorch_attention(q, k, v):
)
try:
out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
out = ops.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
out = out.transpose(2, 3).reshape(orig_shape)
except model_management.OOM_EXCEPTION:
logging.warning("scaled_dot_product_attention OOMed: switched to slice attention")

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):