compressed-tensors 0.3.1__tar.gz → 0.3.2__tar.gz

{compressed-tensors-0.3.1/src/compressed_tensors.egg-info → compressed-tensors-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors
-Version: 0.3.1
+Version: 0.3.2
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
@@ -94,4 +94,7 @@ save_compressed_model(model, "compressed_model.safetensors", compression_format=
 state_dict = dict(load_compressed("compressed_model.safetensors", compression_config))
 ```
 
+For more in-depth tutorial on bitmask compression, refer to the [notebook](https://github.com/neuralmagic/compressed-tensors/blob/d707c5b84bc3fef164aebdcd97cb6eaa571982f8/examples/bitmask_compression.ipynb).
+
+
 

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/README.md

@@ -80,3 +80,6 @@ save_compressed_model(model, "compressed_model.safetensors", compression_format=
 # load compressed model weights (`dict` turns generator into a dictionary)
 state_dict = dict(load_compressed("compressed_model.safetensors", compression_config))
 ```
+
+For more in-depth tutorial on bitmask compression, refer to the [notebook](https://github.com/neuralmagic/compressed-tensors/blob/d707c5b84bc3fef164aebdcd97cb6eaa571982f8/examples/bitmask_compression.ipynb).
+

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/setup.py

@@ -32,7 +32,7 @@ def _setup_extras() -> Dict:
 
 setup(
     name="compressed-tensors",
-    version="0.3.1",
+    version="0.3.2",
     author="Neuralmagic, Inc.",
     author_email="support@neuralmagic.com",
     license="Apache 2.0",

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/compressors/__init__.py

@@ -16,10 +16,5 @@
 
 from .base import ModelCompressor
 from .dense import DenseCompressor
-from .helpers import (
-    infer_compressor_from_model_config,
-    load_compressed,
-    save_compressed,
-    save_compressed_model,
-)
+from .helpers import load_compressed, save_compressed, save_compressed_model
 from .sparse_bitmask import BitmaskCompressor, BitmaskTensor

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/compressors/base.py

@@ -22,6 +22,7 @@ from compressed_tensors.utils import get_safetensors_folder
 from torch import Tensor
 from torch.nn import Module, Parameter
 from tqdm import tqdm
+from transformers import AutoConfig
 
 
 __all__ = ["ModelCompressor"]
@@ -34,6 +35,29 @@ class ModelCompressor(RegistryMixin):
     :param config: config specifying compression parameters
     """
 
+    @classmethod
+    def from_pretrained(
+        cls, pretrained_model_name_or_path: str
+    ) -> Optional["ModelCompressor"]:
+        """
+        Given a path to a model config, extract a sparsity config if it exists and
+        return the associated ModelCompressor
+
+        :param pretrained_model_name_or_path: path to model config on disk or HF hub
+        :return: matching compressor if config contains a sparsity config
+        """
+        config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
+        sparsity_config = getattr(config, SPARSITY_CONFIG_NAME, None)
+        if sparsity_config is None:
+            return None
+
+        format = sparsity_config.get("format")
+        sparsity_config = CompressionConfig.load_from_registry(
+            format, **sparsity_config
+        )
+        compressor = cls.load_from_registry(format, config=sparsity_config)
+        return compressor
+
     def __init__(self, config: Optional[CompressionConfig] = None):
         self.config = config
 
@@ -47,7 +71,7 @@ class ModelCompressor(RegistryMixin):
         raise NotImplementedError()
 
     def decompress(
-        self, path_to_model_or_tensors: str
+        self, path_to_model_or_tensors: str, device: str = "cpu"
     ) -> Generator[Tuple[str, Tensor], None, None]:
         """
         Reads a compressed state dict located at path_to_model_or_tensors

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/compressors/dense.py

@@ -29,6 +29,6 @@ class DenseCompressor(ModelCompressor):
         return model_state
 
     def decompress(
-        self, path_to_model_or_tensors: str, device: str
+        self, path_to_model_or_tensors: str, device: str = "cpu"
     ) -> Generator[Tuple[str, Tensor], None, None]:
         return iter([])

