compressed-tensors 0.10.3a20250724__py3-none-any.whl → 0.10.3a20250731__py3-none-any.whl

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/quantization/lifecycle/forward.py

@@ -112,17 +112,21 @@ def dequantize(
             if scale.shape[1] == 1:
                 args = QuantizationArgs(strategy=QuantizationStrategy.CHANNEL)
             # Scale height matches input or is 1 -> group quantization across columns
-            # 
+            #
             # Example 1: scale.shape[0] == 1
             # x_q: (4, 8), scale: (1, 4) -> 2 columns per group
             #
-            # Example 2: scale.shape[0] == x_q.shape[0] 
+            # Example 2: scale.shape[0] == x_q.shape[0]
             # x_q: (4, 8), scale: (4, 4) -> 2 elements per group (per row)
             elif (scale.shape[0] == 1) or (scale.shape[0] == x_q.shape[0]):
                 group_size = int(x_q.shape[1] / scale.shape[1])
-                args = QuantizationArgs(strategy=QuantizationStrategy.GROUP, group_size=group_size)
+                args = QuantizationArgs(
+                    strategy=QuantizationStrategy.GROUP, group_size=group_size
+                )
             else:
-                args = QuantizationArgs(strategy=QuantizationStrategy.BLOCK, block_structure=scale.shape)
+                args = QuantizationArgs(
+                    strategy=QuantizationStrategy.BLOCK, block_structure=scale.shape
+                )
         else:
             raise ValueError(
                 f"Could not infer a quantization strategy from scale with {scale.ndim} "
@@ -253,13 +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
 
@@ -289,14 +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(
@@ -319,19 +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/quantization/lifecycle/initialize.py

@@ -185,27 +185,29 @@ def _initialize_scale_zero_point(
         elif quantization_args.strategy == QuantizationStrategy.BLOCK:
             # For block quantization, scale shape should match number of blocks - only for weights
             if quantization_args.block_structure is None:
-                raise ValueError("Block quantization requires block_structure to be specified")
+                raise ValueError(
+                    "Block quantization requires block_structure to be specified"
+                )
             block_height, block_width = quantization_args.block_structure
             rows, cols = weight_shape[-2], weight_shape[-1]
             num_rows_blocks = math.ceil(rows / block_height)
             num_cols_blocks = math.ceil(cols / block_width)
-            
+
             # Warn if dimensions don't divide evenly
             if rows % block_height != 0 or cols % block_width != 0:
                 warnings.warn(
                     f"Block quantization: tensor shape {weight_shape} does not divide evenly "
                     f"by block structure {quantization_args.block_structure}. "
                     f"Some blocks will be incomplete which may affect quantization quality.",
-                    UserWarning
+                    UserWarning,
                 )
-            
+
             expected_shape = (num_rows_blocks, num_cols_blocks)
     elif quantization_args.strategy == QuantizationStrategy.BLOCK:
         warnings.warn(
             f"BLOCK quantization not supported for {base_name} activations. "
             f"Falling back to tensor-level quantization.",
-            UserWarning
+            UserWarning,
         )
         expected_shape = 1
 

compressed_tensors/quantization/quant_scheme.py

@@ -64,8 +64,9 @@ class QuantizationScheme(BaseModel):
                 raise ValueError("Cannot apply actorder to output activations")
 
         if (
-            inputs and weights
-            and weights.strategy == QuantizationStrategy.GROUP 
+            inputs
+            and weights
+            and weights.strategy == QuantizationStrategy.GROUP
             and inputs.strategy == QuantizationStrategy.GROUP
             and weights.group_size != inputs.group_size
         ):
@@ -75,7 +76,7 @@ class QuantizationScheme(BaseModel):
                 "may complicate fused kernel implementations. Consider using "
                 "TENSOR_GROUP strategy for both or matching group sizes.",
                 UserWarning,
-                stacklevel=2
+                stacklevel=2,
             )
 
         return model

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/transform/factory/hadamard.py

@@ -12,9 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import math
 from typing import Optional, Union
 
-import math
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
 from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
@@ -103,7 +103,8 @@ class HadamardTransform(TransformBase):
 
         if self.args.inverse:
             weight = weight.T
- 
-        return apply_transform_weight(
-            weight, value, self.args.location, self.module_type
-        ) / self._scale
+
+        return (
+            apply_transform_weight(weight, value, self.args.location, self.module_type)
+            / self._scale
+        )

compressed_tensors/utils/match.py

@@ -18,6 +18,7 @@ from collections.abc import Generator
 from typing import Iterable, Tuple
 
 import torch
+from compressed_tensors.utils.internal import InternalModule
 
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)
@@ -28,8 +29,6 @@ __all__ = [
     "match_named_parameters",
     "match_modules_set",
     "is_match",
-    "match_name",
-    "match_class",
 ]
 
