compressed-tensors-nightly 0.9.1.20250204__tar.gz → 0.9.1.20250205__tar.gz

{compressed-tensors-nightly-0.9.1.20250204/src/compressed_tensors_nightly.egg-info → compressed_tensors_nightly-0.9.1.20250205}/PKG-INFO

@@ -1,15 +1,34 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: compressed-tensors-nightly
-Version: 0.9.1.20250204
+Version: 0.9.1.20250205
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
 Author-email: support@neuralmagic.com
 License: Apache 2.0
 Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=1.7.0
+Requires-Dist: transformers
+Requires-Dist: pydantic>=2.0
 Provides-Extra: dev
+Requires-Dist: black==22.12.0; extra == "dev"
+Requires-Dist: isort==5.8.0; extra == "dev"
+Requires-Dist: wheel>=0.36.2; extra == "dev"
+Requires-Dist: flake8>=3.8.3; extra == "dev"
+Requires-Dist: pytest>=6.0.0; extra == "dev"
+Requires-Dist: nbconvert>=7.16.3; extra == "dev"
 Provides-Extra: accelerate
-License-File: LICENSE
+Requires-Dist: accelerate; extra == "accelerate"
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: license
+Dynamic: provides-extra
+Dynamic: requires-dist
+Dynamic: summary
 
 # compressed-tensors
 

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/base.py

@@ -77,6 +77,15 @@ class BaseCompressor(RegistryMixin, ABC):
         """
         raise NotImplementedError()
 
+    @property
+    @abstractmethod
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        raise NotImplementedError()
+
     @abstractmethod
     def compress(
         self,

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/quantized_compressors/base.py

@@ -144,7 +144,7 @@ class BaseQuantizationCompressor(BaseCompressor):
 
     def _decompress_from_path(self, path_to_model, names_to_scheme, device):
         weight_mappings = get_nested_weight_mappings(
-            path_to_model, self.COMPRESSION_PARAM_NAMES
+            path_to_model, self.compression_param_names
         )
         for weight_name in weight_mappings.keys():
             weight_data = {}
@@ -161,7 +161,7 @@ class BaseQuantizationCompressor(BaseCompressor):
 
     def _decompress_from_state_dict(self, state_dict, names_to_scheme):
         weight_mappings = get_nested_mappings_from_state_dict(
-            state_dict, self.COMPRESSION_PARAM_NAMES
+            state_dict, self.compression_param_names
         )
         for weight_name in weight_mappings.keys():
             weight_data = {}

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/quantized_compressors/naive_quantized.py

@@ -41,12 +41,18 @@ class NaiveQuantizationCompressor(BaseQuantizationCompressor):
     type to the type specified by the layer's QuantizationArgs.
     """
 
-    COMPRESSION_PARAM_NAMES = [
-        "weight",
-        "weight_scale",
-        "weight_zero_point",
-        "weight_g_idx",
-    ]
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return (
+            "weight",
+            "weight_scale",
+            "weight_zero_point",
+            "weight_g_idx",
+        )
 
     def compression_param_info(
         self,

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/quantized_compressors/pack_quantized.py

@@ -36,13 +36,19 @@ class PackedQuantizationCompressor(BaseQuantizationCompressor):
     Compresses a quantized model by packing every eight 4-bit weights into an int32
     """
 
-    COMPRESSION_PARAM_NAMES = [
-        "weight_packed",
-        "weight_scale",
-        "weight_zero_point",
-        "weight_g_idx",
-        "weight_shape",
-    ]
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return (
+            "weight_packed",
+            "weight_scale",
+            "weight_zero_point",
+            "weight_g_idx",
+            "weight_shape",
+        )
 
     def compression_param_info(
         self,

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/sparse_compressors/base.py

@@ -30,8 +30,7 @@ _LOGGER: logging.Logger = logging.getLogger(__name__)
 class BaseSparseCompressor(BaseCompressor):
     """
     Base class representing a sparse compression algorithm. Each child class should
-    implement compression_param_info, compress_weight and decompress_weight; child
-    classes should also define COMPRESSION_PARAM_NAMES.
+    implement compression_param_names, compress_weight and decompress_weight;
 
