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

Feat (brevitas_examples/sdxl): load vae checkpoint #1200

Open
wants to merge 2 commits into
base: dev
Choose a base branch
from
Open

Feat (brevitas_examples/sdxl): load vae checkpoint #1200

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
810402d
Feat (brevitas_examples/sdxl): load vae checkpoint
Giuseppe5 Feb 21, 2025
663a079
Fix
Giuseppe5 Feb 26, 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
22 changes: 19 additions & 3 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 @@ -586,6 +586,7 @@ def sdpa_zp_stats_type():
layerwise=True,
blacklist_layers=blacklist if args.exclude_blacklist_act_eq else None,
add_mul_node=True):

for (inp_args, inp_kwargs) in vae_calibration:
input_args = tuple([
input_arg.cpu() if isinstance(input_arg, torch.Tensor) else input_arg
Expand All @@ -594,6 +595,8 @@ def sdpa_zp_stats_type():
k: (v.cpu() if isinstance(v, torch.Tensor) else v)
for (k, v) in input_kwargs.items()}
pipe.vae.decode(*inp_args, **inp_kwargs)
if args.dry_run or args.vae_load_checkpoint is not None:
break

quantizers = generate_quantizers(
dtype=dtype,
Expand Down Expand Up @@ -646,7 +649,15 @@ def sdpa_zp_stats_type():
k: (v.cuda() if isinstance(v, torch.Tensor) else v)
for (k, v) in vae_calibration[0][1].items()}
pipe.vae.decode(*input_args, **input_kwargs)
if needs_calibration:

if args.vae_load_checkpoint is not None:
with load_quant_model_mode(pipe.vae):
pipe = pipe.to('cpu')
print(f"Loading checkpoint: {args.vae_load_checkpoint}... ", end="")
pipe.vae.load_state_dict(torch.load(args.vae_load_checkpoint, map_location='cpu'))
print(f"Checkpoint loaded!")
pipe = pipe.to(args.device)
if needs_calibration and not (args.dry_run or args.vae_load_checkpoint is not None):
print("Applying activation calibration")
with torch.no_grad(), calibration_mode(pipe.vae):
for (inp_args, inp_kwargs) in vae_calibration:
Expand All @@ -658,7 +669,7 @@ def sdpa_zp_stats_type():
for (k, v) in input_kwargs.items()}
pipe.vae.decode(*inp_args, **inp_kwargs)

if args.vae_gptq:
if args.vae_gptq and not (args.dry_run or args.vae_load_checkpoint is not None):
print("Applying GPTQ")
with torch.no_grad(), gptq_mode(pipe.vae,
create_weight_orig=False,
Expand All @@ -673,7 +684,7 @@ def sdpa_zp_stats_type():
k: (v.cuda() if isinstance(v, torch.Tensor) else v)
for (k, v) in input_kwargs.items()}
pipe.vae.decode(*inp_args, **inp_kwargs)
if args.vae_bias_correction:
if args.vae_bias_correction and not (args.dry_run or args.vae_load_checkpoint is not None):
print("Applying Bias Correction")
with torch.no_grad(), bias_correction_mode(pipe.vae):
for inp_args, inp_kwargs in vae_calibration:
Expand Down Expand Up @@ -830,6 +841,11 @@ def sdpa_zp_stats_type():
type=str,
default=None,
help='Path to checkpoint to load. If provided, PTQ techniques are skipped.')
parser.add_argument(
'--vae-load-checkpoint',
type=str,
default=None,
help='Path to checkpoint vae to load. If provided, PTQ techniques are skipped.')
parser.add_argument(
'--path-to-latents',
type=str,
Expand Down