 
@@ -83,13 +82,16 @@ def match_named_parameters(
     """
     unmatched_targets = set(targets)
     for module_name, module in model.named_modules():
+        if isinstance(module, InternalModule):
+            continue
+
         for param_name, param in module.named_parameters(recurse=False):
             param_fqn = f"{module_name}.{param_name}"
             for target in targets:
-                if match_name(param_fqn, target):
+                if _match_name(param_fqn, target):
                     unmatched_targets -= {target}
 
-                    if not any(match_name(param_fqn, ign) for ign in ignore):
+                    if not any(_match_name(param_fqn, ign) for ign in ignore):
                         yield param_fqn, module, param
 
     if warn_on_fail:
@@ -165,11 +167,14 @@ def match_modules_set(
 def is_match(name: str, module: torch.nn.Module, target: str) -> bool:
     """
     Returns true if either module name or module parent classes match against target
+    and the module is not an internal module
     """
-    return match_name(name, target) or match_class(module, target)
+    return not isinstance(module, InternalModule) and (
+        _match_name(name, target) or _match_class(module, target)
+    )
 
 
-def match_name(name: str, target: str) -> bool:
+def _match_name(name: str, target: str) -> bool:
     """
     Returns true if target string begins with "re:" and
     regex matches or if target string exactly matches name
@@ -180,7 +185,7 @@ def match_name(name: str, target: str) -> bool:
         return target == name
 
 
-def match_class(module: torch.nn.Module, target: str) -> bool:
+def _match_class(module: torch.nn.Module, target: str) -> bool:
     """
     Returns true if any torch parent class names match the target string exactly
     """

