compressed-tensors-nightly 0.3.3.20240523__py3-none-any.whl → 0.3.3.20240524__py3-none-any.whl

compressed_tensors/compressors/int_quantized.py

@@ -78,7 +78,11 @@ class IntQuantizationCompressor(Compressor):
                             args=quant_args,
                             dtype=torch.int8,
                         )
-
+            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
@@ -106,10 +110,16 @@ class IntQuantizationCompressor(Compressor):
                 with safe_open(safe_path, framework="pt", device=device) as f:
                     weight_data[param_name] = f.get_tensor(full_name)
 
-            if len(weight_data) == len(self.COMPRESSION_PARAM_NAMES):
+            if "weight_scale" in weight_data:
+                zero_point = weight_data.get("weight_zero_point", None)
+                scale = weight_data["weight_scale"]
+                if zero_point is None:
+                    # zero_point assumed to be 0 if not included in state_dict
+                    zero_point = torch.zeros_like(scale)
+
                 decompressed = dequantize(
                     x_q=weight_data["weight"],
-                    scale=weight_data["weight_scale"],
-                    zero_point=weight_data["weight_zero_point"],
+                    scale=scale,
+                    zero_point=zero_point,
                 )
                 yield merge_names(weight_name, "weight"), decompressed

compressed_tensors/compressors/model_compressor.py

@@ -249,8 +249,9 @@ class ModelCompressor:
         for name, data in tqdm(dense_weight_generator, desc="Decompressing model"):
             # loading the decompressed weights into the model
             model_device = operator.attrgetter(name)(model).device
-            data_new = Parameter(data.to(model_device))
             data_old = operator.attrgetter(name)(model)
+            data_dtype = data_old.dtype
+            data_new = Parameter(data.to(model_device).to(data_dtype))
             data_old.data = data_new.data
 
 

compressed_tensors/compressors/pack_quantized.py

@@ -87,7 +87,11 @@ class PackedQuantizationCompressor(Compressor):
                         )
                         value = pack_4bit_ints(value.cpu())
                     compressed_dict[merge_names(prefix, "weight_shape")] = shape
-
+            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
@@ -115,14 +119,20 @@ class PackedQuantizationCompressor(Compressor):
                 with safe_open(safe_path, framework="pt", device=device) as f:
                     weight_data[param_name] = f.get_tensor(full_name)
 
-            if len(weight_data) == len(self.COMPRESSION_PARAM_NAMES):
+            if "weight_scale" in weight_data:
+                zero_point = weight_data.get("weight_zero_point", None)
+                scale = weight_data["weight_scale"]
+                if zero_point is None:
+                    # zero_point assumed to be 0 if not included in state_dict
+                    zero_point = torch.zeros_like(scale)
+
                 weight = weight_data["weight"]
                 original_shape = torch.Size(weight_data["weight_shape"])
                 unpacked = unpack_4bit_ints(weight, original_shape)
                 decompressed = dequantize(
                     x_q=unpacked,
-                    scale=weight_data["weight_scale"],
-                    zero_point=weight_data["weight_zero_point"],
+                    scale=scale,
+                    zero_point=zero_point,
                 )
                 yield merge_names(weight_name, "weight"), decompressed
 

compressed_tensors/quantization/lifecycle/apply.py

@@ -16,6 +16,7 @@ import re
 from collections import OrderedDict
 from typing import Dict, Iterable, Optional
 
+import torch
 from compressed_tensors.quantization.lifecycle.calibration import (
     set_module_for_calibration,
 )
@@ -193,7 +194,13 @@ def _load_quant_args_from_state_dict(
     zp_name = f"{base_name}_zero_point"
     device = next(module.parameters()).device
 
-    scale = getattr(module, scale_name)
-    zp = getattr(module, zp_name)
-    scale.data = state_dict[f"{module_name}.{scale_name}"].to(device)
-    zp.data = state_dict[f"{module_name}.{zp_name}"].to(device)
+    scale = getattr(module, scale_name, None)
+    zp = getattr(module, zp_name, None)
+    if scale is not None:
+        scale.data = state_dict[f"{module_name}.{scale_name}"].to(device)
+    if zp is not None:
+        zp_from_state = state_dict.get(f"{module_name}.{zp_name}", None)
+        if zp_from_state is not None:  # load the non-zero zero points
+            zp.data = state_dict[f"{module_name}.{zp_name}"].to(device)
+        else:  # fill with zeros matching scale shape
+            zp.data = torch.zeros_like(scale, dtype=torch.int8).to(device)

compressed_tensors/quantization/lifecycle/initialize.py

@@ -90,7 +90,9 @@ def _initialize_scale_zero_point_observer(
     device = next(module.parameters()).device
 
     # initializes empty scale and zero point parameters for the module
-    init_scale = Parameter(torch.empty(0, device=device), requires_grad=False)
+    init_scale = Parameter(
+        torch.empty(0, dtype=torch.float16, device=device), requires_grad=False
+    )
     module.register_parameter(f"{base_name}_scale", init_scale)
 
     init_zero_point = Parameter(

compressed_tensors/quantization/quant_scheme.py

@@ -22,6 +22,7 @@ from pydantic import BaseModel
 __all__ = [
     "QuantizationScheme",
     "preset_name_to_scheme",
+    "is_preset_scheme",
 ]
 
