compressed-tensors 0.3.3__py3-none-any.whl → 0.5.0__py3-none-any.whl

compressed_tensors/base.py

@@ -13,4 +13,6 @@
 # limitations under the License.
 
 SPARSITY_CONFIG_NAME = "sparsity_config"
-QUANTIZATION_CONFIG_NAME = "sparseml_quantization_config"
+QUANTIZATION_CONFIG_NAME = "quantization_config"
+COMPRESSION_CONFIG_NAME = "compression_config"
+KV_CACHE_SCHEME_NAME = "kv_cache_scheme"

compressed_tensors/compressors/__init__.py

@@ -14,7 +14,15 @@
 
 # flake8: noqa
 
-from .base import ModelCompressor
+from .base import Compressor
 from .dense import DenseCompressor
 from .helpers import load_compressed, save_compressed, save_compressed_model
+from .marlin_24 import Marlin24Compressor
+from .model_compressor import ModelCompressor, map_modules_to_quant_args
+from .naive_quantized import (
+    FloatQuantizationCompressor,
+    IntQuantizationCompressor,
+    QuantizationCompressor,
+)
+from .pack_quantized import PackedQuantizationCompressor
 from .sparse_bitmask import BitmaskCompressor, BitmaskTensor

compressed_tensors/compressors/base.py

@@ -12,56 +12,30 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import operator
-from typing import Dict, Generator, Optional, Tuple
+from typing import Dict, Generator, Tuple, Union
 
-from compressed_tensors.base import SPARSITY_CONFIG_NAME
-from compressed_tensors.config import CompressionConfig
+from compressed_tensors.config import SparsityCompressionConfig
+from compressed_tensors.quantization import QuantizationConfig
 from compressed_tensors.registry import RegistryMixin
-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"]
+__all__ = ["Compressor"]
 
 
-class ModelCompressor(RegistryMixin):
+class Compressor(RegistryMixin):
     """
-    Base class representing a model compression algorithm.
+    Base class representing a model compression algorithm
 
     :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):
+    def __init__(
+        self, config: Union[SparsityCompressionConfig, QuantizationConfig, None] = None
+    ):
         self.config = config
 
-    def compress(self, model_state: Dict[str, Tensor]) -> Dict[str, Tensor]:
+    def compress(self, model_state: Dict[str, Tensor], **kwargs) -> Dict[str, Tensor]:
         """
         Compresses a dense state dict
 
@@ -71,7 +45,7 @@ class ModelCompressor(RegistryMixin):
         raise NotImplementedError()
 
     def decompress(
-        self, path_to_model_or_tensors: str, device: str = "cpu"
+        self, path_to_model_or_tensors: str, device: str = "cpu", **kwargs
     ) -> Generator[Tuple[str, Tensor], None, None]:
         """
         Reads a compressed state dict located at path_to_model_or_tensors
@@ -80,24 +54,7 @@ class ModelCompressor(RegistryMixin):
 
         :param model_path: path to compressed safetensors model (directory with
             one or more safetensors files) or compressed tensors file
+        :param device: optional device to load intermediate weights into
         :return: compressed state dict
         """
         raise NotImplementedError()
-
-    def overwrite_weights(self, model_path: str, model: Module):
-        """
-        Overwrites the weights in model with weights decompressed from model_path
-
-        :param model_path: path to compressed weights
-        :param model: pytorch model to load decompressed weights into
-        """
-        model_path = get_safetensors_folder(model_path)
-        dense_gen = self.decompress(model_path)
-        for name, data in tqdm(dense_gen, 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_old.data = data_new.data
-
-        setattr(model, SPARSITY_CONFIG_NAME, self.config)

compressed_tensors/compressors/dense.py

@@ -14,21 +14,21 @@
 
 from typing import Dict, Generator, Tuple
 
-from compressed_tensors.compressors import ModelCompressor
+from compressed_tensors.compressors import Compressor
 from compressed_tensors.config import CompressionFormat
 from torch import Tensor
 
 
-@ModelCompressor.register(name=CompressionFormat.dense_sparsity.value)
-class DenseCompressor(ModelCompressor):
+@Compressor.register(name=CompressionFormat.dense.value)
+class DenseCompressor(Compressor):
     """
     Identity compressor for dense models, returns the original state_dict
     """
 
