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 (examples/sdxl): Updates to SDXL entry-point #1020

Merged
merged 14 commits into from
Sep 12, 2024
Merged

Feat (examples/sdxl): Updates to SDXL entry-point #1020

Changes from 1 commit
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
ea8cb36
Added conv_in/conv_out to blacklist
nickfraser Sep 5, 2024
fcfedfa
Switch SDPA quantization to FP8 FNUZ
nickfraser Sep 5, 2024
60a94d8
Fix (example/sdxl): Send device to fid computation backend
nickfraser Sep 5, 2024
223700a
Fix (example/sdxl): Allow export of FP8 linear/conv layers.
nickfraser Sep 6, 2024
bec3da1
Fix: formatting
nickfraser Sep 10, 2024
28371a5
feat: (example/sdxl): Added option to specify quantization blacklist …
nickfraser Sep 10, 2024
f7684c3
Feat (example/sdxl): Allow customization of SDPA quant via the comman…
nickfraser Sep 10, 2024
63948f1
precommit
nickfraser Sep 10, 2024
00b1e89
Fix (example/sdxl): Changes how quantization overriding works for conv.
nickfraser Sep 11, 2024
f5ed4b1
precommit
nickfraser Sep 11, 2024
f57905b
support 2 versions of diffusers
Giuseppe5 Sep 11, 2024
8431f59
small fixes
Giuseppe5 Sep 11, 2024
071527f
Fix (example/sdxl): Removed `--quantize-sdp` arg. Improved calibratio…
nickfraser Sep 11, 2024
8396948
precommit
nickfraser Sep 11, 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
17 changes: 12 additions & 5 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 @@ -395,7 +395,7 @@ def sdpa_zp_stats_type():
'weight_quant']
layer_map[torch.nn.Conv2d] = (layer_map[torch.nn.Conv2d][0], conv_qkwargs)

if args.quantize_sdp:
if args.sdpa_bit_width > 0:
# `args.weight_quant_granularity` must be compatible with `args.sdpa_quant_format`
sdpa_quantizers = generate_quantizers(
dtype=dtype,
Expand Down Expand Up @@ -484,9 +484,17 @@ def sdpa_zp_stats_type():
print(f"Checkpoint loaded!")
pipe = pipe.to(args.device)
elif not args.dry_run:
if (args.linear_input_bit_width > 0 or args.conv_input_bit_width > 0 or
args.sdpa_bit_width > 0 or
args.quantize_sdp) and args.input_scale_type == 'static':
# Model needs calibration if any of its activation quantizers are 'static'
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sweet

activation_bw = [
args.linear_input_bit_width,
args.conv_input_bit_width,
args.sdpa_bit_width,]
activation_st = [
args.input_scale_type,
args.input_scale_type,
args.sdpa_scale_type,]
needs_calibration = any(map(lambda b, st: (b > 0) and st == 'static', activation_bw, activation_st))
if needs_calibration:
print("Applying activation calibration")
with torch.no_grad(), calibration_mode(pipe.unet):
run_val_inference(
Expand Down Expand Up @@ -952,7 +960,6 @@ def sdpa_zp_stats_type():
'dry-run',
default=False,
help='Generate a quantized model without any calibration. Default: Disabled')
add_bool_arg(parser, 'quantize-sdp', default=False, help='Quantize SDP. Default: Disabled')
add_bool_arg(
parser,
'override-conv-quant-config',
Expand Down
Loading