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/compressors/helpers.py

@@ -16,45 +16,21 @@ from pathlib import Path
 from typing import Dict, Generator, Optional, Tuple, Union
 
 import torch
-from compressed_tensors.base import SPARSITY_CONFIG_NAME
 from compressed_tensors.compressors import ModelCompressor
 from compressed_tensors.config import CompressionConfig, CompressionFormat
 from compressed_tensors.utils.safetensors_load import get_weight_mappings
 from safetensors import safe_open
 from safetensors.torch import save_file
 from torch import Tensor
-from transformers import AutoConfig
 
 
 __all__ = [
-    "infer_compressor_from_model_config",
     "load_compressed",
     "save_compressed",
     "save_compressed_model",
 ]
 
 
-def infer_compressor_from_model_config(
-    pretrained_model_name_or_path: str,
-) -> Optional[ModelCompressor]:
-    """
-    Given a path to a model config, extract a sparsity config if it exists and return
-    the associated ModelCompressor
-
-    :param pretrained_model_name_or_path: path to model config on disk or HF hub
-    :return: matching compressor if config contains a sparsity config
-    """
-    config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
-    sparsity_config = getattr(config, SPARSITY_CONFIG_NAME, None)
-    if sparsity_config is None:
-        return None
-
-    format = sparsity_config.get("format")
-    sparsity_config = CompressionConfig.load_from_registry(format, **sparsity_config)
-    compressor = ModelCompressor.load_from_registry(format, config=sparsity_config)
-    return compressor
-
-
 def save_compressed(
     tensors: Dict[str, Tensor],
     save_path: Union[str, Path],

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/compressors/sparse_bitmask.py

@@ -75,8 +75,9 @@ class BitmaskCompressor(ModelCompressor):
         self, path_to_model_or_tensors: str, device: str = "cpu"
     ) -> Generator[Tuple[str, Tensor], None, None]:
         """
-        Reads a bitmask compressed state dict located at path_to_model_or_tensors
-        and returns a generator for sequentially decompressing back to a dense state dict
+        Reads a bitmask compressed state dict located
+        at path_to_model_or_tensors and returns a generator
+        for sequentially decompressing back to a dense state dict
 
         :param model_path: path to compressed safetensors model (directory with
             one or more safetensors files) or compressed tensors file

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/quantization/lifecycle/forward.py

@@ -111,7 +111,7 @@ def wrap_module_forward_quantized(module: Module, scheme: QuantizationScheme):
 
 
 def _maybe_calibrate_or_quantize(
-    module: Module, value: Module, base_name: str, args: "QuantizationArgs"
+    module: Module, value: torch.Tensor, base_name: str, args: "QuantizationArgs"
 ) -> torch.Tensor:
     # only run quantized for the included stages
     if module.quantization_status not in {
@@ -120,17 +120,23 @@ def _maybe_calibrate_or_quantize(
     }:
         return value
 
-    device = next(module.parameters()).device
-    scale = getattr(module, f"{base_name}_scale")
-    zero_point = getattr(module, f"{base_name}_zero_point")
-
-    if module.quantization_status == QuantizationStatus.CALIBRATION:
-        # get observer and get new quant params from observation
+    if args.dynamic:
+        # dynamic quantization - get scale and zero point directly from observer
         observer = getattr(module, f"{base_name}_observer")
-        updated_scale, updated_zero_point = observer(value)
-
-        # update scale and zero point
-        scale.data = updated_scale.to(device)
-        zero_point.data = updated_zero_point.to(device)
+        scale, zero_point = observer(value)
+    else:
+        # static quantization - get previous scale and zero point from layer
+        scale = getattr(module, f"{base_name}_scale")
+        zero_point = getattr(module, f"{base_name}_zero_point")
+
+        if module.quantization_status == QuantizationStatus.CALIBRATION:
+            # calibration mode - get new quant params from observer
+            observer = getattr(module, f"{base_name}_observer")
+            updated_scale, updated_zero_point = observer(value)
+
+            # update scale and zero point
+            device = next(module.parameters()).device
+            scale.data = updated_scale.to(device)
+            zero_point.data = updated_zero_point.to(device)
 
     return fake_quantize(value, scale, zero_point, args)

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/quantization/lifecycle/frozen.py

@@ -30,17 +30,17 @@ def freeze_module_quantization(module: Module):
 