     Compressors support compressing/decompressing a full module state dict or a single
     quantized PyTorch leaf module.
@@ -113,7 +112,7 @@ class BaseSparseCompressor(BaseCompressor):
         """
         weight_mappings, ignored_params = get_nested_weight_mappings(
             path_to_model_or_tensors,
-            self.COMPRESSION_PARAM_NAMES,
+            self.compression_param_names,
             return_unmatched_params=True,
         )
         for weight_name in weight_mappings.keys():

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/sparse_compressors/dense.py

@@ -25,6 +25,14 @@ class DenseCompressor(BaseCompressor):
     Identity compressor for dense models, returns the original state_dict
     """
 
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return ()
+
     def compress(self, model_state: Dict[str, Tensor], **kwargs) -> Dict[str, Tensor]:
         return model_state
 

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

@@ -40,11 +40,17 @@ class Sparse24BitMaskCompressor(BaseSparseCompressor):
     values tensor, with their locations stored in a 2d bitmask
     """
 
-    COMPRESSION_PARAM_NAMES = [
-        "shape",
-        "compressed",
-        "bitmask",
-    ]
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return (
+            "shape",
+            "compressed",
+            "bitmask",
+        )
 
     def compress_weight(self, name, value):
         bitmask_tensor = Sparse24BitMaskTensor.from_dense(

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py

@@ -38,7 +38,13 @@ class BitmaskCompressor(BaseSparseCompressor):
     values tensor, with their locations stored in a 2d bitmask
     """
 
-    COMPRESSION_PARAM_NAMES = ["shape", "compressed", "bitmask", "row_offsets"]
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return ("shape", "compressed", "bitmask", "row_offsets")
 
     def compress_weight(self, name, value):
         bitmask_tensor = BitmaskTensor.from_dense(value)

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py

@@ -42,8 +42,6 @@ class Marlin24Compressor(BaseCompressor):
     Marlin24 kernel. Decompression is not implemented for this compressor.
     """
 
-    COMPRESSION_PARAM_NAMES = ["weight_packed", "scale_packed", "meta"]
-
     @staticmethod
     def validate_quant_compatability(
         model_quant_args: Dict[str, QuantizationArgs]
@@ -105,6 +103,14 @@ class Marlin24Compressor(BaseCompressor):
 
         return True
 
+    @property
+    def compression_param_names(self) -> Tuple[str]:
+        """
+        Returns a tuple of compression parameter names introduced by
+        the compressor during compression
+        """
+        return ("weight_packed", "scale_packed", "meta")
+
     def compress(
         self,
         model_state: Dict[str, Tensor],

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors/utils/safetensors_load.py

@@ -16,7 +16,7 @@ import json
 import os
 import re
 import struct
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Dict, Iterable, Optional, Tuple, Union
 
 from safetensors import safe_open
 from torch import Tensor
@@ -180,7 +180,9 @@ def get_weight_mappings(path_to_model_or_tensors: str) -> Dict[str, str]:
 
 
 def get_nested_weight_mappings(
-    model_path: str, params_to_nest: List[str], return_unmatched_params: bool = False
+    model_path: str,
+    params_to_nest: Iterable[str],
+    return_unmatched_params: bool = False,
 ) -> Union[NestedWeightMappingType, Tuple[NestedWeightMappingType, WeightMappingType]]:
     """
     Takes a path to a state dict saved in safetensors format and returns a nested
@@ -211,7 +213,7 @@ def get_nested_weight_mappings(
 
     :param model_path: Path to the safetensors state dict, must contain either a
         single safetensors file or multiple files with an index.
-    :param params_to_nest: List of parameter names to nest.
+    :param params_to_nest: Iterable of parameter names to nest.
     :param return_unmatched_params: If True, return a second dictionary containing
         the remaining parameters that were not matched to the params_to_nest.
     :return:
@@ -247,7 +249,7 @@ def get_nested_weight_mappings(
 
 
 def get_nested_mappings_from_state_dict(
-    state_dict, params_to_nest
+    state_dict, params_to_nest: Iterable[str]
 ) -> NestedWeightMappingType:
     """
     Takes a state dict and returns a nested mapping from uncompressed
@@ -262,7 +264,7 @@ def get_nested_mappings_from_state_dict(
     }
 
