compressed-tensors 0.13.0__tar.gz → 0.13.1a20260108__tar.gz

{compressed_tensors-0.13.0/src/compressed_tensors.egg-info → compressed_tensors-0.13.1a20260108}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.13.0
+Version: 0.13.1a20260108
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108}/src/compressed_tensors/compressors/base.py

@@ -19,7 +19,7 @@ import torch
 from compressed_tensors.config import SparsityCompressionConfig
 from compressed_tensors.quantization import QuantizationArgs, QuantizationConfig
 from compressed_tensors.registry import RegistryMixin
-from compressed_tensors.utils import has_offloaded_params
+from compressed_tensors.utils import has_offloaded_params, register_offload_parameter
 from torch import Tensor
 from torch.nn import Module
 
@@ -185,10 +185,16 @@ class BaseCompressor(RegistryMixin, ABC):
         for name, parameter in module.named_parameters():
             compressed_data[name] = parameter
 
-        return self.decompress_weight(
+        decompressed_weight = self.decompress_weight(
             compressed_data=compressed_data, quantization_args=quantization_args
         ).to(device)
 
+        for name in ("weight_scale", "weight_zero_point"):
+            if hasattr(module, name):
+                register_offload_parameter(module, name, compressed_data[name])
+
+        return decompressed_weight
+
     def decompress_weight(
         self, compressed_data: Dict[str, Tensor], **kwargs
     ) -> torch.Tensor:

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108}/src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -36,6 +36,7 @@ from compressed_tensors.config import CompressionFormat, SparsityCompressionConf
 from compressed_tensors.config.format import (
     infer_and_set_per_module_quantization_format,
 )
+from compressed_tensors.linear.compressed_linear import CompressedLinear
 from compressed_tensors.quantization import (
     DEFAULT_QUANTIZATION_METHOD,
     QuantizationConfig,
@@ -474,6 +475,9 @@ class ModelCompressor:
             ),
             desc="Compressing model",
         ):
+            if isinstance(module, CompressedLinear):
+                continue  # already compressed
+
             module_device = get_execution_device(module)
             is_meta = module_device.type == "meta"
 

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108}/src/compressed_tensors/compressors/quantized_compressors/fp4_quantized.py

@@ -127,6 +127,11 @@ class NVFP4PackedCompressor(BaseQuantizationCompressor):
         m, n = weight.shape
         # TODO: use a user provided dequant dtype
         unpacked = unpack_fp4_from_uint8(weight, m, n * 2)
+
+        # decompress scale
+        scale = scale.to(unpacked.dtype)
+        compressed_data["weight_scale"] = torch.nn.Parameter(scale, requires_grad=False)
+
         decompressed_weight = dequantize(
             x_q=unpacked, scale=scale, global_scale=global_scale, dtype=unpacked.dtype
         )

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108}/src/compressed_tensors/compressors/quantized_compressors/pack_quantized.py

@@ -175,6 +175,10 @@ class PackedQuantizationCompressor(BaseQuantizationCompressor):
             zero_point = unpack_from_int32(
                 zero_point, num_bits, original_zp_shape, packed_dim=0
             )
+            # Update the compressed_data dict with unpacked zero_point
+            compressed_data["weight_zero_point"] = torch.nn.Parameter(
+                zero_point, requires_grad=False
+            )
 
         decompressed_weight = dequantize(
             x_q=unpacked, scale=scale, zero_point=zero_point, g_idx=g_idx

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.13.0'
-__version_tuple__ = version_tuple = (0, 13, 0)
+__version__ = version = '0.13.1.a20260108'
+__version_tuple__ = version_tuple = (0, 13, 1)

{compressed_tensors-0.13.0 → compressed_tensors-0.13.1a20260108/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.13.0
+Version: 0.13.1a20260108
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/vllm-project/compressed-tensors
 Author: Neuralmagic, Inc.