[Bugfix] Fix shape calculation for group quantization (#308)

* use ceil of group size

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

* use math, not torch

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

* fix false assumption in tests

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

---------

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>

Author: kylesayrs

src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -14,6 +14,7 @@
 
 
 import logging
+import math
 from enum import Enum
 from typing import Optional
 
@@ -162,7 +163,7 @@ def _initialize_scale_zero_point(
             # (output_channels, 1)
             expected_shape = (weight_shape[0], 1)
         elif quantization_args.strategy == QuantizationStrategy.GROUP:
-            num_groups = weight_shape[1] // quantization_args.group_size
+            num_groups = math.ceil(weight_shape[1] / quantization_args.group_size)
             expected_shape = (weight_shape[0], max(num_groups, 1))
 
     scale_dtype = scale_dtype if scale_dtype is not None else module.weight.dtype

tests/test_quantization/lifecycle/test_initialize.py

@@ -13,6 +13,8 @@
 # limitations under the License.
 
 
+import math
+
 import pytest
 from compressed_tensors.quantization import (
     ActivationOrdering,
@@ -183,7 +185,7 @@ def test_initialize_quantization_parameters(weights, input_activations):
             expected_shape = (layer.weight.shape[0], 1)
 
         elif args.strategy == QuantizationStrategy.GROUP:  # only weight
-            num_groups = layer.weight.shape[1] // args.group_size
+            num_groups = math.ceil(layer.weight.shape[1] / args.group_size)
             expected_shape = (layer.weight.shape[0], max(num_groups, 1))
 
         elif args.strategy == QuantizationStrategy.BLOCK: