Skip to content

Update int8-int4 QAT config to use Int8DynamicActivationIntxWeightConfig #3391

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 1 commit into from
Oct 2, 2025
Merged

Update int8-int4 QAT config to use Int8DynamicActivationIntxWeightConfig #3391

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
5 changes: 3 additions & 2 deletions unsloth/models/_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1653,11 +1653,12 @@ def _prepare_model_for_qat(model: torch.nn.Module, qat_scheme: str) -> torch.nn.
from torchao.quantization import (
Float8DynamicActivationInt4WeightConfig,
Float8DynamicActivationFloat8WeightConfig,
Int8DynamicActivationInt4WeightConfig,
Int8DynamicActivationIntxWeightConfig,
Int4WeightOnlyConfig,
PerRow,
quantize_,
)
from torchao.quantization.granularity import PerGroup
from torchao.quantization.qat import QATConfig
filter_fn = None
if qat_scheme == "fp8-int4":
Expand All @@ -1668,7 +1669,7 @@ def _prepare_model_for_qat(model: torch.nn.Module, qat_scheme: str) -> torch.nn.
base_config = Float8DynamicActivationFloat8WeightConfig(granularity=PerRow())
elif qat_scheme == "int8-int4":
group_size = 32
base_config = Int8DynamicActivationInt4WeightConfig(group_size=group_size)
base_config = Int8DynamicActivationIntxWeightConfig(weight_dtype=torch.int4, weight_granularity=PerGroup(group_size))
filter_fn = lambda m, _: isinstance(m, torch.nn.Linear) and m.in_features >= group_size
elif qat_scheme == "int4":
group_size = 128
Expand Down