compressed-tensors 0.10.3a20250715__tar.gz → 0.10.3a20250721__tar.gz

{compressed_tensors-0.10.3a20250715/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250721}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250715
+Version: 0.10.3a20250721
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/quantization/lifecycle/forward.py

@@ -111,11 +111,18 @@ def dequantize(
         elif scale.ndim == 2:
             if scale.shape[1] == 1:
                 args = QuantizationArgs(strategy=QuantizationStrategy.CHANNEL)
-            else:
+            # 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] 
+            # 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)
         else:
             raise ValueError(
                 f"Could not infer a quantization strategy from scale with {scale.ndim} "
@@ -189,7 +196,63 @@ def _process_quantization(
     q_min, q_max = calculate_range(args, x.device)
     group_size = args.group_size
 
-    if args.strategy in (QuantizationStrategy.GROUP, QuantizationStrategy.TENSOR_GROUP):
+    # blockwise FP8: quantize per 2D block, supports block_structure for static block quant
+    if args.strategy == QuantizationStrategy.BLOCK:
+        original_shape = x.shape
+        rows, cols = x.shape[-2], x.shape[-1]
+        block_height, block_width = args.block_structure
+
+        # Ensure exact division (tensor dimensions must be divisible by block size)
+        if rows % block_height != 0:
+            raise ValueError(
+                f"Tensor height {rows} is not divisible by block_height {block_height}. "
+                f"Block quantization requires exact division."
+            )
+        if cols % block_width != 0:
+            raise ValueError(
+                f"Tensor width {cols} is not divisible by block_width {block_width}. "
+                f"Block quantization requires exact division."
+            )
+
+        # reshape into blocks and transpose to make each block contiguous
+        num_rows_blocks = rows // block_height
+        num_cols_blocks = cols // block_width
+        x_blocks = x.reshape(
+            num_rows_blocks,
+            block_height,
+            num_cols_blocks,
+            block_width,
+        ).transpose(1, 2)
+
+        # expand scale/zero_point for blocks
+        sb = scale.unsqueeze(-1).unsqueeze(-1)
+        zb = zero_point.unsqueeze(-1).unsqueeze(-1) if zero_point is not None else None
+        if do_quantize:
+            # quantize blocks
+            x_blocks = _quantize(
+                x=x_blocks,
+                scale=sb,
+                zero_point=zb,
+                q_min=q_min,
+                q_max=q_max,
+                args=args,
+                dtype=dtype,
+                global_scale=global_scale,
+            )
+        if do_dequantize:
+            # dequantize blocks
+            x_blocks = _dequantize(
+                x_q=x_blocks,
+                scale=sb,
+                zero_point=zb,
+                global_scale=global_scale,
+            )
+        # restore original shape
+        output = x_blocks.transpose(1, 2).reshape(original_shape)
+    elif args.strategy in (
+        QuantizationStrategy.GROUP,
+        QuantizationStrategy.TENSOR_GROUP,
+    ):
         n_dims = x.shape
         if len(n_dims) > 2:
             x = x.squeeze(0)

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/quantization/quant_args.py

@@ -14,7 +14,7 @@
 
 import warnings
 from enum import Enum
-from typing import Any, Dict, Optional, Union
+from typing import Any, Dict, List, Optional, Union
 
 import torch
 from compressed_tensors.utils import Aliasable
@@ -153,8 +153,8 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
     :param symmetric: whether or not quantization scale is symmetric about zero-point
     :param strategy: string id determining the scope of scale/zero-point to apply
     :param group_size: group length to use for the group strategy
-    :param block_structure: 2d block structure to use for the block strategy, must be
-    of the format "2x4", "8x16", etc.
+    :param block_structure: 2d block structure to use for the block strategy; must be
+        a list of two ints [rows, cols] like [128, 128].
     :param dynamic: set True to perform dynamic quantization - values will not be
         calibrated during calibration phase, instead during inference new quantization
         ranges will be observed with every sample. Defaults to False for static
@@ -169,7 +169,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
     symmetric: bool = True
     group_size: Optional[int] = None
     strategy: Optional[QuantizationStrategy] = None
-    block_structure: Optional[str] = None
+    block_structure: Optional[List[int]] = None
     dynamic: Union[DynamicType, bool] = False
     actorder: Union[ActivationOrdering, bool, None] = None
     observer: Optional[str] = Field(
@@ -207,6 +207,28 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
 