 
@@ -98,6 +99,14 @@ def preset_name_to_scheme(name: str, targets: List[str]) -> QuantizationScheme:
     )
 
 
+def is_preset_scheme(name: str) -> bool:
+    """
+    :param name: preset quantization settings name
+    :return: True if the name is a preset scheme name
+    """
+    return name.upper() in PRESET_SCHEMES
+
+
 W8A8 = dict(
     weights=QuantizationArgs(), input_activations=QuantizationArgs(symmetric=False)
 )

{compressed_tensors_nightly-0.3.3.20240523.dist-info → compressed_tensors_nightly-0.3.3.20240524.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.3.3.20240523
+Version: 0.3.3.20240524
 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-0.3.3.20240523.dist-info → compressed_tensors_nightly-0.3.3.20240524.dist-info}/RECORD

@@ -5,9 +5,9 @@ compressed_tensors/compressors/__init__.py,sha256=3yyoNICHll3F4HS6Yu-cgNZpDhfuob
 compressed_tensors/compressors/base.py,sha256=LWEgbpgTxzmoqQ7Xhq2OQszUgWoDtFuGCiV1Y8nlBGw,2134
 compressed_tensors/compressors/dense.py,sha256=G_XHbvuENyupIKlXSITOQgvPkNkcMEOLcLWQr70V9EE,1257
 compressed_tensors/compressors/helpers.py,sha256=k9avlkmeYj6vkOAvl-MgcixtP7ib24SCfhzZ-RusXfw,5403
-compressed_tensors/compressors/int_quantized.py,sha256=I0FqnjtwCiJvQxi9YyfA8aBeaR5csqtq1bOrVvRqJ1I,4744
-compressed_tensors/compressors/model_compressor.py,sha256=teohd0xTbcIDIuEfZrH-bZyAzHn2UZH2KJXT-7Gk3sw,10426
-compressed_tensors/compressors/pack_quantized.py,sha256=K03l8kFqejpapgcMU5hMm1-JIX1cUVvU-VybGSN6RWA,7885
+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/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
@@ -16,14 +16,14 @@ compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5y
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=A6b2V8lhsM8Ho8RjlPBQdxRUDNWhqq-ie5E3RR2_GNg,4360
 compressed_tensors/quantization/quant_config.py,sha256=3BcbQ8-Ah7LbTDSSkRu29Yiid33xo0C1ki6NVhxLiaY,8727
-compressed_tensors/quantization/quant_scheme.py,sha256=QwZsCo8QR9ISB_d58WhIngk2gsMM8ooX-LcRPR-JDRw,3341
+compressed_tensors/quantization/quant_scheme.py,sha256=-hAK1-C67_wJl10eaVLUvbslPBTV04WyzL_J-u9f1ck,3571
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=ggRGWRqhCxCaTTDWRcgTVX3axnS2xV6rc5YvdzK7fSg,798
-compressed_tensors/quantization/lifecycle/apply.py,sha256=whKfNGC_EZm0BC23AP7qWfjRe5OJVWmcZOpX7lryZZc,7625
+compressed_tensors/quantization/lifecycle/apply.py,sha256=yLTDT1zkJp1Nti-aKZGOMW8-TELanF8dXiqDvAkVUQo,7984
 compressed_tensors/quantization/lifecycle/calibration.py,sha256=mLns4jlaWmBwOW8Jtlm5bMX-JET1AiZYUBO7qa-XuxI,1776
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=VreB10xPwgSLQQlTu20UCrFpRS--cA7-lx5s7nrPPrg,2247
 compressed_tensors/quantization/lifecycle/forward.py,sha256=x9JaIX3TK7cb_-0aCOTTYtA4At9l6v5YOY_70GzIeFU,10520
 compressed_tensors/quantization/lifecycle/frozen.py,sha256=h1XYt89MouBTf3jTYLG_6OdFxIu5q2N8tPjsy6J4E6Y,1726
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=U6g9qifSF6pagQZQZEwd-rwWC6uQ_dZXn1wg6nr1Abg,3697
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=pFfcu-pxdQKzlnn-18-RlkEktt2yDi6woNXJsiv1A2c,3732
 compressed_tensors/quantization/observers/__init__.py,sha256=DNH31NQYrIBBcmHsMyFA6whh4pbRsLwuNa6L8AeXaGc,745
 compressed_tensors/quantization/observers/base.py,sha256=kywLVwycFvGxuZMU2cy8-KYyNrZCHkinN6YzCL7boLE,5121
 compressed_tensors/quantization/observers/helpers.py,sha256=JwALNfBYY9Eyl8Q180t0lGh8szumQj8TygfNl-isErs,2166
@@ -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-0.3.3.20240523.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.3.3.20240523.dist-info/METADATA,sha256=_c67GXEm0cMZ_AGWhcLqsMZ3hSbFB4KdQ3lL9Dg7M8M,5633
-compressed_tensors_nightly-0.3.3.20240523.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-compressed_tensors_nightly-0.3.3.20240523.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.3.3.20240523.dist-info/RECORD,,
+compressed_tensors_nightly-0.3.3.20240524.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.3.3.20240524.dist-info/METADATA,sha256=UKjwrUdq2hJGGcA2_ZGO0us811fMjVun9scVSPVXxTI,5633
+compressed_tensors_nightly-0.3.3.20240524.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+compressed_tensors_nightly-0.3.3.20240524.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.3.3.20240524.dist-info/RECORD,,