compressed-tensors 0.9.2__py3-none-any.whl → 0.9.3__py3-none-any.whl

compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -19,7 +19,7 @@ import os
 import re
 from contextlib import contextmanager
 from copy import deepcopy
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Set, TypeVar, Union
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Set, Tuple, TypeVar, Union
 
 import compressed_tensors
 import torch
@@ -522,10 +522,13 @@ class ModelCompressor:
                 update_parameter_data(module, data, param_name)
 
 
-def map_modules_to_quant_args(model: Module) -> Dict[str, QuantizationArgs]:
+def map_modules_to_quant_args(
+    model: Module,
+) -> Dict[str, Union[QuantizationArgs, Tuple[QuantizationArgs, QuantizationArgs]]]:
     """
     Given a pytorch model, map out the submodule name (usually linear layers)
-     to the QuantizationArgs
+    to the weight QuantizationArgs. If running input activation quantization, will also
+    map to the input QuantizationArgs in a tuple.
 
     :param model: pytorch model
     """
@@ -535,6 +538,12 @@ def map_modules_to_quant_args(model: Module) -> Dict[str, QuantizationArgs]:
             if submodule.quantization_scheme.weights is not None:
                 name = fix_fsdp_module_name(name)
                 quantized_modules_to_args[name] = submodule.quantization_scheme.weights
+                if submodule.quantization_scheme.input_activations is not None:
+                    weight_args = quantized_modules_to_args.get(name)
+                    quantized_modules_to_args[name] = (
+                        weight_args,
+                        submodule.quantization_scheme.input_activations,
+                    )
 
     return quantized_modules_to_args
 

compressed_tensors/compressors/quantized_compressors/base.py

@@ -82,11 +82,32 @@ class BaseQuantizationCompressor(BaseCompressor):
         """
         compressed_dict = {}
         weight_suffix = ".weight"
+        input_zp_suffix = ".input_zero_point"
+        weight_zp_suffix = ".weight_zero_point"
         _LOGGER.debug(
             f"Compressing model with {len(model_state)} parameterized layers..."
         )
 
         for name, value in tqdm(model_state.items(), desc="Quantized Compression"):
+            # check if the parameter we're compressing is the weight zp
+            # or the input zp
+            is_weight_zp = name.endswith(weight_zp_suffix)
+            is_input_zp = name.endswith(input_zp_suffix)
+
+            # if we're saving the weight zp, fetch weight quant args
+            if is_weight_zp:
+                quant_args_zp = names_to_scheme.get(name[: -(len(weight_zp_suffix))])
+                if isinstance(quant_args_zp, tuple):
+                    # If tuple, first value is weight args, second is input args
+                    quant_args_zp = quant_args_zp[0]
+
+            # if we're saving the input zp, fetch input quant args
+            if is_input_zp:
+                input_args_zp = names_to_scheme.get(name[: -(len(input_zp_suffix))])
+                if isinstance(input_args_zp, tuple):
+                    # If tuple, first value is weight args, second is input args
+                    input_args_zp = input_args_zp[-1]
+
             if name.endswith(weight_suffix):
                 prefix = name[: -(len(weight_suffix))]
                 scale = model_state.get(merge_names(prefix, "weight_scale"), None)
@@ -94,7 +115,11 @@ class BaseQuantizationCompressor(BaseCompressor):
                 g_idx = model_state.get(merge_names(prefix, "weight_g_idx"), None)
                 if scale is not None:
                     # weight is quantized, compress it
-                    quant_args = names_to_scheme[prefix]
+                    if isinstance(names_to_scheme[prefix], tuple):
+                        quant_args = names_to_scheme[prefix][0]
+                    else:
+                        quant_args = names_to_scheme[prefix]
+
                     compressed_data = self.compress_weight(
                         weight=value,
                         scale=scale,
@@ -107,7 +132,11 @@ class BaseQuantizationCompressor(BaseCompressor):
                         compressed_dict[merge_names(prefix, key)] = value
                 else:
                     compressed_dict[name] = value.to("cpu")
-            elif name.endswith("zero_point") and torch.all(value == 0):
+            # only save if asym
+            elif is_weight_zp and quant_args_zp.symmetric:
+                continue
+            # only save if asym
+            elif is_input_zp and input_args_zp.symmetric:
                 continue
             elif name.endswith("g_idx") and torch.any(value <= -1):
                 continue

compressed_tensors/linear/compressed_linear.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import warnings
 from typing import Dict, Tuple
 
 import torch
@@ -21,6 +22,7 @@ from compressed_tensors.quantization import (
     QuantizationStatus,
     initialize_module_for_quantization,
 )
+from compressed_tensors.utils import register_offload_parameter
 from torch import Tensor
 from torch.nn import Parameter
 from torch.nn.functional import linear
@@ -32,11 +34,16 @@ class CompressedLinear(Linear):
     Wrapper module for running a compressed forward pass of a quantized Linear module.
     The wrapped layer will decompressed on each forward call.
 