         return value
 
+    @field_validator("block_structure", mode="before")
+    def validate_block_structure(cls, value) -> Optional[List[int]]:
+        if value is None:
+            return value
+        # For backward compatibility, allow string format "2x4", "8x16", etc.
+        if isinstance(value, str):
+            try:
+                return [int(x) for x in value.split("x")]
+            except Exception:
+                raise ValueError(
+                    f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+                )
+        if isinstance(value, (list, tuple)):
+            if len(value) != 2 or not all(isinstance(v, int) for v in value):
+                raise ValueError(
+                    f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+                )
+            return list(value)
+        raise ValueError(
+            f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+        )
+
     @field_validator("strategy", mode="before")
     def validate_strategy(cls, value) -> Union[QuantizationStrategy, None]:
         if isinstance(value, str):
@@ -277,14 +299,15 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
 
         # infer observer w.r.t. dynamic
         if dynamic:
-            if strategy not in (
+            supported_strategies = (
                 QuantizationStrategy.TOKEN,
                 QuantizationStrategy.TENSOR,
                 QuantizationStrategy.TENSOR_GROUP,
-            ):
+                QuantizationStrategy.GROUP,
+            )
+            if strategy not in supported_strategies:
                 raise ValueError(
-                    f"One of {(QuantizationStrategy.TOKEN, QuantizationStrategy.TENSOR, QuantizationStrategy.TENSOR_GROUP)} "
-                    "must be used for dynamic quantization",
+                    f"One of {supported_strategies} must be used for dynamic quantization"
                 )
 
             if (

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/quantization/quant_scheme.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import warnings
 from copy import deepcopy
 from typing import Any, Dict, List, Optional
 
@@ -52,6 +53,7 @@ class QuantizationScheme(BaseModel):
     def validate_model_after(model: "QuantizationScheme") -> "QuantizationScheme":
         inputs = model.input_activations
         outputs = model.output_activations
+        weights = model.weights
 
         if inputs is not None:
             if inputs.actorder is not None:
@@ -61,6 +63,21 @@ class QuantizationScheme(BaseModel):
             if outputs.actorder is not None:
                 raise ValueError("Cannot apply actorder to output activations")
 
+        if (
+            inputs and weights
+            and weights.strategy == QuantizationStrategy.GROUP 
+            and inputs.strategy == QuantizationStrategy.GROUP
+            and weights.group_size != inputs.group_size
+        ):
+            warnings.warn(
+                "Using GROUP strategy for both weights and input_activations "
+                f"with different group sizes ({weights.group_size} vs {inputs.group_size}) "
+                "may complicate fused kernel implementations. Consider using "
+                "TENSOR_GROUP strategy for both or matching group sizes.",
+                UserWarning,
+                stacklevel=2
+            )
+
         return model
 
 
@@ -243,6 +260,29 @@ FP8_DYNAMIC = dict(
     ),
 )
 
+# Block‐wise FP8 (deepseekv3-style quantization):
+# static 128x128 per‐block weights and
+# dynamic per‐token‐group activations
+FP8_BLOCK = dict(
+    weights=QuantizationArgs(
+        num_bits=8,
+        type=QuantizationType.FLOAT,
+        strategy=QuantizationStrategy.BLOCK,
+        symmetric=True,
+        dynamic=False,
+        block_structure=[128, 128],
+    ),
+    input_activations=QuantizationArgs(
+        num_bits=8,
+        type=QuantizationType.FLOAT,
+        strategy=QuantizationStrategy.GROUP,
+        symmetric=True,
+        dynamic=True,
+        observer=None,
+        group_size=128,
+    ),
+)
+
 PRESET_SCHEMES = {
     # Unquantized (no-op)
     "UNQUANTIZED": UNQUANTIZED,
@@ -257,6 +297,7 @@ PRESET_SCHEMES = {
     # Float weight and activation schemes
     "FP8": FP8,
     "FP8_DYNAMIC": FP8_DYNAMIC,
+    "FP8_BLOCK": FP8_BLOCK,
     "NVFP4A16": NVFP4A16,
     "NVFP4": NVFP4,
 }

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/quantization/utils/helpers.py