     :param module: module to freeze quantization for
     """
-    if not getattr(module, "quantization_scheme", None):
+    scheme = getattr(module, "quantization_scheme", None)
+    if not scheme:
         # no quantization scheme nothing to do
         return
 
-    # delete observers from module
-    observer_names = []
-    for submodule_name, _ in module.named_modules():
-        if "." not in submodule_name and submodule_name.endswith("_observer"):
-            # delete any observers that belong directly to this module
-            observer_names.append(submodule_name)
-    for observer_name in observer_names:
-        delattr(module, observer_name)
+    # delete observers from module if not dynamic
+    if scheme.input_activations and not scheme.input_activations.dynamic:
+        delattr(module, "input_observer")
+    if scheme.weights and not scheme.weights.dynamic:
+        delattr(module, "weight_observer")
+    if scheme.output_activations and not scheme.output_activations.dynamic:
+        delattr(module, "output_observer")
 
     module.quantization_status = QuantizationStatus.FROZEN

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -80,6 +80,13 @@ def initialize_module_for_quantization(
 def _initialize_scale_zero_point_observer(
     module: Module, base_name: str, quantization_args: QuantizationArgs
 ):
+    # initialize observer module and attach as submodule
+    observer = quantization_args.get_observer()
+    module.register_module(f"{base_name}_observer", observer)
+
+    if quantization_args.dynamic:
+        return  # no need to register a scale and zero point for a dynamic observer
+
     device = next(module.parameters()).device
 
     # initializes empty scale and zero point parameters for the module
@@ -90,7 +97,3 @@ def _initialize_scale_zero_point_observer(
         torch.empty(0, device=device, dtype=int), requires_grad=False
     )
     module.register_parameter(f"{base_name}_zero_point", init_zero_point)
-
-    # initialize observer module and attach as submodule
-    observer = quantization_args.get_observer()
-    module.register_module(f"{base_name}_observer", observer)

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/quantization/observers/memoryless.py

@@ -23,10 +23,10 @@ from torch import FloatTensor, IntTensor, Tensor
 __all__ = ["MemorylessObserver"]
 
 
-@Observer.register("memoryless")
+@Observer.register("memoryless", alias=["dynamic"])
 class MemorylessObserver(Observer):
     """
-    Implements a dynamic quantization observer that sets the scale and
+    Implements a quantization observer that sets the scale and
     zero point based on the latest observed value without tracking state
     """
 

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2}/src/compressed_tensors/quantization/quant_args.py

@@ -53,6 +53,11 @@ class QuantizationArgs(BaseModel):
     :param group_size: group length to use for the group strategy
     :param block_structure: 2d block structure to use for the block strategy, must be
     of the format "2x4", "8x16", etc.
+    :param dynamic: set True to perform dynamic quantization - values will not be
+        calibrated during calibration phase, instead during inference new quantization
+        ranges will be observed with every sample. Defaults to False for static
+        quantization. Note that enabling dynamic quantization will change the default
+        observer to a memoryless one
     """
 
     num_bits: int = 8
@@ -61,6 +66,7 @@ class QuantizationArgs(BaseModel):
     strategy: QuantizationStrategy = QuantizationStrategy.TENSOR
     group_size: Optional[int] = None
     block_structure: Optional[str] = None
+    dynamic: bool = False
     observer: str = Field(
         default="minmax",
         description=(
@@ -82,4 +88,9 @@ class QuantizationArgs(BaseModel):
         """
         from compressed_tensors.quantization.observers.base import Observer
 
+        if self.observer == "minmax" and self.dynamic:
+            # override defualt observer for dynamic, you never want minmax which
+            # keeps state across samples for dynamic
+            self.observer = "memoryless"
+
         return Observer.load_from_registry(self.observer, quantization_args=self)

{compressed-tensors-0.3.1 → compressed-tensors-0.3.2/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors
-Version: 0.3.1
+Version: 0.3.2
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
@@ -94,4 +94,7 @@ save_compressed_model(model, "compressed_model.safetensors", compression_format=
 state_dict = dict(load_compressed("compressed_model.safetensors", compression_config))
 ```
 
+For more in-depth tutorial on bitmask compression, refer to the [notebook](https://github.com/neuralmagic/compressed-tensors/blob/d707c5b84bc3fef164aebdcd97cb6eaa571982f8/examples/bitmask_compression.ipynb).
+
+