-    :param module: dense linear module to replace
-    :param quantization_scheme: quantization config for the module to wrap
-    :param quantization_format: compression format module is stored as
     """
 
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+        warnings.warn(
+            "CompressedLinear should not be initialized directly. "
+            "Use the from_linear method instead.",
+            UserWarning,
+        )
+
     @classmethod
     @torch.no_grad()
     def from_linear(
@@ -45,6 +52,12 @@ class CompressedLinear(Linear):
         quantization_scheme: QuantizationScheme,
         quantization_format: str,
     ):
+        """
+        :param module: dense linear module to replace
+        :param quantization_scheme: quantization config for the module to wrap
+        :param quantization_format: compression format module is stored as
+        :return: CompressedLinear module wrapping the input module
+        """
         module.__class__ = CompressedLinear
         module.compressor = BaseCompressor.load_from_registry(quantization_format)
         device = next(module.parameters()).device
@@ -68,7 +81,7 @@ class CompressedLinear(Linear):
             param = Parameter(
                 torch.empty(shape, device=device, dtype=dtype), requires_grad=False
             )
-            module.register_parameter(name, param)
+            register_offload_parameter(module, name, param)
 
         # mark module as compressed
         module.quantization_status = QuantizationStatus.COMPRESSED
@@ -85,5 +98,11 @@ class CompressedLinear(Linear):
         """
         Decompresses the weight, then runs the wrapped forward pass
         """
-        uncompressed_weight = self.compressor.decompress_module(self)
-        return linear(input, uncompressed_weight, self.bias)
+        if self.quantization_status == QuantizationStatus.COMPRESSED:
+            weight_data = self.compressor.decompress_module(self)
+            param = Parameter(weight_data, requires_grad=False)
+            register_offload_parameter(self, "weight", param)
+
+            self.quantization_status = QuantizationStatus.FROZEN
+
+        return linear(input, self.weight, self.bias)

compressed_tensors/quantization/lifecycle/initialize.py

@@ -203,11 +203,10 @@ def _initialize_attn_scales(module: Module) -> None:
         torch.empty(expected_shape, dtype=scale_dtype, device=device),
         requires_grad=False,
     )
-
-    module.register_parameter(KVCacheScaleType.KEY.value, init_scale)
+    register_offload_parameter(module, KVCacheScaleType.KEY.value, init_scale)
 
     init_scale = Parameter(
         torch.empty(expected_shape, dtype=scale_dtype, device=device),
         requires_grad=False,
     )
-    module.register_parameter(KVCacheScaleType.VALUE.value, init_scale)
+    register_offload_parameter(module, KVCacheScaleType.VALUE.value, init_scale)

compressed_tensors/version.py

@@ -17,7 +17,7 @@ Functionality for storing and setting the version info for SparseML
 """
 
 
-version_base = "0.9.2"
+version_base = "0.9.3"
 is_release = True  # change to True to set the generated version as a release version
 
 

{compressed_tensors-0.9.2.dist-info → compressed_tensors-0.9.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.2
+Version: 0.9.3
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
@@ -26,6 +26,7 @@ Dynamic: description
 Dynamic: description-content-type
 Dynamic: home-page
 Dynamic: license
+Dynamic: license-file
 Dynamic: provides-extra
 Dynamic: requires-dist
 Dynamic: summary