@@ -171,7 +171,10 @@ def compute_dynamic_scales_and_zp(
         reduce_dims = tuple(idx for idx in range(value.ndim) if idx not in dim)
     elif args.strategy == QuantizationStrategy.TENSOR:
         reduce_dims = None
-    elif args.strategy == QuantizationStrategy.TENSOR_GROUP:
+    elif args.strategy in (
+        QuantizationStrategy.TENSOR_GROUP,
+        QuantizationStrategy.GROUP,
+    ):
         if len(value.shape) > 2:
             value = value.squeeze(0)
 
@@ -187,9 +190,15 @@ def compute_dynamic_scales_and_zp(
             ),
         )
     else:
+        supported_strategies = (
+            QuantizationStrategy.TOKEN,
+            QuantizationStrategy.TENSOR,
+            QuantizationStrategy.TENSOR_GROUP,
+            QuantizationStrategy.GROUP,
+        )
         raise ValueError(
             "Dynamic quantization is only supported for ",
-            f"{QuantizationStrategy.TOKEN, QuantizationStrategy.TENSOR, QuantizationStrategy.TENSOR_GROUP}",
+            f"{supported_strategies}",
         )
 
     if not reduce_dims:

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/transform/factory/base.py

@@ -117,10 +117,8 @@ class TransformFactory(RegistryMixin, ABC):
             TransformLocation.WEIGHT_INPUT,
             TransformLocation.WEIGHT_OUTPUT,
         ):
-            assert isinstance(module, torch.nn.Linear)
-            assert module.bias is None
-
             # fuse transform into weight
+            assert hasattr(module, "weight")
             with torch.no_grad(), align_module_device(module):
                 update_offload_parameter(module, "weight", transform(module.weight))
 

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/transform/factory/hadamard.py

@@ -14,13 +14,14 @@
 
 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
 from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from compressed_tensors.transform.utils.utils import (
+from compressed_tensors.transform.utils.matrix import (
     apply_transform_weight,
-    get_matrix_size,
+    get_transform_size,
 )
 from compressed_tensors.utils import get_execution_device, get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
@@ -51,8 +52,8 @@ class HadamardFactory(TransformFactory):
         :param module: parent module that transform will be applied to
         :param args: defines how the transform will be applied to the module
         """
-        assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        assert hasattr(module, "weight")
+        size = get_transform_size(module, args.location, self.scheme.head_dim)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
         exec_device = get_execution_device(module)
@@ -60,7 +61,7 @@ class HadamardFactory(TransformFactory):
         factory_kwargs = {"construct_device": exec_device}
         weight = self.weights.get(size, dtype, device, factory_kwargs=factory_kwargs)
         perm = self.perms[weight] if self.scheme.randomize else None
-        return HadamardTransform(weight, perm, args)
+        return HadamardTransform(weight, perm, args, type(module))
 
     def _create_weight(
         self,
@@ -81,12 +82,18 @@ class HadamardFactory(TransformFactory):
 
 class HadamardTransform(TransformBase):
     def __init__(
-        self, weight: Parameter, perm: Union[Parameter, None], args: TransformArgs
+        self,
+        weight: Parameter,
+        perm: Optional[Parameter],
+        args: TransformArgs,
+        module_type: type[torch.nn.Module],
     ):
         super().__init__()
         self.weight = weight
         self.perm = perm
         self.args = args
+        self.module_type = module_type
+        self._scale = math.sqrt(weight.size(0))
 
     def forward(self, value: Tensor) -> Tensor:
         weight = self.weight
@@ -96,5 +103,7 @@ class HadamardTransform(TransformBase):
 
         if self.args.inverse:
             weight = weight.T
-
-        return apply_transform_weight(weight, value, self.args.location)
+ 
+        return apply_transform_weight(
+            weight, value, self.args.location, self.module_type
+        ) / self._scale

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -17,9 +17,9 @@ from typing import Optional
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
 from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
-from compressed_tensors.transform.utils.utils import (
+from compressed_tensors.transform.utils.matrix import (
     apply_transform_weight,
-    get_matrix_size,
+    get_transform_size,
 )
 from compressed_tensors.utils import get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
@@ -50,8 +50,8 @@ class RandomMatrixFactory(TransformFactory):
         :param module: parent module that transform will be applied to
         :param args: defines how the transform will be applied to the module
         """
-        assert isinstance(module, Linear)
-        size = get_matrix_size(module, args.location)
+        assert hasattr(module, "weight")
+        size = get_transform_size(module, args.location, self.scheme.head_dim)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
 
@@ -59,7 +59,7 @@ class RandomMatrixFactory(TransformFactory):
         if args.inverse:
             weight = self.inverses[weight]
 