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250724'
+__version__ = version = '0.10.3.a20250731'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250724.dist-info → compressed_tensors-0.10.3a20250731.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250724
+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-0.10.3a20250724.dist-info → compressed_tensors-0.10.3a20250731.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=LKiXh8O_XB2unUsk0HmC-_PgpfbOswj5PZqtDHOPnRg,523
+compressed_tensors/version.py,sha256=cuOuj6FL5GE-iPKjLVFuRjlwW0_6uDC3tDxFkkHyXFg,523
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -9,7 +9,7 @@ compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=e-2n
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=KvaFBL_Q84LxRGJOV035M8OBoCkAx8kOkfphswgkKWk,745
 compressed_tensors/compressors/quantized_compressors/base.py,sha256=YGUMzbxekj_36ChgQnVZN6T8uDjXtGG1zfMIBGBLWco,10354
 compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=0ANDcuD8aXPqTYNPY6GnX9iS6eXJw6P0TzNV_rYS2l8,5369
-compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=Gw-lVzk5jrKUlM5UTCiJBmhM5gHzB9mn8r298MVUbDI,6395
+compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=tKEaYom4SdMwZWg4MDMMMLNGTLgcVT20lPzewboVpMM,7145
 compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=47W1hFTi5YHVNKEWptzztsSutwI1kxy2Troh-NW1y14,11244
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
 compressed_tensors/compressors/sparse_compressors/base.py,sha256=YNZWcHjDleAlqbgRZQ6oJf44MQb_UDNvJGOqhl26uFA,8098
@@ -28,15 +28,15 @@ compressed_tensors/linear/compressed_linear.py,sha256=1yo9RyjA0aQ--iuIknFfcSorJn
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=yKTj_4lAy_pnXeTCyUADpyz2qAzJXYJU2P03NF_TP68,12835
 compressed_tensors/quantization/quant_config.py,sha256=w6sEEZGVGIF0Ub2r_cqRfZwbkBT8WzfY3ug52olmjGY,10049
-compressed_tensors/quantization/quant_scheme.py,sha256=qApRLsPxELe5S2qFv8OVyAZ5TpRL7gT35i4U3c9PAwI,8461
+compressed_tensors/quantization/quant_scheme.py,sha256=xk2LPn18tjS1PEOyf0WKvavBq3rzAVHFLB3H2mQQWnc,8473
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
 compressed_tensors/quantization/lifecycle/apply.py,sha256=wM8mVcbKvZjBo18pSXMp28i30YWwUXJPSS7_HCakH9U,17892
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
-compressed_tensors/quantization/lifecycle/forward.py,sha256=jT70Mbbu9pH10vu5ALVD7VWGoFdMEUpxmihGrf4frjM,17432
+compressed_tensors/quantization/lifecycle/forward.py,sha256=lQwibkDGroJqONhP9ATZWwaZF9suPmCZMQEagFlFc94,17329
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=3Vuj1a-Y7f_7QXagG7BAeAPnDGtbWGFJXBATg6eT-O0,10241
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=BM7bR_uNa-Ex4T-roHonWiRaxCi5sFysXyl0cFh1ZVs,10257
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
-compressed_tensors/quantization/utils/helpers.py,sha256=Je96Wai9SOizbdE5ph0nsJ86zS96lE4fkf_9q9o2tpA,17212
+compressed_tensors/quantization/utils/helpers.py,sha256=7a89X0kg6xDGplw6trOrkRQzMRPu-txY_qvEt07Vcgc,17036
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=0s15BxdGgzBv8RL4kUJCYcuDOFUh_KZYvNvLEeRqWTc,11956
 compressed_tensors/transform/__init__.py,sha256=v2wfl4CMfA6KbD7Hxx_MbRev63y_6QLDlccZq-WTtdw,907
@@ -46,7 +46,7 @@ compressed_tensors/transform/transform_config.py,sha256=A3RuLNDqBNEByQNeu40Kg7sI
 compressed_tensors/transform/transform_scheme.py,sha256=uGLC4avdbhrVqNC3-Eo0p7WzNRQK92Fpg0N9hWiuCRQ,1752
 compressed_tensors/transform/factory/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/transform/factory/base.py,sha256=Zplf8QO-mFqGwDEhLdYL_afSu7v4nMa79oNhidRNPvY,5880
-compressed_tensors/transform/factory/hadamard.py,sha256=iJ2OyKitR2Duw0z5Jqj69GTih2C1WtHRXQCTtATaTtw,4180
+compressed_tensors/transform/factory/hadamard.py,sha256=B0BVjbF3y707MO6L2XfEoZJTQU965vU9dUPLOiUSXII,4193
 compressed_tensors/transform/factory/matrix_multiply.py,sha256=LdoV2E12HTucmUWcw7UKOpRNnL8QhOOIUnNVlpOpGiI,3925
 compressed_tensors/transform/factory/random_hadamard.py,sha256=nUhTlFa4ikSpcl4Umme71pnjMPgwYoGlwjKlU27UHZ4,1634
 compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
@@ -56,14 +56,14 @@ compressed_tensors/transform/utils/matrix.py,sha256=FIHCUlpWVIIhdr3c6EbQec41JeiP
 compressed_tensors/utils/__init__.py,sha256=KZctuotCmX4byXhwDvSeXgp-Ny_awpziAX-WUkZfodI,853
 compressed_tensors/utils/helpers.py,sha256=Q3iRAa2XSdmmn4vSpUplnvKOmWwn4Clao9ZkPBHXtpI,12604
 compressed_tensors/utils/internal.py,sha256=7SSWgDoNFRnlfadwkoFhLW-T2jOc7Po_WzWv5h32Sa8,982
-compressed_tensors/utils/match.py,sha256=DjqTH-J9-E7ULVXPLV-HBRhdi07JhK-H90PbFK-DRAY,7017
+compressed_tensors/utils/match.py,sha256=ZVBPzrGYExq7-6RRUlU5XeCjl0ooLaNUoDO6Cgnn9cY,7220
 compressed_tensors/utils/offload.py,sha256=3XiBuWbUkBAt8v1t5i57qDcbB3VJQs_FDeayi-JzIWg,23896
 compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.10.3a20250724.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.3a20250724.dist-info/METADATA,sha256=ZH66sWeKBfvuLUe-ArnII1LYXG3UAEYUt6D6YPQ_W-M,7031
-compressed_tensors-0.10.3a20250724.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.3a20250724.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.3a20250724.dist-info/RECORD,,
+compressed_tensors-0.10.3a20250731.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.3a20250731.dist-info/METADATA,sha256=1NCpfVbLTf6aGJ38rJz3Lmu9DptHpuYm5vTRxIB9PB8,7031
+compressed_tensors-0.10.3a20250731.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.3a20250731.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.3a20250731.dist-info/RECORD,,