{compressed_tensors-0.9.2.dist-info → compressed_tensors-0.9.3.dist-info}/RECORD

@@ -1,13 +1,13 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=BaFkY2H2ed1_O6ZxJgT-GUlT7oI0xJIeXGPO8yXqBE0,1585
+compressed_tensors/version.py,sha256=X4y5lqlF1QFUgl25iumzagpg3dzyVoLP6i82HZEhCJA,1585
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=x8dQrWVEurynXw03yHJZTaAmrRTOsdZJoHjmvs0IKwk,7002
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
 compressed_tensors/compressors/model_compressors/__init__.py,sha256=5RGGPFu4YqEt_aOdFSQYFYFDjcZFJN0CsMqRtDZz3Js,666
-compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=AmIE1SoNRH1fNgQALfNkdQo8y5tePVpdWUgLIOtf5rg,22569
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=n0gcrKwefJuO6b4LNjCynJQf7NNqNHDcoLlzZgTCPGc,23080
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=09UJq68Pht6Bf-4iP9xYl3tetKsncNPHD8IAGbePsr4,714
-compressed_tensors/compressors/quantized_compressors/base.py,sha256=cp8S1Kr3HhlMHIz7k4vGo-qxxdknEC3qP1QLIhNnwRA,7217
+compressed_tensors/compressors/quantized_compressors/base.py,sha256=GXTSWgFAhksbno94Ulpth9-YM4a7NsDlx4oQGGB0swQ,8567
 compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=fd0KlkSx6bvZ3xwIkK3jEUdPSUPs56Eua4dEDOtzKW0,5150
 compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=zH2PocRe_T5yt1-3kLdZH9AUQWQyaVOi4U9nEJiYaWA,8509
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
@@ -23,7 +23,7 @@ compressed_tensors/config/dense.py,sha256=NgSxnFCnckU9-iunxEaqiFwqgdO7YYxlWKR74j
 compressed_tensors/config/sparse_24_bitmask.py,sha256=Lhj39zT2V1hxftprvxvneyhv45ShlXOKd75DBbDTyTE,1401
 compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5ynVAUeiiYpS1Gt8,1308
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
-compressed_tensors/linear/compressed_linear.py,sha256=MJa-UfoKhIkdUWRD1shrXXri2cOwR5GK0a4t4bNYosM,3268
+compressed_tensors/linear/compressed_linear.py,sha256=_m6XpNcI53eeSHO8VdiuAM6UBTdpDhn5Ivd8iRMwEKc,3980
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=sKpb8DcNObidjXjNol1Tn_Iih3ZXBycSp-fyz68TGhY,9117
 compressed_tensors/quantization/quant_config.py,sha256=vx06wBo91p4LCb3Vzd-2eCTUeIf_Sz2ZXRP263eQyjQ,10385
@@ -33,7 +33,7 @@ compressed_tensors/quantization/lifecycle/apply.py,sha256=lZmCCSm1_o79iUAy460w6B
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=DOWouUqfaLA4Qhg-ojVVBdhhSAlgZqFC26vZARxE0ko,12961
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=hymYtayTSumm8KCYAYPY267aWmlsJpt8oQFiRblk8qE,7452
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=sK3PLm69N91QepBuq-83Qd2Br6XcOmRDpD5qo_WWNJo,7469
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
 compressed_tensors/quantization/utils/helpers.py,sha256=DBP-sGRpGAY01K0LFE7qqonNj4hkTYL_mXrMs2LtAD8,14100
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
@@ -45,8 +45,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=5SeM2hzLh77Ne8Vk7qR6-km7cf8bhov41ExpWITqX3A,11470
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.2.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.2.dist-info/METADATA,sha256=BmuThcrHjHQVScbgaWfyCRGuLo-cB6OnujZJPOKe1bQ,6975
-compressed_tensors-0.9.2.dist-info/WHEEL,sha256=In9FTNxeP60KnTkGw7wk6mJPYd_dQSjEZmXdBdMCI-8,91
-compressed_tensors-0.9.2.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.2.dist-info/RECORD,,
+compressed_tensors-0.9.3.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.3.dist-info/METADATA,sha256=zs3aFaG-BGV9hqJbW9Zwzex0TVcM5sPZhiaeVx2qjR0,6997
+compressed_tensors-0.9.3.dist-info/WHEEL,sha256=CmyFI0kx5cdEMTLiONQRbGQwjIoR1aIYB7eCAQ4KPJ0,91
+compressed_tensors-0.9.3.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.3.dist-info/RECORD,,

{compressed_tensors-0.9.2.dist-info → compressed_tensors-0.9.3.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.8.0)
+Generator: setuptools (78.1.0)
 Root-Is-Purelib: true
 Tag: py3-none-any