     :param state_dict: state dict of the model
-    :param params_to_nest: List of parameter names to nest.
+    :param params_to_nest: Iterable of parameter names to nest.
     :return: Nested mapping of parameterized layer names to the value of
         each layer's compression parameters.
     """

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205/src/compressed_tensors_nightly.egg-info}/PKG-INFO

@@ -1,15 +1,34 @@
-Metadata-Version: 2.1
+Metadata-Version: 2.2
 Name: compressed-tensors-nightly
-Version: 0.9.1.20250204
+Version: 0.9.1.20250205
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
 Author-email: support@neuralmagic.com
 License: Apache 2.0
 Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=1.7.0
+Requires-Dist: transformers
+Requires-Dist: pydantic>=2.0
 Provides-Extra: dev
+Requires-Dist: black==22.12.0; extra == "dev"
+Requires-Dist: isort==5.8.0; extra == "dev"
+Requires-Dist: wheel>=0.36.2; extra == "dev"
+Requires-Dist: flake8>=3.8.3; extra == "dev"
+Requires-Dist: pytest>=6.0.0; extra == "dev"
+Requires-Dist: nbconvert>=7.16.3; extra == "dev"
 Provides-Extra: accelerate
-License-File: LICENSE
+Requires-Dist: accelerate; extra == "accelerate"
+Dynamic: author
+Dynamic: author-email
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: license
+Dynamic: provides-extra
+Dynamic: requires-dist
+Dynamic: summary
 
 # compressed-tensors
 

{compressed-tensors-nightly-0.9.1.20250204 → compressed_tensors_nightly-0.9.1.20250205}/src/compressed_tensors_nightly.egg-info/SOURCES.txt

@@ -54,4 +54,6 @@ src/compressed_tensors_nightly.egg-info/PKG-INFO
 src/compressed_tensors_nightly.egg-info/SOURCES.txt
 src/compressed_tensors_nightly.egg-info/dependency_links.txt
 src/compressed_tensors_nightly.egg-info/requires.txt
-src/compressed_tensors_nightly.egg-info/top_level.txt
+src/compressed_tensors_nightly.egg-info/top_level.txt
+tests/test_registry.py
+tests/testing_utils.py

compressed_tensors_nightly-0.9.1.20250205/tests/test_registry.py

@@ -0,0 +1,53 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import pytest
+from compressed_tensors import (
+    BaseCompressor,
+    BitmaskCompressor,
+    BitmaskConfig,
+    CompressionFormat,
+    DenseCompressor,
+    DenseSparsityConfig,
+    SparsityCompressionConfig,
+)
+
+
+@pytest.mark.parametrize(
+    "name,type",
+    [
+        [CompressionFormat.sparse_bitmask.value, BitmaskConfig],
+        [CompressionFormat.dense.value, DenseSparsityConfig],
+    ],
+)
+def test_configs(name, type):
+    config = SparsityCompressionConfig.load_from_registry(name)
+    assert isinstance(config, type)
+    assert config.format == name
+
+
+@pytest.mark.parametrize(
+    "name,type",
+    [
+        [CompressionFormat.sparse_bitmask.value, BitmaskCompressor],
+        [CompressionFormat.dense.value, DenseCompressor],
+    ],
+)
+def test_compressors(name, type):
+    compressor = BaseCompressor.load_from_registry(
+        name, config=SparsityCompressionConfig(format="none")
+    )
+    assert isinstance(compressor, type)
+    assert isinstance(compressor.config, SparsityCompressionConfig)
+    assert compressor.config.format == "none"