-    def compress(self, model_state: Dict[str, Tensor]) -> Dict[str, Tensor]:
+    def compress(self, model_state: Dict[str, Tensor], **kwargs) -> Dict[str, Tensor]:
         return model_state
 
     def decompress(
-        self, path_to_model_or_tensors: str, device: str = "cpu"
+        self, path_to_model_or_tensors: str, device: str = "cpu", **kwargs
     ) -> Generator[Tuple[str, Tensor], None, None]:
         return iter([])

compressed_tensors/compressors/helpers.py

@@ -16,8 +16,8 @@ from pathlib import Path
 from typing import Dict, Generator, Optional, Tuple, Union
 
 import torch
-from compressed_tensors.compressors import ModelCompressor
-from compressed_tensors.config import CompressionConfig, CompressionFormat
+from compressed_tensors.compressors import Compressor
+from compressed_tensors.config import CompressionFormat, SparsityCompressionConfig
 from compressed_tensors.utils.safetensors_load import get_weight_mappings
 from safetensors import safe_open
 from safetensors.torch import save_file
@@ -48,20 +48,20 @@ def save_compressed(
     if tensors is None or len(tensors) == 0:
         raise ValueError("No tensors or empty tensors provided to compress")
 
-    # if no compression_format specified, default to `dense_sparsity`
-    compression_format = compression_format or CompressionFormat.dense_sparsity.value
+    # if no compression_format specified, default to `dense`
+    compression_format = compression_format or CompressionFormat.dense.value
 
     if not (
-        compression_format in ModelCompressor.registered_names()
-        or compression_format in ModelCompressor.registered_aliases()
+        compression_format in Compressor.registered_names()
+        or compression_format in Compressor.registered_aliases()
     ):
         raise ValueError(
             f"Unknown compression format: {compression_format}. "
-            f"Must be one of {set(ModelCompressor.registered_names() + ModelCompressor.registered_aliases())}"  # noqa E501
+            f"Must be one of {set(Compressor.registered_names() + Compressor.registered_aliases())}"  # noqa E501
         )
 
     # compress
-    compressor = ModelCompressor.load_from_registry(compression_format)
+    compressor = Compressor.load_from_registry(compression_format)
     # save compressed tensors
     compressed_tensors = compressor.compress(tensors)
     save_file(compressed_tensors, save_path)
@@ -69,7 +69,7 @@ def save_compressed(
 
 def load_compressed(
     compressed_tensors: Union[str, Path],
-    compression_config: CompressionConfig = None,
+    compression_config: SparsityCompressionConfig = None,
     device: Optional[str] = "cpu",
 ) -> Generator[Tuple[str, Tensor], None, None]:
     """
@@ -90,9 +90,9 @@ def load_compressed(
 
     if (
         compression_config is None
-        or compression_config.format == CompressionFormat.dense_sparsity.value
+        or compression_config.format == CompressionFormat.dense.value
     ):
-        # if no compression_config specified, or `dense_sparsity` format specified,
+        # if no compression_config specified, or `dense` format specified,
         # assume tensors are not compressed on disk
         weight_mappings = get_weight_mappings(compressed_tensors)
         for weight_name, file_with_weight_name in weight_mappings.items():
@@ -102,7 +102,7 @@ def load_compressed(
     else:
         # decompress tensors
         compression_format = compression_config.format
-        compressor = ModelCompressor.load_from_registry(
+        compressor = Compressor.load_from_registry(
             compression_format, config=compression_config
         )
         yield from compressor.decompress(compressed_tensors, device=device)

