Skip to content

[Issue 543] [Bug fix] Fix dynamic input quant for AWQ #726

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

Merged
meenchen merged 1 commit into from
Dec 30, 2025
Merged

[Issue 543] [Bug fix] Fix dynamic input quant for AWQ #726

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
20 changes: 12 additions & 8 deletions modelopt/torch/quantization/model_calib.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -751,14 +751,18 @@ def postprocess(module, name):
delattr(module.weight_quantizer, "_pre_quant_scale")
if hasattr(module.input_quantizer, "_pre_quant_scale"):
delattr(module.input_quantizer, "_pre_quant_scale")
if module.awq_lite.is_input_quantized and module.input_quantizer.amax is not None:
act_amax = module.input_quantizer.amax
# TODO: make this a buffer after we support only heterogeneous checkpointing for MCore
module.input_quantizer._amax_for_smoothing = act_amax.cpu()
module.input_quantizer.reset_amax()
module.input_quantizer.axis = None
module.input_quantizer.amax = act_amax.amax()
module.input_quantizer.enable()
if module.awq_lite.is_input_quantized:
if module.input_quantizer.amax is not None:
act_amax = module.input_quantizer.amax
# TODO: make this a buffer after we support only heterogeneous checkpointing for MCore
module.input_quantizer._amax_for_smoothing = act_amax.cpu()
module.input_quantizer.reset_amax()
module.input_quantizer.axis = None
module.input_quantizer.amax = act_amax.amax()
module.input_quantizer.enable()
# for dynamic quantization, there is no amax, so we just enable the quantizer
else:
module.input_quantizer.enable()

if module.awq_lite.is_enabled:
apply_pre_quant_scale_and_smooth(module, 1.0 / module.awq_lite.best_scale)
Expand Down