Skip to content

Quantization: Add 2bit to precision constants, Fix -1 group_size handling in ModelBuilder #2338

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
jambayk merged 2 commits into from
Feb 18, 2026
+12 −13
Merged

Quantization: Add 2bit to precision constants, Fix -1 group_size handling in ModelBuilder #2338

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7f6c4b0
2bit precision, mb -1 blocksize
jambayk Feb 17, 2026
3f046d3
ut
jambayk Feb 17, 2026
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
1 change: 1 addition & 0 deletions olive/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,6 +55,7 @@ class Precision(StrEnumBase):


class PrecisionBits(IntEnum):
BITS2 = 2
BITS4 = 4
BITS8 = 8
BITS16 = 16
Expand Down
15 changes: 8 additions & 7 deletions olive/passes/onnx/model_builder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -420,13 +420,14 @@ def set_tensor(module, tensor_name, tensor_value, local_bits, local_group_size):
for q_attr, q_value in [("bits", local_bits), ("_group_size", local_group_size)]:
setattr(submodule, q_attr, q_value)
# in_features is always a multiple of group_size, group_size is a power of 2
if tensor_name.endswith("scales"):
submodule.out_features = tensor_value.shape[0]
submodule.in_features = tensor_value.shape[1] * local_group_size
elif tensor_name.endswith("qweight"):
tensor_value = tensor_value.reshape(
tensor_value.shape[0], (tensor_value.shape[1] * 8 // local_bits) // local_group_size, -1
)
# assumes no padding
if tensor_name.endswith("qweight"):
out_features, in_features_packed = tensor_value.shape
in_features = in_features_packed * 8 // local_bits
submodule.in_features = in_features
submodule.out_features = out_features
num_blocks = in_features // local_group_size if local_group_size != -1 else 1
tensor_value = tensor_value.reshape(out_features, num_blocks, -1)
setattr(submodule, tensor_name.split(".")[-1], tensor_value)

for weight_file in Path(input_path).iterdir():
Expand Down
9 changes: 3 additions & 6 deletions test/passes/onnx/test_model_builder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,18 +33,15 @@ def test_model_builder(tmp_path, metadata_only):
assert Path(output_folder / "genai_config.json").exists()


@pytest.mark.skip(
reason="Skip for now, need a fix in genai to support new Olive quant format "
"https://github.com/microsoft/onnxruntime-genai/pull/1916"
)
@pytest.mark.parametrize("embeds", [True, False])
def test_model_builder_olive_quant(tmp_path, embeds):
@pytest.mark.parametrize("group_size", [16, -1])
def test_model_builder_olive_quant(tmp_path, embeds, group_size):
# set up quantized model
input_model = create_pass_from_dict(
Rtn,
{
"bits": 4,
"group_size": 16,
"group_size": group_size,
"symmetric": False,
"lm_head": True,
"embeds": embeds,
Expand Down
Loading