Skip to content

add weight_scale_inv implementation #10930

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

Closed
yoinked-h wants to merge 3 commits into from
Closed

add weight_scale_inv implementation #10930

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
701e499
add weight_scale_inv impl
yoinked-h Nov 27, 2025
a4c2b6f
divide when dividing
yoinked-h Nov 27, 2025
af7becd
lint
yoinked-h Nov 27, 2025
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
28 changes: 26 additions & 2 deletions comfy/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -464,6 +464,22 @@ def forward_comfy_cast_weights(self, input):
uncast_bias_weight(self, weight, bias, offload_stream)
return x


def scale_hadamard(larger, smaller, k_value=None, divide=False):
if smaller.shape == torch.Size([]):
# unset
return larger
h, w = smaller.shape
if k_value is None:
#calculate from larger compared to smaller
k_value = larger.shape[-1] // smaller.shape[-1]
expected_shape = (h * k_value, w * k_value)
assert larger.shape == expected_shape, "weight_scale_inv mismatch, skipping"
if divide:
smaller = 1.0 / smaller
result = larger.view(h, k_value, w, k_value) * smaller.view(h, 1, w, 1)
return result.reshape(h * k_value, w * k_value)

def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
class scaled_fp8_op(manual_cast):
Expand All @@ -477,6 +493,9 @@ def reset_parameters(self):
if not hasattr(self, 'scale_weight'):
self.scale_weight = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)

if not hasattr(self, 'weight_scale_inv'):
self.weight_scale_inv = torch.nn.parameter.Parameter(data=torch.ones((), device=self.weight.device, dtype=torch.float32), requires_grad=False)

if not scale_input:
self.scale_input = None

Expand All @@ -502,12 +521,17 @@ def forward_comfy_cast_weights(self, input):
def convert_weight(self, weight, inplace=False, **kwargs):
if inplace:
weight *= self.scale_weight.to(device=weight.device, dtype=weight.dtype)
if self.weight_scale_inv.shape == torch.Size([]):
return weight
weight = scale_hadamard(weight, self.weight_scale_inv.to(device=weight.device, dtype=weight.dtype), k_value=None, divide=False)
return weight
else:
return weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32)
return scale_hadamard(weight.to(dtype=torch.float32) * self.scale_weight.to(device=weight.device, dtype=torch.float32),
self.weight_scale_inv.to(device=weight.device, dtype=torch.float32), k_value=None, divide=False)

def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
weight = comfy.float.stochastic_rounding(scale_hadamard(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype),
self.weight_scale_inv.to(device=weight.device, dtype=weight.dtype), k_value=None, divide=True), self.weight.dtype, seed=seed)
if return_weight:
return weight
if inplace_update:
Expand Down