compressed-tensors 0.12.3a20251007__tar.gz → 0.12.3a20251008__tar.gz

{compressed_tensors-0.12.3a20251007 → compressed_tensors-0.12.3a20251008}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251007
+Version: 0.12.3a20251008
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.3a20251007 → compressed_tensors-0.12.3a20251008}/src/compressed_tensors/quantization/quant_args.py

@@ -259,6 +259,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
         # extract user-passed values from dictionary
         strategy = model.strategy
         group_size = model.group_size
+        block_structure = model.block_structure
         actorder = model.actorder
         dynamic = model.dynamic
         observer = model.observer
@@ -277,7 +278,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                     "strategy='group' and group_size = -1 for 'channel'"
                 )
 
-        # validate strategy and group
+        # validate group strategy
         if strategy == QuantizationStrategy.GROUP:
             if group_size is None or group_size <= 0:
                 raise ValueError(
@@ -292,6 +293,14 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
         ):
             raise ValueError("group_size requires strategy to be set to 'group'")
 
+        # validate block strategy
+        has_block_strategy = strategy == QuantizationStrategy.BLOCK
+        has_block_structure = block_structure is not None
+        if has_block_strategy and not has_block_structure:
+            raise ValueError(f"Block strategy requires block structure\n{model}")
+        if has_block_structure and not has_block_strategy:
+            raise ValueError(f"Block structure requires block strategy\n{model}")
+
         # validate activation ordering and strategy
         if actorder is not None and strategy != QuantizationStrategy.GROUP:
             raise ValueError(

{compressed_tensors-0.12.3a20251007 → compressed_tensors-0.12.3a20251008}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.12.3.a20251007'
+__version__ = version = '0.12.3.a20251008'
 __version_tuple__ = version_tuple = (0, 12, 3)

{compressed_tensors-0.12.3a20251007 → compressed_tensors-0.12.3a20251008}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.12.3a20251007
+Version: 0.12.3a20251008
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.12.3a20251007 → compressed_tensors-0.12.3a20251008}/tests/test_quantization/test_utils/test_helpers.py

@@ -32,36 +32,42 @@ from compressed_tensors.quantization.utils import (
     [
         (
             False,
-            QuantizationStrategy.TENSOR,
+            "tensor",
             torch.Size(
                 [
                     1,
                 ]
             ),
         ),
-        (True, QuantizationStrategy.CHANNEL, torch.Size([1, 1])),
-        (True, QuantizationStrategy.GROUP, torch.Size([1, 1])),
+        (True, "channel", torch.Size([1, 1])),
+        (True, "group", torch.Size([1, 1])),
         (
             False,
-            QuantizationStrategy.BLOCK,
+            "block",
             torch.Size(
                 [
                     1,
                 ]
             ),
         ),
-        (True, QuantizationStrategy.TOKEN, torch.Size([1, 1])),
+        (True, "token", torch.Size([1, 1])),
     ],
 )
 def test_calculate_qparams(keepdims, strategy, exp_shape):
-    value = torch.randn(14, 5)
+    value = torch.empty(5, 6)
     min_val = torch.amin(value, dim=tuple(), keepdims=keepdims)
     max_val = torch.amax(value, dim=tuple(), keepdims=keepdims)
 
     if strategy == QuantizationStrategy.GROUP:
         args = QuantizationArgs(strategy=strategy, group_size=2)
+    elif strategy == QuantizationStrategy.BLOCK:
+        args = QuantizationArgs(strategy=strategy, block_structure=[1, 3])
     else:
-        args = QuantizationArgs(strategy=strategy)
+        args = QuantizationArgs(
+            strategy=strategy,
+            group_size=(2 if strategy == "group" else None),
+            block_structure=([1, 3] if strategy == "block" else None),
+        )
         scale, zp = calculate_qparams(min_val, max_val, args)
         assert scale.shape == exp_shape
         assert zp.shape == exp_shape