github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

DFE optimizations

Browse Source
This commit is contained in:
Jaret Burkett
2026-08-01 10:06:16 -06:00
parent fc78b07332
commit 5baa495585
1 changed files with 20 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
toolkit/models/diffusion_feature_extraction.py
View File

@@ -690,6 +690,11 @@ class DiffusionFeatureExtractor6(nn.Module):
self.log_every = 100 self.log_every = 100
self.step = 0 self.step = 0
# cache normalization constants once so prepare_inputs doesn't rebuild
# them (CPU tensor construction + H2D copy) on every call
self.image_mean = torch.tensor(self.processor.image_mean, device=device, dtype=dtype).view(1, 3, 1, 1)
self.image_std = torch.tensor(self.processor.image_std, device=device, dtype=dtype).view(1, 3, 1, 1)
def prepare_inputs(self, tensor_0_1: torch.Tensor): def prepare_inputs(self, tensor_0_1: torch.Tensor):
""" """
tensor_0_1: (bs, 3, h, w), float, values in [0, 1] tensor_0_1: (bs, 3, h, w), float, values in [0, 1]
@@ -703,8 +708,10 @@ class DiffusionFeatureExtractor6(nn.Module):
if not torch.is_floating_point(x): if not torch.is_floating_point(x):
x = x.float() x = x.float()
# VAE decode can overshoot [0, 1] slightly; clamp so the rescale # VAE decode can overshoot [0, 1] slightly; clamp back into range.
# heuristic below never mistakes an overshoot for 0..255 input # Inputs are documented as 0..1 and clamped here, so no 0..255 rescale
# heuristic is needed (the old max().item() check forced a GPU sync
# every call and could never trigger after this clamp anyway).
x = torch.clamp(x, 0.0, 1.0) x = torch.clamp(x, 0.0, 1.0)
# Resize # Resize
@@ -722,17 +729,10 @@ class DiffusionFeatureExtractor6(nn.Module):
target_w = (target_w // p) * p target_w = (target_w // p) * p
x = F.interpolate(x, size=(target_h, target_w), mode="bilinear", align_corners=False) x = F.interpolate(x, size=(target_h, target_w), mode="bilinear", align_corners=False)
# Rescale (HF processors usually assume uint8 0..255 inputs; your inputs are already 0..1)
if self.processor.do_rescale:
# If it looks like [0..1], skip to avoid double-scaling.
# If user accidentally passed 0..255 floats, this will fix it.
if x.detach().max().item() > 1.0 + 1e-6:
x = x * float(self.processor.rescale_factor or 1.0 / 255.0)
# Normalize # Normalize
if self.processor.do_normalize: if self.processor.do_normalize:
mean = torch.tensor(self.processor.image_mean, device=x.device, dtype=x.dtype).view(1, 3, 1, 1) mean = self.image_mean.to(device=x.device, dtype=x.dtype)
std = torch.tensor(self.processor.image_std, device=x.device, dtype=x.dtype).view(1, 3, 1, 1) std = self.image_std.to(device=x.device, dtype=x.dtype)
x = (x - mean) / std x = (x - mean) / std
return {"pixel_values": x} return {"pixel_values": x}
@@ -814,18 +814,22 @@ class DiffusionFeatureExtractor6(nn.Module):
pred_dino_output.float(), target_dino_output.float() pred_dino_output.float(), target_dino_output.float()
) )
# accumulate on-GPU; .item() every step forces a full pipeline sync,
# so only sync when we actually log
if 'dinov3' not in self.losses: if 'dinov3' not in self.losses:
self.losses['dinov3'] = dino_loss.item() self.losses['dinov3'] = dino_loss.detach()
else: else:
self.losses['dinov3'] += dino_loss.item() self.losses['dinov3'] = self.losses['dinov3'] + dino_loss.detach()
with torch.no_grad(): with torch.no_grad():
if self.step % self.log_every == 0 and self.step > 0: if self.step % self.log_every == 0 and self.step > 0:
print(f"DFE losses:") print(f"DFE losses:")
for key in self.losses: for key in self.losses:
self.losses[key] /= self.log_every avg = self.losses[key] / self.log_every
if torch.is_tensor(avg):
avg = avg.item()
# print in 2.000e-01 format # print in 2.000e-01 format
print(f" - {key}: {self.losses[key]:.3e}") print(f" - {key}: {avg:.3e}")
self.losses[key] = 0.0 self.losses[key] = 0.0
# total_loss += mse_loss # total_loss += mse_loss