compressed-tensors 0.11.1a20250908__tar.gz → 0.11.1a20250909__tar.gz

{compressed_tensors-0.11.1a20250908 → compressed_tensors-0.11.1a20250909}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250908
+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-0.11.1a20250908 → 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-0.11.1a20250908 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors/quantization/quant_scheme.py

@@ -63,9 +63,19 @@ class QuantizationScheme(BaseModel):
             if inputs.strategy not in (
                 QuantizationStrategy.TOKEN,
                 QuantizationStrategy.TENSOR,
+                QuantizationStrategy.GROUP,
                 QuantizationStrategy.TENSOR_GROUP,
             ):
-                raise ValueError(
+                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"
                 )

{compressed_tensors-0.11.1a20250908 → 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
@@ -137,25 +130,6 @@ def get_offloaded_device(module: torch.nn.Module) -> torch.device:
         return get_execution_device(module)
 
 
-@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)
-
-
 def update_parameter_data(
     module: torch.nn.Module, new_param_data: torch.Tensor, param_name: str
 ):

{compressed_tensors-0.11.1a20250908 → 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.a20250908'
+__version__ = version = '0.11.1.a20250909'
 __version_tuple__ = version_tuple = (0, 11, 1)

{compressed_tensors-0.11.1a20250908 → compressed_tensors-0.11.1a20250909}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250908
+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-0.11.1a20250908 → 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)