compressed_tensors/compressors/marlin_24.py

@@ -0,0 +1,251 @@
+# 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 logging
+from typing import Dict, Generator, Tuple
+
+import numpy as np
+import torch
+from compressed_tensors.compressors import Compressor
+from compressed_tensors.config import CompressionFormat
+from compressed_tensors.quantization import QuantizationArgs, QuantizationStrategy
+from compressed_tensors.quantization.lifecycle.forward import quantize
+from compressed_tensors.utils import (
+    get_permutations_24,
+    is_quantization_param,
+    merge_names,
+    sparse_semi_structured_from_dense_cutlass,
+    tensor_follows_mask_structure,
+)
+from torch import Tensor
+from tqdm import tqdm
+
+
+_LOGGER: logging.Logger = logging.getLogger(__name__)
+
+
+@Compressor.register(name=CompressionFormat.marlin_24.value)
+class Marlin24Compressor(Compressor):
+    """
+    Compresses a quantized model with 2:4 sparsity structure for inference with the
+    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]
+    ) -> bool:
+        """
+        Checks if every quantized module in the model is compatible with Marlin24
+        compression. Quantization must be channel or group strategy with group_size
+        of 128. Only symmetric quantization is supported
+
+        :param model_quant_args: dictionary of mapping module names to their
+            quantization configuration
+        :return: True if all modules are compatible with Marlin24 compression, raises
+            a ValueError otherwise
+        """
+        for name, quant_args in model_quant_args.items():
+            strategy = quant_args.strategy
+            group_size = quant_args.group_size
+            symmetric = quant_args.symmetric
+            if (
+                strategy is not QuantizationStrategy.GROUP.value
+                and strategy is not QuantizationStrategy.CHANNEL.value
+            ):
+                raise ValueError(
+                    f"Marlin24 Compressor is only valid for group and channel "
+                    f"quantization strategies, got {strategy} in {name}"
+                )
+
+            if group_size is not None and group_size != 128:
+                raise ValueError(
+                    f"Marlin24 Compressor is only valid for group size 128, "
+                    f"got {group_size} in {name}"
+                )
+
+            if not symmetric:
+                raise ValueError(
+                    f"Marlin24 Compressor is only valid for symmetric quantzation, "
+                    f"got symmetric={symmetric} in {name}"
+                )
+
+        return True
+
+    @staticmethod
+    def validate_sparsity_structure(name: str, weight: Tensor) -> bool:
+        """
+        Checks if a tensor fits the required 2:4 sparsity structure
+
+        :param name: name of the tensor to check
+        :param weight: tensor to check for sparsity structure
+        :return: True if all rows match the 2:4 sparsity structure, raises
+            ValueError otherwise
+        """
+
+        if not tensor_follows_mask_structure(weight):
+            raise ValueError(
+                "Marlin24 Compressor is only compatible with weights that have "
+                f"a 2:4 sparsity structure. Found segments in {name} "
+                "that do not match the expected structure."
+            )
+
+        return True
+
+    def compress(
+        self,
+        model_state: Dict[str, Tensor],
+        names_to_scheme: Dict[str, QuantizationArgs],
+        **kwargs,
+    ) -> Dict[str, Tensor]:
+        """
+        Compresses a quantized state_dict with 2:4 sparsity structure for inference
+        with the Marlin24 kernel
+
+        :param model_state: state dict of uncompressed model
+        :param names_to_scheme: quantization args for each quantized weight, needed for
+           quantize function to calculate bit depth
+        :return: compressed state dict
+        """
+        self.validate_quant_compatability(names_to_scheme)
+
+        compressed_dict = {}
+        weight_suffix = ".weight"
+        _LOGGER.debug(
+            f"Compressing model with {len(model_state)} parameterized layers..."
+        )
+
+        for name, value in tqdm(model_state.items(), desc="Compressing model"):
+            if name.endswith(weight_suffix):
+                prefix = name[: -(len(weight_suffix))]
+                scale = model_state.get(merge_names(prefix, "weight_scale"), None)
