compressed-tensors-nightly 0.6.0.20240926__tar.gz → 0.6.0.20240928__tar.gz

{compressed-tensors-nightly-0.6.0.20240926/src/compressed_tensors_nightly.egg-info → compressed-tensors-nightly-0.6.0.20240928}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.6.0.20240926
+Version: 0.6.0.20240928
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed-tensors-nightly-0.6.0.20240926 → compressed-tensors-nightly-0.6.0.20240928}/src/compressed_tensors/compressors/model_compressor.py

@@ -216,6 +216,11 @@ class ModelCompressor:
         self.sparsity_compressor = None
         self.quantization_compressor = None
 
+
+        if sparsity_config and sparsity_config.format == CompressionFormat.dense.value:
+            # ignore dense sparsity config
+            self.sparsity_config = None
+
         if sparsity_config is not None:
             self.sparsity_compressor = Compressor.load_from_registry(
                 sparsity_config.format, config=sparsity_config
@@ -306,16 +311,16 @@ class ModelCompressor:
         with open(config_file_path, "r") as config_file:
             config_data = json.load(config_file)
 
-        config_data[COMPRESSION_CONFIG_NAME] = {}
+        config_data[QUANTIZATION_CONFIG_NAME] = {}
         if self.quantization_config is not None:
             quant_config_data = self.quantization_config.model_dump()
-            config_data[COMPRESSION_CONFIG_NAME] = quant_config_data
+            config_data[QUANTIZATION_CONFIG_NAME] = quant_config_data
         if self.sparsity_config is not None:
             sparsity_config_data = self.sparsity_config.model_dump()
-            config_data[COMPRESSION_CONFIG_NAME][
+            config_data[QUANTIZATION_CONFIG_NAME][
                 SPARSITY_CONFIG_NAME
             ] = sparsity_config_data
-        config_data[COMPRESSION_CONFIG_NAME][
+        config_data[QUANTIZATION_CONFIG_NAME][
             COMPRESSION_VERSION_NAME
         ] = compressed_tensors.__version__
 

{compressed-tensors-nightly-0.6.0.20240926 → compressed-tensors-nightly-0.6.0.20240928}/src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -75,7 +75,10 @@ def initialize_module_for_quantization(
 
         if scheme.input_activations is not None:
             _initialize_scale_zero_point_observer(
-                module, "input", scheme.input_activations
+                module,
+                "input",
+                scheme.input_activations,
+                force_zero_point=force_zero_point,
             )
         if scheme.weights is not None:
             if hasattr(module, "weight"):
@@ -83,7 +86,11 @@ def initialize_module_for_quantization(
                 if isinstance(module, torch.nn.Linear):
                     weight_shape = module.weight.shape
                 _initialize_scale_zero_point_observer(
-                    module, "weight", scheme.weights, weight_shape=weight_shape
+                    module,
+                    "weight",
+                    scheme.weights,
+                    weight_shape=weight_shape,
+                    force_zero_point=force_zero_point,
                 )
             else:
                 _LOGGER.warning(

{compressed-tensors-nightly-0.6.0.20240926 → compressed-tensors-nightly-0.6.0.20240928/src/compressed_tensors_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.6.0.20240926
+Version: 0.6.0.20240928
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.