Skip to content

Fix pad tokens passed to quantization capture #1968

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
merged 3 commits into from
Oct 2, 2025
Merged

Fix pad tokens passed to quantization capture #1968

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
383d76f
enable deterministic ci tests
Qubitium Oct 2, 2025
64532d4
always capture attention masks or we process pad tokens for lower but…
Qubitium Oct 2, 2025
28e796b
diable slower deterministic tests
Qubitium Oct 2, 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
5 changes: 4 additions & 1 deletion gptqmodel/looper/module_looper.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -493,7 +493,8 @@ def store_input_hook(module, args, kwargs):
layer_inputs.append(layer_input)

# Keyword arguments.
if kwargs.get("attention_mask") is not None and self.gptq_model.ATTENTION_MASKS_REQUIRED_FOR_INPUT:
# Always capture attention_mask so downstream masking can drop padded tokens
if kwargs.get("attention_mask") is not None:
attention_masks.append(kwargs["attention_mask"].to(device=data_device))
else:
attention_masks.append(None)
Expand Down Expand Up @@ -632,6 +633,8 @@ def loop(self, fail_safe: bool = False, **kwargs):

layer_modules = self.gptq_model.simple_layer_modules(model_config=self.gptq_model.model.config, quantize_config=self.gptq_model.quantize_config)

# true-sequential will replay the quantized activations after each subset has been quantized to be used for next subset quantization
# this should always be true for gptq unless you want lower but misleading error_loss that is misleading and will lead to lower post-quantized model
if not self.gptq_model.quantize_config.true_sequential:
layer_modules = [sum(layer_modules, [])]

Expand Down
7 changes: 7 additions & 0 deletions tests/models/model_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,13 @@
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

# Following makes test results more deterministic but much slower
# # the CUBLAS env is required for use_deterministic_algorithms
# os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
#
# import torch
# torch.use_deterministic_algorithms(True)

# -- end do not touch

from pathlib import Path # noqa: E402
Expand Down