compressed-tensors 0.11.1a20250904__tar.gz → 0.11.1a20250909__tar.gz

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250904
+Version: 0.11.1a20250909
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py

@@ -140,6 +140,11 @@ def pack_fp4_to_uint8(x: torch.Tensor) -> torch.Tensor:
     m, n = x.shape
     device = x.device
 
+    if n % 2 != 0:
+        raise ValueError(
+            "tensor must have an even number of columns for nvfp4 compression"
+        )
+
     # Create lookup table for FP4 values to indices
     # Map the absolute values to 0-7 indices
     kE2M1 = torch.tensor(FLOAT_TO_E2M1, device=device, dtype=x.dtype)
@@ -155,10 +160,6 @@ def pack_fp4_to_uint8(x: torch.Tensor) -> torch.Tensor:
     # Reshape to prepare for packing pairs of values
     indices = indices.reshape(-1)
 
-    # Handle odd length by padding if necessary
-    if indices.numel() % 2 != 0:
-        indices = torch.cat([indices, torch.zeros(1, dtype=torch.long, device=device)])
-
     # Reshape to pair consecutive elements
     indices = indices.reshape(-1, 2)
 

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/quantization/quant_scheme.py

@@ -60,6 +60,26 @@ class QuantizationScheme(BaseModel):
         format = model.format
 
         if inputs is not None:
+            if inputs.strategy not in (
+                QuantizationStrategy.TOKEN,
+                QuantizationStrategy.TENSOR,
+                QuantizationStrategy.GROUP,
+                QuantizationStrategy.TENSOR_GROUP,
+            ):
+                if (
+                    inputs.strategy == QuantizationStrategy.GROUP
+                    and inputs.dynamic is True
+                ):
+                    raise NotImplementedError(
+                        "Static and local group-wise activation "
+                        "quantization is not supported"
+                    )
+
+                raise NotImplementedError(
+                    f"Using {inputs.strategy} strategy is not supported for "
+                    "activation quantization"
+                )
+
             if inputs.actorder is not None:
                 raise ValueError("Cannot apply actorder to input activations")
 

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/transform/factory/base.py

@@ -18,6 +18,7 @@ from typing import List, Optional, Set, Tuple
 
 import torch
 import torch.nn.utils.parametrize as P
+import tqdm
 from compressed_tensors.registry.registry import RegistryMixin, T
 from compressed_tensors.transform import (
     TransformArgs,
@@ -84,15 +85,21 @@ class TransformFactory(RegistryMixin, ABC):
         """
         raise NotImplementedError()
 
-    def apply_to_model(self, model: Module):
+    def apply_to_model(self, model: Module, use_tqdm=True):
         """
         Create transforms and apply them to the model
 
         :param model: module to apply transforms to
         """
-        for arg in self.scheme.apply:
-            for _, module in match_named_modules(model, arg.targets, arg.ignore):
-                self._apply_to_module(module, arg)
+        modules_args = [
+            (module, arg)
+            for arg in self.scheme.apply
+            for _, module in match_named_modules(model, arg.targets, arg.ignore)
+        ]
+
+        desc = f"Applying {self.name} transforms"
+        for module, arg in tqdm.tqdm(modules_args, desc=desc, disable=(not use_tqdm)):
+            self._apply_to_module(module, arg)
 
         self._update_tied_weights()
 

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/transform/factory/hadamard.py

