Skip to content

Workaround for nvidia issue where VAE uses 3x more memory on torch 2.9 #10373

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

Merged
merged 1 commit into from
Oct 16, 2025
Merged

Workaround for nvidia issue where VAE uses 3x more memory on torch 2.9 #10373

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
19 changes: 19 additions & 0 deletions comfy/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,16 @@ def scaled_dot_product_attention(q, k, v, *args, **kwargs):
except (ModuleNotFoundError, TypeError):
logging.warning("Could not set sdpa backend priority.")

NVIDIA_MEMORY_CONV_BUG_WORKAROUND = False
try:
if comfy.model_management.is_nvidia():
if torch.backends.cudnn.version() >= 91300 and comfy.model_management.torch_version_numeric >= (2, 9) and comfy.model_management.torch_version_numeric <= (2, 10):
#TODO: change upper bound version once it's fixed'
NVIDIA_MEMORY_CONV_BUG_WORKAROUND = True
logging.info("working around nvidia conv3d memory bug.")
except:
pass

cast_to = comfy.model_management.cast_to #TODO: remove once no more references

if torch.cuda.is_available() and torch.backends.cudnn.is_available() and PerformanceFeature.AutoTune in args.fast:
Expand Down Expand Up @@ -151,6 +161,15 @@ class Conv3d(torch.nn.Conv3d, CastWeightBiasOp):
def reset_parameters(self):
return None

def _conv_forward(self, input, weight, bias, *args, **kwargs):
if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16):
out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True)
if bias is not None:
out += bias.reshape((1, -1) + (1,) * (out.ndim - 2))
return out
else:
return super()._conv_forward(input, weight, bias, *args, **kwargs)

def forward_comfy_cast_weights(self, input):
weight, bias = cast_bias_weight(self, input)
return self._conv_forward(input, weight, bias)
Expand Down