Skip to content

Fix small performance regression with fp8 fast and scaled fp8. #10537

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
comfyanonymous merged 1 commit into from
Oct 29, 2025
Merged

Fix small performance regression with fp8 fast and scaled fp8. #10537

Show file tree
Hide file tree
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
6 changes: 5 additions & 1 deletion comfy/ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -421,14 +421,18 @@ def fp8_linear(self, input):

if scale_input is None:
scale_input = torch.ones((), device=input.device, dtype=torch.float32)
input = torch.clamp(input, min=-448, max=448, out=input)
input = input.reshape(-1, input_shape[2]).to(dtype).contiguous()
layout_params_weight = {'scale': scale_input, 'orig_dtype': input_dtype}
quantized_input = QuantizedTensor(input.reshape(-1, input_shape[2]).to(dtype).contiguous(), TensorCoreFP8Layout, layout_params_weight)
else:
scale_input = scale_input.to(input.device)
quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype)

# Wrap weight in QuantizedTensor - this enables unified dispatch
# Call F.linear - __torch_dispatch__ routes to fp8_linear handler in quant_ops.py!
layout_params_weight = {'scale': scale_weight, 'orig_dtype': input_dtype}
quantized_weight = QuantizedTensor(w, TensorCoreFP8Layout, layout_params_weight)
quantized_input = QuantizedTensor.from_float(input.reshape(-1, input_shape[2]), TensorCoreFP8Layout, scale=scale_input, dtype=dtype)
o = torch.nn.functional.linear(quantized_input, quantized_weight, bias)

uncast_bias_weight(self, w, bias, offload_stream)
Expand Down
5 changes: 3 additions & 2 deletions comfy/quant_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -357,9 +357,10 @@ def quantize(cls, tensor, scale=None, dtype=torch.float8_e4m3fn):
scale = torch.tensor(scale)
scale = scale.to(device=tensor.device, dtype=torch.float32)

lp_amax = torch.finfo(dtype).max
tensor_scaled = tensor * (1.0 / scale).to(tensor.dtype)
torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
# TODO: uncomment this if it's actually needed because the clamp has a small performance penality'
# lp_amax = torch.finfo(dtype).max
# torch.clamp(tensor_scaled, min=-lp_amax, max=lp_amax, out=tensor_scaled)
qdata = tensor_scaled.to(dtype, memory_format=torch.contiguous_format)

layout_params = {
Expand Down