@@ -53,24 +53,28 @@ class HadamardFactory(TransformFactory):
         """
         assert hasattr(module, "weight")
         size = get_transform_size(module, args.location, self.scheme.head_dim)
-        dtype = self.scheme.precision
-        device = get_offloaded_device(module)
         exec_device = get_execution_device(module)
-
-        factory_kwargs = {"construct_device": exec_device}
-        weight = self.weights.get(size, dtype, device, factory_kwargs=factory_kwargs)
+        device = get_offloaded_device(module)
+        precision = self.scheme.precision if args.is_online() else torch.float64
+
+        factory_kwargs = {
+            "device": device,
+            "construct_device": exec_device,
+            "precision": precision,
+        }
+        weight = self.weights.get(size, factory_kwargs=factory_kwargs)
+        # TODO: permutations should be keyed by fused modules, not weight
         perm = self.perms[weight] if self.scheme.randomize else None
         return HadamardTransform(weight, perm, self.scheme, args, type(module))
 
     def _create_weight(
         self,
         size: int,
-        dtype: dtype,
         device: device,
         construct_device: device,
+        precision: dtype,
     ) -> Parameter:
-        # construct on execution device, cache on offload device
-        data = deterministic_hadamard_matrix(size, dtype, construct_device)
+        data = deterministic_hadamard_matrix(size, precision, construct_device)
         data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
@@ -94,8 +98,7 @@ class HadamardTransform(TransformBase):
         self.scheme = scheme
         self.args = args
         self.module_type = module_type
-        self._scale = torch.tensor(weight.size(0), dtype=self.scheme.precision).sqrt()
-        self._precision = scheme.precision if args.is_online() else torch.float64
+        self._scale = torch.tensor(weight.size(0), dtype=torch.float64).sqrt()
 
     def forward(self, value: Tensor) -> Tensor:
         weight = self.weight
@@ -108,8 +111,8 @@ class HadamardTransform(TransformBase):
 
         return (
             apply_transform_weight(
-                weight.to(self._precision),
-                value.to(self._precision),
+                weight.to(device=value.device),
+                value.to(dtype=weight.dtype),
                 self.args.location,
                 self.module_type,
             )

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -21,8 +21,8 @@ from compressed_tensors.transform.utils.matrix import (
     apply_transform_weight,
     get_transform_size,
 )
-from compressed_tensors.utils import get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
+from compressed_tensors.utils.offload import get_offloaded_device
 from torch import Tensor, device, dtype
 from torch.nn import Module, Parameter
 
@@ -52,19 +52,23 @@ class RandomMatrixFactory(TransformFactory):
         """
         assert hasattr(module, "weight")
         size = get_transform_size(module, args.location, self.scheme.head_dim)
-        dtype = self.scheme.precision
         device = get_offloaded_device(module)
+        precision = self.scheme.precision if args.is_online() else torch.float64
 
-        weight = self.weights[size, dtype, device]
+        factory_kwargs = {"device": device, "precision": precision}
+        weight = self.weights.get(size, factory_kwargs=factory_kwargs)
         if args.inverse:
             weight = self.inverses[weight]
 
         return RandomMatrixTransform(weight, self.scheme, args, type(module))
 
-    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        # TODO: verify that weight is invertible (has non-zero determinant)
+    def _create_weight(self, size: int, device: device, precision: dtype) -> Parameter:
+        # TODO: construct such that weight is invertible (has non-zero determinant)
         data = torch.rand(
-            (size, size), generator=self.generator, dtype=dtype, device=device
+            (size, size),
+            generator=self.generator,
+            dtype=precision,
+            device=device,
         )
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
@@ -87,12 +91,11 @@ class RandomMatrixTransform(TransformBase):
         self.scheme = scheme
         self.args = args
         self.module_type = module_type
-        self._precision = scheme.precision if args.is_online() else torch.float64
 
     def forward(self, value: Tensor) -> Parameter:
         return apply_transform_weight(
-            self.weight.to(self._precision),
-            value.to(self._precision),
+            self.weight.to(device=value.device),
+            value.to(dtype=self.weight.dtype),
             self.args.location,
             self.module_type,
         ).to(value.dtype)
@@ -100,8 +103,8 @@ class RandomMatrixTransform(TransformBase):
     def right_inverse(self, value: Tensor) -> Tensor:
         inverse = high_precision_invert(self.weight)
         return apply_transform_weight(
-            inverse.to(self._precision),
-            value.to(self._precision),
+            inverse.to(device=value.device),
+            value.to(dtype=inverse.dtype),
             self.args.location,
             self.module_type,
         ).to(value.dtype)

