Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Optimize nvfp4 lora applying. (#11856)

comfyanonymouseff2b9d41293
diff --git a/comfy/float.py b/comfy/float.py
index 1a6070bff..8c303bea0 100644
--- a/comfy/float.py
+++ b/comfy/float.py
@@ -161,10 +161,7 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
block_size = 16
x = x.reshape(orig_shape[0], -1, block_size)
- max_abs = torch.amax(torch.abs(x), dim=-1)
- block_scale = max_abs / F4_E2M1_MAX
- scaled_block_scales = block_scale / per_tensor_scale.to(block_scale.dtype)
- scaled_block_scales_fp8 = torch.clamp(scaled_block_scales, max=F8_E4M3_MAX).to(torch.float8_e4m3fn)
+ scaled_block_scales_fp8 = torch.clamp(((torch.amax(torch.abs(x), dim=-1)) / F4_E2M1_MAX) / per_tensor_scale.to(x.dtype), max=F8_E4M3_MAX).to(torch.float8_e4m3fn)
x /= (per_tensor_scale.to(x.dtype) * scaled_block_scales_fp8.to(x.dtype)).unsqueeze(-1)
generator = torch.Generator(device=x.device)
@@ -172,6 +169,5 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
x = x.view(orig_shape).nan_to_num()
data_lp = stochastic_float_to_fp4_e2m1(x, generator=generator)
-
blocked_scales = to_blocked(scaled_block_scales_fp8, flatten=False)
return data_lp, blocked_scales