+                zp = model_state.get(merge_names(prefix, "weight_zero_point"), None)
+                if scale is not None:  # weight is quantized, compress it
+
+                    # Marlin24 kernel requires float16 inputs
+                    scale = scale.to(torch.float16)
+                    value = value.to(torch.float16)
+
+                    # quantize weight, keeping it as a float16 for now
+                    quant_args = names_to_scheme[prefix]
+                    value = quantize(
+                        x=value, scale=scale, zero_point=zp, args=quant_args
+                    )
+
+                    # compress based on sparsity structure
+                    self.validate_sparsity_structure(prefix, value)
+                    value, meta = compress_weight_24(value)
+                    meta = meta.cpu()
+
+                    # Marlin24 kernel expects input dim first
+                    value = value.t().contiguous().cpu()
+                    scale = scale.t().contiguous().cpu()
+                    og_weight_shape = value.shape
+
+                    # Marlin24 kernel expects unsigned values, shift zero-point
+                    value += (1 << quant_args.num_bits) // 2
+
+                    # pack quantized weight and scale
+                    value = pack_weight_24(value, quant_args)
+                    packed_scale = pack_scales_24(scale, quant_args, og_weight_shape)
+                    meta = meta.resize_(meta.shape[1] // 2, meta.shape[0] * 2)
+
+                    # save compressed values
+                    compressed_dict[merge_names(prefix, "scale_packed")] = packed_scale
+                    compressed_dict[merge_names(prefix, "weight_packed")] = value
+                    compressed_dict[merge_names(prefix, "meta")] = meta
+                    continue
+
+            if not is_quantization_param(name):
+                # export unquantized parameters without modifying
+                compressed_dict[name] = value.to("cpu")
+
+        return compressed_dict
+
+    def decompress(
+        self, path_to_model_or_tensors: str, device: str = "cpu", **kwargs
+    ) -> Generator[Tuple[str, Tensor], None, None]:
+        raise NotImplementedError(
+            "Decompression is not implemented for the Marlin24 Compressor."
+        )
+
+
+def compress_weight_24(weight: Tensor):
+    weight = weight.contiguous()
+    w_comp, meta = sparse_semi_structured_from_dense_cutlass(weight)
+    w_comp = w_comp.contiguous()
+    return w_comp, meta
+
+
+def marlin_permute_weights(q_w, size_k, size_n, perm, tile):
+    assert q_w.shape == (size_k, size_n)
+    assert size_k % tile == 0, f"size_k = {size_k}, tile = {tile}"
+    assert size_n % tile == 0, f"size_k = {size_n}, tile = {tile}"
+
+    # Permute weights to 16x64 marlin tiles
+    q_w = q_w.reshape((size_k // tile, tile, size_n // tile, tile))
+    q_w = q_w.permute((0, 2, 1, 3))
+    q_w = q_w.reshape((size_k // tile, size_n * tile))
+
+    q_w = q_w.reshape((-1, perm.numel()))[:, perm].reshape(q_w.shape)
+
+    return q_w
+
+
+def pack_weight_24(
+    weight: Tensor,
+    quantization_args: QuantizationArgs,
+    tile: int = 16,
+):
+    size_k = weight.shape[0]
+    size_n = weight.shape[1]
+    num_bits = quantization_args.num_bits
+    pack_factor = 32 // num_bits
+
+    # Reshuffle to marlin_24 format
+    perm, _, _ = get_permutations_24(num_bits)
+    q_w = marlin_permute_weights(weight, size_k, size_n, perm, tile)
+
+    q_w = q_w.cpu().numpy().astype(np.uint32)
+
+    q_packed = np.zeros((q_w.shape[0], q_w.shape[1] // pack_factor), dtype=np.uint32)
+    for i in range(pack_factor):
+        q_packed |= q_w[:, i::pack_factor] << num_bits * i
+
+    q_packed = torch.from_numpy(q_packed.astype(np.int32))
+
+    return q_packed
+
+
+def pack_scales_24(scales, quantization_args, w_shape):
+    size_k = w_shape[0]
+    size_n = w_shape[1]
+    num_bits = quantization_args.num_bits
+
+    _, scale_perm_2_4, scale_perm_single_2_4 = get_permutations_24(num_bits)
+
+    if (
+        quantization_args.strategy is QuantizationStrategy.GROUP
+        and quantization_args.group_size < size_k
+    ):
+        scales = scales.reshape((-1, len(scale_perm_2_4)))[:, scale_perm_2_4]
+    else:  # channelwise
+        scales = scales.reshape((-1, len(scale_perm_single_2_4)))[
+            :, scale_perm_single_2_4
+        ]
+    scales = scales.reshape((-1, size_n)).contiguous()
+
+    return scales