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/transform/factory/random_hadamard.py

@@ -31,11 +31,10 @@ class RandomHadamardFactory(HadamardFactory):
     def _create_weight(
         self,
         size: int,
-        dtype: dtype,
         device: device,
         construct_device: device,
+        precision: dtype,
     ) -> Parameter:
-        # construct on execution device, cache on offload device
-        data = random_hadamard_matrix(size, dtype, construct_device, self.generator)
+        data = random_hadamard_matrix(size, precision, construct_device, self.generator)
         data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/utils/offload.py

@@ -68,10 +68,8 @@ except ImportError:
 
 
 __all__ = [
-    "is_module_offloaded",
     "get_execution_device",
     "get_offloaded_device",
-    "update_prefix_dict",
     "update_parameter_data",
     "register_offload_parameter",
     "update_offload_parameter",
@@ -117,11 +115,6 @@ def check_accelerate(fallback: Any):
 """ Candidates for Depreciation """
 
 
-@check_accelerate(fallback=False)
-def is_module_offloaded(module: torch.nn.Module) -> bool:
-    return has_offloaded_params(module)
-
-
 def get_offloaded_device(module: torch.nn.Module) -> torch.device:
     """
     :param module: module to check
@@ -131,26 +124,10 @@ def get_offloaded_device(module: torch.nn.Module) -> torch.device:
         first_key = list(module._hf_hook.weights_map.keys())[0]
         prefix_dataset = module._hf_hook.weights_map.dataset
         return prefix_dataset[first_key].device
-    return next(module.parameters()).device
-
-
-@check_accelerate(fallback=None)
-def update_prefix_dict(module: torch.nn.Module, key: str, data: torch.Tensor):
-    """
-    Updates the offloaded state dict for a given module. Parameter named key is replaced
-    by data. This is neccesary because parameter updates for offloaded modules do not
-    persist automatically between loads. This function only affects the offloaded
-    state dict and not the current state of the loaded module.
-
-    :param module: module containing the parameter to update
-    :param key: name of parameter to update
-    :param data: tensor to update parameter with in the offloaded state dict
-    """
-    if not has_offloaded_params(module):
-        raise ValueError("Prefix dict is only applicable to offloaded modules")
-
-    weights_map = module._hf_hook.weights_map
-    offload_to_weights_map(weights_map, key, data)
+    else:
+        # if the module is not offloaded, then any addded weights
+        # should be placed the module's execution device
+        return get_execution_device(module)
 
 
 def update_parameter_data(

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.11.1.a20250904'
+__version__ = version = '0.11.1.a20250909'
 __version_tuple__ = version_tuple = (0, 11, 1)

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250904
+Version: 0.11.1a20250909
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.11.1a20250904 → compressed_tensors-0.11.1a20250909}/tests/test_compressors/quantized_compressors/test_nvfp4_quant.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import pytest
 import torch
 from compressed_tensors.compressors.quantized_compressors.nvfp4_quantized import (
     pack_fp4_to_uint8,
@@ -41,3 +42,16 @@ def test_pack_unpack():
     sign_bitx = torch.signbit(x)
     sign_bitout = torch.signbit(unpacked)
     assert torch.equal(sign_bitout, sign_bitx)
+
+
+def test_pack_unpack_odd_dims():
+    x = torch.Tensor(
+        [
+            [-0.5000, -6.0000, -0.5000, -1.5000, -1.0000, 6.0000, 0.0000],
+            [-1.0000, -6.0000, -0.5000, -0.0000, 0.5000, 0.5000, -0.0000],
+            [1.5000, 6.0000, -0.0000, -0.5000, 1.0000, 1.0000, -0.0000],
+        ]
+    )
+
+    with pytest.raises((ValueError, torch._dynamo.exc.Unsupported)):
+        _ = pack_fp4_to_uint8(x)