Skip to content
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

Fix (sdxl): avoid suppressing checkpoint errors #1034

Merged
merged 2 commits into from
Oct 1, 2024
Merged

Fix (sdxl): avoid suppressing checkpoint errors #1034

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ba248a2
Fix (calibrate): avoid suppressing checkpoint errors
Giuseppe5 Oct 1, 2024
a8a609c
Feat (examples/sdxl): move and consolidate export
Giuseppe5 Sep 27, 2024
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
1 change: 0 additions & 1 deletion src/brevitas/graph/calibrate.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,7 +149,6 @@ def __exit__(self, *args, **kwargs):
for module in self.model.modules():
if issubclass(type(module), QuantWBIOL):
module._quant_load_model_mode = False
return True


class ClipFloatWeights(Transform):
Expand Down
87 changes: 43 additions & 44 deletions src/brevitas_examples/stable_diffusion/main.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -463,18 +463,17 @@ def sdpa_zp_stats_type():

pipe.set_progress_bar_config(disable=True)

if args.dry_run:
with torch.no_grad():
run_val_inference(
pipe,
args.resolution, [calibration_prompts[0]],
test_seeds,
args.device,
dtype,
total_steps=1,
use_negative_prompts=args.use_negative_prompts,
test_latents=latents,
guidance_scale=args.guidance_scale)
with torch.no_grad():
run_val_inference(
pipe,
args.resolution, [calibration_prompts[0]],
test_seeds,
args.device,
dtype,
total_steps=1,
use_negative_prompts=args.use_negative_prompts,
test_latents=latents,
guidance_scale=args.guidance_scale)

if args.load_checkpoint is not None:
with load_quant_model_mode(pipe.unet):
Expand Down Expand Up @@ -574,6 +573,38 @@ def sdpa_zp_stats_type():
torch.save(
pipe.vae.state_dict(), os.path.join(output_dir, f"vae_{args.checkpoint_name}"))

if args.export_target:
# Move to cpu and to float32 to enable CPU export
if args.export_cpu_float32:
pipe.unet.to('cpu').to(torch.float32)
pipe.unet.eval()
device = next(iter(pipe.unet.parameters())).device
dtype = next(iter(pipe.unet.parameters())).dtype

# Define tracing input
if is_sd_xl:
generate_fn = generate_unet_xl_rand_inputs
shape = SD_XL_EMBEDDINGS_SHAPE
else:
generate_fn = generate_unet_21_rand_inputs
shape = SD_2_1_EMBEDDINGS_SHAPE
trace_inputs = generate_fn(
embedding_shape=shape,
unet_input_shape=unet_input_shape(args.resolution),
device=device,
dtype=dtype)

if args.export_target == 'onnx':
if args.weight_quant_granularity == 'per_group':
export_manager = BlockQuantProxyLevelManager
else:
export_manager = StdQCDQONNXManager
export_manager.change_weight_export(export_weight_q_node=args.export_weight_q_node)
export_onnx(pipe, trace_inputs, output_dir, export_manager)
if args.export_target == 'params_only':
pipe.to('cpu')
export_quant_params(pipe, output_dir, export_vae=args.vae_fp16_fix)

# Perform inference
if args.prompt > 0 and not args.dry_run:
# with brevitas_proxy_inference_mode(pipe.unet):
Expand Down Expand Up @@ -619,38 +650,6 @@ def sdpa_zp_stats_type():
fid.update(quant_images_values, real=False)
print(f"FID: {float(fid.compute())}")

if args.export_target:
# Move to cpu and to float32 to enable CPU export
if args.export_cpu_float32:
pipe.unet.to('cpu').to(torch.float32)
pipe.unet.eval()
device = next(iter(pipe.unet.parameters())).device
dtype = next(iter(pipe.unet.parameters())).dtype

# Define tracing input
if is_sd_xl:
generate_fn = generate_unet_xl_rand_inputs
shape = SD_XL_EMBEDDINGS_SHAPE
else:
generate_fn = generate_unet_21_rand_inputs
shape = SD_2_1_EMBEDDINGS_SHAPE
trace_inputs = generate_fn(
embedding_shape=shape,
unet_input_shape=unet_input_shape(args.resolution),
device=device,
dtype=dtype)

if args.export_target == 'onnx':
if args.weight_quant_granularity == 'per_group':
export_manager = BlockQuantProxyLevelManager
else:
export_manager = StdQCDQONNXManager
export_manager.change_weight_export(export_weight_q_node=args.export_weight_q_node)
export_onnx(pipe, trace_inputs, output_dir, export_manager)
if args.export_target == 'params_only':
pipe.to('cpu')
export_quant_params(pipe, output_dir, export_vae=args.vae_fp16_fix)


if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Stable Diffusion quantization')
Expand Down
12 changes: 6 additions & 6 deletions src/brevitas_examples/stable_diffusion/sd_quant/export.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,9 +36,9 @@ def handle_quant_param(layer, layer_dict):
output_scale = layer.output_quant.export_handler.symbolic_kwargs[
'dequantize_symbolic_kwargs']['scale'].data

layer_dict['output_scale'] = output_scale.numpy().tolist()
layer_dict['output_scale'] = output_scale.cpu().numpy().tolist()
layer_dict['output_scale_shape'] = output_scale.shape
layer_dict['input_scale'] = input_scale.numpy().tolist()
layer_dict['input_scale'] = input_scale.cpu().numpy().tolist()
layer_dict['input_scale_shape'] = input_scale.shape
layer_dict['input_zp'] = input_zp.to(torch.float32).cpu().numpy().tolist()
layer_dict['input_zp_shape'] = input_zp.shape
Expand Down Expand Up @@ -83,7 +83,7 @@ def export_quant_params(pipe, output_dir, export_vae=False):
full_name = name
smoothquant_param = module.scale.weight

layer_dict['smoothquant_mul'] = smoothquant_param.data.numpy().tolist()
layer_dict['smoothquant_mul'] = smoothquant_param.data.cpu().numpy().tolist()
layer_dict['smoothquant_mul_shape'] = module.scale.runtime_shape
layer_dict = handle_quant_param(module.layer, layer_dict)

Expand All @@ -94,7 +94,7 @@ def export_quant_params(pipe, output_dir, export_vae=False):
full_name = name
smoothquant_param = module.scale.weight

layer_dict['smoothquant_mul'] = smoothquant_param.data.numpy().tolist()
layer_dict['smoothquant_mul'] = smoothquant_param.data.cpu().numpy().tolist()
layer_dict['smoothquant_mul_shape'] = module.scale.runtime_shape
quant_params[full_name] = layer_dict
handled_quant_layers.add(id(module.layer))
Expand All @@ -113,9 +113,9 @@ def export_quant_params(pipe, output_dir, export_vae=False):
'dequantize_symbolic_kwargs']['scale'].data
act_zp = module.act_quant.export_handler.symbolic_kwargs[
'dequantize_symbolic_kwargs']['zero_point'].data
layer_dict['act_scale'] = act_scale.numpy().tolist()
layer_dict['act_scale'] = act_scale.cpu().numpy().tolist()
layer_dict['act_scale_shape'] = act_scale.shape
layer_dict['act_zp'] = act_zp.to(torch.float32).numpy().tolist()
layer_dict['act_zp'] = act_zp.to(torch.float32).cpu().numpy().tolist()
layer_dict['act_zp_shape'] = act_zp.shape
layer_dict['act_zp_dtype'] = str(act_zp.dtype)
quant_params[full_name] = layer_dict
Expand Down
Loading