compressed-tensors-nightly 0.3.3.20240526__py3-none-any.whl → 0.3.3.20240528__py3-none-any.whl

compressed_tensors/compressors/pack_quantized.py

@@ -41,7 +41,7 @@ class PackedQuantizationCompressor(Compressor):
     """
 
     COMPRESSION_PARAM_NAMES = [
-        "weight",
+        "weight_packed",
         "weight_scale",
         "weight_zero_point",
         "weight_shape",
@@ -73,7 +73,6 @@ class PackedQuantizationCompressor(Compressor):
                 zp = model_state.get(merge_names(prefix, "weight_zero_point"), None)
                 shape = torch.tensor(value.shape)
                 if scale is not None and zp is not None:
-                    # weight is quantized, compress it
                     # weight is quantized, compress it
                     quant_args = model_quant_args[prefix]
                     if can_quantize(value, quant_args):
@@ -85,13 +84,17 @@ class PackedQuantizationCompressor(Compressor):
                             args=quant_args,
                             dtype=torch.int8,
                         )
-                        value = pack_4bit_ints(value.cpu())
+                    value = pack_4bit_ints(value.cpu())
                     compressed_dict[merge_names(prefix, "weight_shape")] = shape
+                    compressed_dict[merge_names(prefix, "weight_packed")] = value
+                    continue
+
             elif name.endswith("zero_point"):
                 if torch.all(value == 0):
                     # all zero_points are 0, no need to include in
                     # compressed state_dict
                     continue
+
             compressed_dict[name] = value.to("cpu")
 
         return compressed_dict
@@ -126,7 +129,7 @@ class PackedQuantizationCompressor(Compressor):
                     # zero_point assumed to be 0 if not included in state_dict
                     zero_point = torch.zeros_like(scale)
 
-                weight = weight_data["weight"]
+                weight = weight_data["weight_packed"]
                 original_shape = torch.Size(weight_data["weight_shape"])
                 unpacked = unpack_4bit_ints(weight, original_shape)
                 decompressed = dequantize(

{compressed_tensors_nightly-0.3.3.20240526.dist-info → compressed_tensors_nightly-0.3.3.20240528.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.3.3.20240526
+Version: 0.3.3.20240528
 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-0.3.3.20240526.dist-info → compressed_tensors_nightly-0.3.3.20240528.dist-info}/RECORD

@@ -7,7 +7,7 @@ compressed_tensors/compressors/dense.py,sha256=G_XHbvuENyupIKlXSITOQgvPkNkcMEOLc
 compressed_tensors/compressors/helpers.py,sha256=k9avlkmeYj6vkOAvl-MgcixtP7ib24SCfhzZ-RusXfw,5403
 compressed_tensors/compressors/int_quantized.py,sha256=bPi62n1MjySOeBat_yWMyc_LvDNDeSihu1gxzo_YrNY,5203
 compressed_tensors/compressors/model_compressor.py,sha256=gHD2VMbXkXaZiJu3ibOaWiYb4oJDz2hxX03wDuu1yhI,10481
-compressed_tensors/compressors/pack_quantized.py,sha256=VFaHQU-f1QuXuTyOtn19p015KHveXe-NeNJ97ATuOR8,8344
+compressed_tensors/compressors/pack_quantized.py,sha256=zNQnnefK3qWpIzeGTBldQrRc8LM0p-nfSAN9Q9zUAIE,8412
 compressed_tensors/compressors/sparse_bitmask.py,sha256=H9oZSTYI1oRCzAMbd4zThUnZd1h2rfs8DmA3tPcvuNE,8637
 compressed_tensors/config/__init__.py,sha256=ZBqWn3r6ku1qfmlHHYp0mQueY0i7Pwhr9rbQk9dDlMc,704
 compressed_tensors/config/base.py,sha256=grf5tDaLep8i2-W_p7H-fW9DOGXDi4Zz7su7zjs1Qqc,1454
@@ -36,8 +36,8 @@ compressed_tensors/registry/registry.py,sha256=fxjOjh2wklCvJhQxwofdy-zV8q7MkQ85S
 compressed_tensors/utils/__init__.py,sha256=5DrYjoZbaEvSkJcC-GRSbM_RBHVF4tG9gMd3zsJnjLw,665
 compressed_tensors/utils/helpers.py,sha256=h0jfl9drs5FAx40tCHRcVtJqXixB5hT5yq_IG2aY_-w,1735
 compressed_tensors/utils/safetensors_load.py,sha256=wo9UirGrGlenBqZeqotvpCT7D5MEdjCo2J3HeRaIFoU,8502
-compressed_tensors_nightly-0.3.3.20240526.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.3.3.20240526.dist-info/METADATA,sha256=TMoyMn2nDVzv5poJ6neJkosQzlYuvgTKLG3_MkerReA,5633
-compressed_tensors_nightly-0.3.3.20240526.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-compressed_tensors_nightly-0.3.3.20240526.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.3.3.20240526.dist-info/RECORD,,
+compressed_tensors_nightly-0.3.3.20240528.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.3.3.20240528.dist-info/METADATA,sha256=Xl260iCZ0lSPgSbvxBHUqPPC7c75L1nIxfOtXoc08kU,5633
+compressed_tensors_nightly-0.3.3.20240528.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+compressed_tensors_nightly-0.3.3.20240528.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.3.3.20240528.dist-info/RECORD,,