compressed-tensors-nightly 0.6.0.20240923__tar.gz → 0.6.0.20240925__tar.gz

{compressed-tensors-nightly-0.6.0.20240923/src/compressed_tensors_nightly.egg-info → compressed-tensors-nightly-0.6.0.20240925}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.6.0.20240923
+Version: 0.6.0.20240925
 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-0.6.0.20240923 → compressed-tensors-nightly-0.6.0.20240925}/src/compressed_tensors/__init__.py

@@ -19,3 +19,4 @@ from .compressors import *
 from .config import *
 from .quantization import QuantizationConfig, QuantizationStatus
 from .utils import *
+from .version import *

{compressed-tensors-nightly-0.6.0.20240923 → compressed-tensors-nightly-0.6.0.20240925}/src/compressed_tensors/compressors/model_compressor.py

@@ -22,6 +22,7 @@ from typing import Any, Dict, Optional, Union
 
 import torch
 import transformers
+import compressed_tensors
 from compressed_tensors.base import (
     COMPRESSION_CONFIG_NAME,
     QUANTIZATION_CONFIG_NAME,
@@ -368,6 +369,7 @@ class ModelCompressor:
             config_data[COMPRESSION_CONFIG_NAME][
                 SPARSITY_CONFIG_NAME
             ] = sparsity_config_data
+        config_data[COMPRESSION_CONFIG_NAME]["version"] = compressed_tensors.__version__
 
         with open(config_file_path, "w") as config_file:
             json.dump(config_data, config_file, indent=2, sort_keys=True)

{compressed-tensors-nightly-0.6.0.20240923 → compressed-tensors-nightly-0.6.0.20240925}/src/compressed_tensors/quantization/quant_scheme.py

@@ -108,7 +108,7 @@ def is_preset_scheme(name: str) -> bool:
 UNQUANTIZED = dict()
 
 # 8 bit integer weights and 8 bit activations quantization
-W8A8 = dict(
+INT8_W8A8 = dict(
     weights=QuantizationArgs(
         num_bits=8,
         type=QuantizationType.INT,
@@ -149,7 +149,7 @@ W4A16 = dict(
 )
 
 # 4 bit integer weights and 8 bit activations quantization
-W4A8 = dict(
+INT8_W4A8 = dict(
     weights=QuantizationArgs(
         num_bits=4,
         type=QuantizationType.INT,
@@ -210,8 +210,9 @@ PRESET_SCHEMES = {
     "W8A16": W8A16,
     "W4A16": W4A16,
     # Integer weight and activation schemes
-    "W8A8": W8A8,
-    "W4A8": W4A8,
+    "W8A8": INT8_W8A8,
+    "INT8": INT8_W8A8, # alias for W8A8
+    "W4A8": INT8_W4A8,
     # Float weight and activation schemes
     "FP8": FP8,
     "FP8_DYNAMIC": FP8_DYNAMIC,

{compressed-tensors-nightly-0.6.0.20240923 → compressed-tensors-nightly-0.6.0.20240925/src/compressed_tensors_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.6.0.20240923
+Version: 0.6.0.20240925
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.