-        return RandomMatrixTransform(weight, args)
+        return RandomMatrixTransform(weight, args, type(module))
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
         # TODO: verify that weight is invertible (has non-zero determinant)
@@ -74,17 +74,27 @@ class RandomMatrixFactory(TransformFactory):
 
 
 class RandomMatrixTransform(TransformBase):
-    def __init__(self, weight: Tensor, args: TransformArgs):
+    def __init__(
+        self,
+        weight: Tensor,
+        args: TransformArgs,
+        module_type: type[torch.nn.Module],
+    ):
         super().__init__()
         self.weight = weight  # is an inverse if args.inverse
         self.args = args
+        self.module_type = module_type
 
     def forward(self, value: Tensor) -> Parameter:
-        return apply_transform_weight(self.weight, value, self.args.location)
+        return apply_transform_weight(
+            self.weight, value, self.args.location, self.module_type
+        )
 
     def right_inverse(self, value: Tensor) -> Tensor:
         inverse = high_precision_invert(self.weight)
-        return apply_transform_weight(inverse, value, self.args.location)
+        return apply_transform_weight(
+            inverse, value, self.args.location, self.module_type
+        )
 
 
 def high_precision_invert(weight: Tensor) -> Tensor:

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/transform/transform_scheme.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import List
+from typing import List, Optional
 
 from compressed_tensors.transform import TransformArgs
 from pydantic import BaseModel, Field
@@ -40,3 +40,4 @@ class TransformScheme(BaseModel):
     apply: List[TransformArgs] = Field(default_factory=list)
     randomize: bool = Field(default=False)
     requires_grad: bool = Field(default=False)
+    head_dim: Optional[int] = Field(default=None)

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/transform/utils/hadamard.py

@@ -59,7 +59,7 @@ def deterministic_hadamard_matrix(
     for _ in range(log2):
         H = torch.vstack((torch.hstack((H, H)), torch.hstack((H, -H))))
 
-    return H / math.sqrt(size)
+    return H
 
 
 def random_hadamard_matrix(
@@ -86,7 +86,7 @@ def random_hadamard_matrix(
     Q = Q.to(device=device)
     Q = Q * 2 - 1
     Q = torch.diag(Q)
-    return _matmul_hadU(Q) / math.sqrt(size)
+    return _matmul_hadU(Q)
 
 
 def is_pow2(n: int) -> bool:

compressed_tensors-0.10.3a20250721/src/compressed_tensors/transform/utils/matrix.py

