compressed-tensors 0.10.3a20250728__tar.gz → 0.10.3a20250731__tar.gz

{compressed_tensors-0.10.3a20250728/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250731}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250728
+Version: 0.10.3a20250731
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731}/src/compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py

@@ -61,6 +61,27 @@ class NVFP4PackedCompressor(BaseQuantizationCompressor):
             "weight_global_scale",
         )
 
+    def compression_param_info(
+        self,
+        weight_shape: torch.Size,
+        quantization_args: Optional[QuantizationArgs] = None,
+    ) -> Dict[str, Tuple[torch.Size, torch.dtype]]:
+        """
+        Creates a dictionary of expected shapes and dtypes for each compression
+            parameter used by the compressor
+
+        :param weight_shape: uncompressed weight shape
+        :param quantization_args: quantization parameters for the weight
+        :return: dictionary mapping compressed parameter names to shape and dtype
+        """
+        output = {
+            "weight_packed": (
+                torch.Size((weight_shape[0], weight_shape[1] // 2)),
+                torch.uint8,
+            ),
+        }
+        return output
+
     def compress_weight(
         self,
         weight: Tensor,

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731}/src/compressed_tensors/quantization/lifecycle/forward.py

@@ -257,13 +257,10 @@ def _process_quantization(
         QuantizationStrategy.GROUP,
         QuantizationStrategy.TENSOR_GROUP,
     ):
-        n_dims = x.shape
-        if len(n_dims) > 2:
-            x = x.squeeze(0)
 
         output_dtype = dtype if dtype is not None else x.dtype
         output = torch.zeros_like(x).to(output_dtype)
-        columns = output.shape[1]
+        columns = output.shape[-1]
 
         # TODO: make validation step for inputs
 
@@ -293,14 +290,12 @@ def _process_quantization(
             perm = torch.argsort(g_idx)
             x = safe_permute(x, perm, dim=1)
 
-        x = torch.reshape(
-            x,
-            (
-                x.shape[0],
-                ceil(x.shape[1] / group_size),
-                group_size,
-            ),
+        # Maintain all dimensions apart from the last dim, which is divided by the group_size
+        reshaped_dims = (
+            ceil(x.shape[-1] / group_size),
+            group_size,
         )
+        x = x.unflatten(-1, reshaped_dims)
 
         if do_quantize:
             output = _quantize(
@@ -323,19 +318,12 @@ def _process_quantization(
                 global_scale=global_scale,
             )
 
-        output = torch.reshape(
-            output,
-            (output.shape[0], output.shape[1] * output.shape[2]),
-        )
-
+        output = output.flatten(start_dim=-2)
         output = output.to(output_dtype)
 
         if not is_column_order:
             output = safe_permute(output, torch.argsort(perm), dim=1)
 
-        if len(n_dims) > 2:
-            output = output.unsqueeze(0)
-
     else:  # covers channel, token and tensor strategies
         if do_quantize:
             output = _quantize(

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731}/src/compressed_tensors/quantization/utils/helpers.py

@@ -175,20 +175,16 @@ def compute_dynamic_scales_and_zp(
         QuantizationStrategy.TENSOR_GROUP,
         QuantizationStrategy.GROUP,
     ):
-        if len(value.shape) > 2:
-            value = value.squeeze(0)
 
-        dim = {0, 1}
-        reduce_dims = tuple(idx for idx in range(3) if idx not in dim)
+        reduce_dims = -1
         keep_dims = False
-        value = torch.reshape(
-            value,
-            (
-                value.shape[0],
-                math.ceil(value.shape[1] / args.group_size),
-                args.group_size,
-            ),
+
+        reshaped_dims = (
+            math.ceil(value.shape[-1] / args.group_size),
+            args.group_size,
         )
+        value = value.unflatten(-1, reshaped_dims)
+
     else:
         supported_strategies = (
             QuantizationStrategy.TOKEN,

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250728'
+__version__ = version = '0.10.3.a20250731'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250728
+Version: 0.10.3a20250731
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250728 → compressed_tensors-0.10.3a20250731}/tests/test_quantization/test_utils/test_helpers.py

@@ -83,7 +83,7 @@ def test_fused_global_scales():
     "shape,group_size,exp_shape",
     [
         # Only batch size =1 is supported for dynamic GROUP quantization
-        ((1, 4, 8), 4, torch.Size([4, 2])),
+        ((1, 4, 8), 4, torch.Size([1, 4, 2])),
     ],
 )
 def test_compute_dynamic_scales_and_zp_group(shape, group_size, exp_shape):