compressed_tensors_nightly-0.9.1.20250205/tests/testing_utils.py

@@ -0,0 +1,144 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# flake8: noqa
+import unittest
+
+import pytest
+
+
+def compressed_tensors_config_available():
+    try:
+        from transformers.utils.quantization_config import (  # noqa: F401
+            CompressedTensorsConfig,
+        )
+
+        return True
+    except ImportError:
+        return False
+
+
+def accelerate_availabe():
+    try:
+        import accelerate  # noqa: F401
+
+        return True
+
+    except ImportError:
+        return False
+
+
+_is_compressed_tensors_config_available = compressed_tensors_config_available()
+_is_accelerate_available = accelerate_availabe()
+
+
+def requires_hf_quantizer():
+    return pytest.mark.skipif(
+        not _is_compressed_tensors_config_available,
+        reason="requires transformers>=4.45 to support CompressedTensorsHfQuantizer",
+    )
+
+
+def requires_accelerate():
+    return pytest.mark.skipif(
+        not _is_accelerate_available,
+        reason="requires accelerate",
+    )
+
+
+def get_random_mat(M, K, dtype) -> "torch.Tensor":
+    """
+    :param M: number of rows
+    :param K: number of columns
+    :param dtype: data type of the matrix
+    :return: random matrix of shape (M, K) with non-zero values
+    """
+    import torch
+    from compressed_tensors.quantization import FP8_DTYPE
+
+    rand_tensor_dtype = dtype
+    if dtype in [torch.int8, FP8_DTYPE]:
+        rand_tensor_dtype = torch.float16
+    mat = torch.rand(M, K, dtype=rand_tensor_dtype).cuda()
+    mat = mat.masked_fill_(mat == 0, 1)
+    return mat.to(dtype)
+
+
+def generate_pruned_semi_structured_mat(M, K, dtype) -> "torch.Tensor":
+    """
+    :param M: number of rows
+    :param K: number of columns
+    :param dtype: data type of the matrix
+    :return: random matrix of shape (M, K) with 2:4 sparsity pattern
+    """
+    import torch
+    from compressed_tensors.quantization import FP8_DTYPE
+
+    mask = torch.Tensor([0, 0, 1, 1]).tile((M, K // 4)).bool()
+    rand_tensor_dtype = dtype
+    if dtype in [torch.int8, FP8_DTYPE]:
+        rand_tensor_dtype = torch.float16
+    mat = torch.rand(M, K, dtype=rand_tensor_dtype)
+    mat = mat.masked_fill_(mat == 0, 1)
+    if dtype == FP8_DTYPE:
+        # some float8_e4m3fn operations are not supported on CPU
+        mat = mat.cuda()
+        mask = mask.cuda()
+    mat = mat * mask
+    return mat.to(dtype)
+
+
+def induce_sparsity(tensor, sparsity_ratio) -> "torch.Tensor":
+    """
+    Makes a tensor sparse by zeroing out a given fraction
+    of its smallest absolute values.
+
+    :param: weight_tensor (torch.Tensor): The input weight tensor.
+    :param: sparsity_ratio (float): Fraction of weights to be zeroed
+        (0 <= sparsity_ratio <= 1).
+    :returns: torch.Tensor: Sparse version of the input tensor.
+    """
+    import torch
+
+    if not (0 <= sparsity_ratio <= 1):
+        raise ValueError("Sparsity ratio must be between 0 and 1.")
+
+    # Flatten the tensor and compute the threshold for sparsity
+    flattened = tensor.view(-1)
+    k = int(sparsity_ratio * flattened.numel())
+
+    if k > 0:
+        threshold = torch.topk(flattened.abs(), k, largest=False).values.max()
+        sparse_tensor = torch.where(
+            tensor.abs() > threshold, tensor, torch.zeros_like(tensor)
+        )
+    else:
+        sparse_tensor = tensor
+
+    return sparse_tensor
+
+
+def is_gpu_available():
+    """
+    :return: True if a GPU is available, False otherwise
+    """
+    try:
+        import torch  # noqa: F401
+
+        return torch.cuda.device_count() > 0
+    except ImportError:
+        return False
+
+
+def requires_gpu(test_case):
+    return unittest.skipUnless(is_gpu_available(), "test requires GPU")(test_case)