@@ -0,0 +1,179 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Optional, Tuple
+
+import torch
+from compressed_tensors.transform import TransformLocation
+
+
+__all__ = ["get_transform_size", "apply_transform_weight"]
+
+
+def get_transform_size(
+    module: torch.nn.Module,
+    location: TransformLocation,
+    head_dim: Optional[int] = None,
+) -> int:
+    """
+    Determine the size of a transform matrix given its location on the module
+
+    :param module: module that matrix will be applied to
+    :param location: location on module
+    :param head_dim: size of head when transform is applied to mha
+    :return: size of matrix
+    """
+    if isinstance(module, torch.nn.Linear):
+        if location in (TransformLocation.INPUT, TransformLocation.WEIGHT_INPUT):
+            size = module.in_features
+        else:
+            size = module.out_features
+    elif isinstance(module, torch.nn.Embedding):
+        if location in (TransformLocation.INPUT, TransformLocation.WEIGHT_INPUT):
+            size = module.num_embeddings
+        else:
+            size = module.embedding_dim
+    else:
+        raise NotImplementedError(f"Transforms on {type(module)} are not supported")
+
+    if head_dim is not None:
+        if size % head_dim != 0:
+            raise ValueError(
+                f"{head_dim} must divide {size} for {type(module)} at {location}"
+            )
+
+        size = head_dim
+
+    return size
+
+
+def apply_transform_weight(
+    transform_weight: torch.Tensor,
+    value: torch.Tensor,
+    location: TransformLocation,
+    module_type: type[torch.nn.Module],
+) -> torch.Tensor:
+    """
+    Using the transform location, apply the transform_weight to the
+    given value wrt linear weights. For more info on input and output transforms,
+    see `TransformLocation`
+
+    The following explains how weights should be applied to values according to location
+
+    let  x          be input activation
+         W          be weight,
+         yh, xh, Wh be transformed output, input, weight
+
+    note that
+         y  = (x W.T)        // torch.nn.Linear
+
+    Choose values for yh, xh, and Wh which incorporate matrix transforms
+
+    let  V, Vi      be transform matrices on input side
+         U, Ui      be transform matrices on output side
+
+    pick xh = (x V)
+         Wh = (U.T W Vi.T)
+         yh = (y U)
+
+    The following shows that `yh = (xh) (Wh).T` for the chosen values of yh, xh, and Wh
+
+    (xh) (Wh).T = (x V) (U.T W Vi.T).T
+                = (x V) (Vi W.T U)        // transpose matrix product identity
+                = (x W.T) U
+                = y U
+                = yh
+
+    :param transform_weight: transform weight to apply
+    :param value: value to apply transform_weight to
+    :param location: determines how weight should be applied
+    :param model_type: result of type(module), passed in to determine application of
+        weight transform
+    :return: value after transform_weight has been applied
+    """
+
+    assert transform_weight.shape[0] == transform_weight.shape[1]
+
+    if module_type == torch.nn.Linear:
+        if location == TransformLocation.INPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.WEIGHT_INPUT:
+            # equivalent to (transform_weight @ value.T).T
+            return _multihead_matmul(value, transform_weight.T)
+
+        elif location == TransformLocation.WEIGHT_OUTPUT:
+            # equivalent to (value.T @ transform_weight).T
+            return _multihead_matmul(transform_weight.T, value)
+
+        elif location == TransformLocation.OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+    # similar derivation to torch.nn.Linear, but `y = (x W)`
+    elif module_type == torch.nn.Embedding:
+        if location == TransformLocation.INPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.WEIGHT_INPUT:
+            return _multihead_matmul(
+                transform_weight,
+                value,
+            )
+
+        elif location == TransformLocation.WEIGHT_OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+        elif location == TransformLocation.OUTPUT:
+            return _multihead_matmul(value, transform_weight)
+
+    raise NotImplementedError(
+        f"Applying transforms to {module_type} {location} is not supported"
+    )
+
+
+def _multihead_matmul(A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
+    """
+    Performs A @ B for last two dims of two matrices A and B that possibly
+    have different shapes, as is the case in multi-headed dimension. If
+    shapes are different, this is equivalent to converting the last two dims
+    of the smaller matrix into a block-diagonal matrix with the same shape as
+    the last two dims of the larger matrix.
+
+    E.g. if A is half the size of B, this function will perform
+    [[A  ]  @ B
+     [  A]]
+
+    If B is a third of the size of A, this function will perform
+    A @ [[B    ]
+         [  B  ]
+         [    B]]
+
+    This function will error out if the shapes are not evenly divisble
+
+    :param A: left-hand tensor
+    :param B: right-hand tensor
+    :return: result
+    """
+    if A.shape[-1] > B.shape[-2]:
+        head_dim = B.shape[-2]
+        num_heads = A.shape[-1] // head_dim
+        A = A.unflatten(-1, (num_heads, head_dim))
+        return (A @ B).flatten(-2, -1)
+    elif A.shape[-1] < B.shape[-2]:
+        head_dim = A.shape[-1]
+        num_heads = B.shape[-2] // head_dim
+        B = B.unflatten(-2, (num_heads, head_dim))
+        return (A @ B).flatten(-3, -2)
+    else:
+        return A @ B

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250715'
+__version__ = version = '0.10.3.a20250721'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250715
+Version: 0.10.3a20250721
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -84,7 +84,7 @@ src/compressed_tensors/transform/factory/random_hadamard.py
 src/compressed_tensors/transform/utils/__init__.py
 src/compressed_tensors/transform/utils/hadamard.py
 src/compressed_tensors/transform/utils/hadamards.safetensors
-src/compressed_tensors/transform/utils/utils.py
+src/compressed_tensors/transform/utils/matrix.py
 src/compressed_tensors/utils/__init__.py
 src/compressed_tensors/utils/helpers.py
 src/compressed_tensors/utils/internal.py

{compressed_tensors-0.10.3a20250715 → compressed_tensors-0.10.3a20250721}/tests/test_examples/test_bitmask_compression_ipynb.py

@@ -12,8 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import nbformat
 import pytest
+
+
+nbformat = pytest.importorskip("nbformat")
 from nbconvert.preprocessors import ExecutePreprocessor