compressed-tensors 0.8.1__py3-none-any.whl → 0.9.0__py3-none-any.whl

compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py

@@ -14,12 +14,12 @@
 
 from typing import Dict, List, Tuple, Union
 
-import numpy
 import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.compressors.sparse_compressors.base import BaseSparseCompressor
 from compressed_tensors.config import CompressionFormat
-from compressed_tensors.utils import merge_names
+from compressed_tensors.quantization import FP8_DTYPE
+from compressed_tensors.utils import merge_names, pack_bitmasks, unpack_bitmasks
 from torch import Tensor
 
 
@@ -28,8 +28,6 @@ __all__ = [
     "BitmaskTensor",
     "bitmask_compress",
     "bitmask_decompress",
-    "pack_bitmasks",
-    "unpack_bitmasks",
 ]
 
 
@@ -134,9 +132,14 @@ def bitmask_compress(tensor: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
     bytemasks = tensor != 0
     row_counts = bytemasks.sum(dim=-1)
     row_offsets = torch.cumsum(row_counts, 0) - row_counts
-    values = tensor[bytemasks]
+    if tensor.dtype == FP8_DTYPE:
+        # acces raw bytes of the tensor
+        tensor_view = tensor.view(torch.int8)
+        values = tensor_view[bytemasks]
+        values = values.view(FP8_DTYPE)
+    else:
+        values = tensor[bytemasks]
     bitmasks_packed = pack_bitmasks(bytemasks)
-
     return values, bitmasks_packed, row_offsets
 
 
@@ -158,37 +161,3 @@ def bitmask_decompress(
     decompressed_tensor[bytemasks_unpacked] = values
 
     return decompressed_tensor
-
-
-def pack_bitmasks(bytemasks: Tensor) -> Tensor:
-    """
-    Converts a bytemask tensor to a bitmask tensor to reduce memory. Shape RxC will be
-    compressed to R x ceil(C/8)
-    :param bytemasks: mask tensor where each byte corresponds to a weight
-    :return: mask tensor where each bit corresounds to a weight
-    """
-    packed_bits_numpy = numpy.packbits(bytemasks.numpy(), axis=-1, bitorder="little")
-    packed_bits_torch = torch.from_numpy(packed_bits_numpy)
-
-    return packed_bits_torch
-
-
-def unpack_bitmasks(packed_bitmasks: Tensor, original_shape: torch.Size) -> Tensor:
-    """
-    Converts a bitmask tensor back to a bytemask tensor for use during decompression
-
-    :param packed_bitmasks: mask tensor where each bit corresponds to a weight
-    :param original_shape: dense shape to decompress to
-    :return: boolean mask of weights in the original dense shape
-    """
-    # Unpack the bits
-    unpacked_bits = numpy.unpackbits(
-        packed_bitmasks.numpy(), axis=-1, count=original_shape[-1], bitorder="little"
-    )
-
-    # Reshape to match the original shape
-    unpacked_bitmasks_torch = torch.from_numpy(
-        unpacked_bits.reshape(original_shape).astype(bool)
-    )
-
-    return unpacked_bitmasks_torch

compressed_tensors/config/__init__.py

@@ -15,4 +15,5 @@
 # flake8: noqa
 from .base import *
 from .dense import *
+from .sparse_24_bitmask import *
 from .sparse_bitmask import *

compressed_tensors/config/base.py

@@ -26,6 +26,7 @@ __all__ = ["SparsityCompressionConfig", "CompressionFormat", "SparsityStructure"
 class CompressionFormat(Enum):
     dense = "dense"
     sparse_bitmask = "sparse-bitmask"
+    sparse_24_bitmask = "sparse-24-bitmask"
     int_quantized = "int-quantized"
     float_quantized = "float-quantized"
     naive_quantized = "naive-quantized"

compressed_tensors/config/sparse_24_bitmask.py

@@ -0,0 +1,40 @@
+# 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.
+
+from typing import Optional
+
+from compressed_tensors.config import (
+    CompressionFormat,
+    SparsityCompressionConfig,
+    SparsityStructure,
+)
+
+
+__all__ = ["Sparse24BitMaskConfig"]
+
+
+@SparsityCompressionConfig.register(name=CompressionFormat.sparse_24_bitmask.value)
+class Sparse24BitMaskConfig(SparsityCompressionConfig):
+    """
+    Configuration for storing a 24 sparse model using
+    bytemask compression
+
+    :param global_sparsity: average sparsity of the entire model
+    :param sparsity_structure: structure of the sparsity, should always be
+        "2:4" for this compression format
+    """
+
+    format: str = CompressionFormat.sparse_24_bitmask.value
+    global_sparsity: Optional[float] = 0.0
+    sparsity_structure: Optional[str] = SparsityStructure.TWO_FOUR.value

compressed_tensors/quantization/lifecycle/apply.py

@@ -18,7 +18,7 @@ from collections import OrderedDict, defaultdict
 from copy import deepcopy
 from typing import Dict, Iterable, List, Optional
 from typing import OrderedDict as OrderedDictType
-from typing import Union
+from typing import Set, Union
 
 import torch
 from compressed_tensors.config import CompressionFormat
@@ -52,6 +52,8 @@ __all__ = [
     "apply_quantization_config",
     "apply_quantization_status",
     "find_name_or_class_matches",
+    "expand_sparse_target_names",
+    "is_sparse_target",
 ]
 
 from compressed_tensors.quantization.utils.helpers import is_module_quantized
@@ -245,6 +247,49 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
         model.apply(compress_quantized_weights)
 
 
+def expand_sparse_target_names(
+    model: Module, targets: Iterable[str], ignore: Iterable[str]
+) -> Set[str]:
+    """
+    Finds all unique module names in the model that match the given
+    targets and ignore lists.
+
+    Note: Targets must be regexes, layer types, or full layer names.
+
+    :param model: model to search for targets in
+    :param targets: list of targets to search for
+    :param ignore: list of targets to ignore
+    :return: set of all targets that match the given targets and should
+        not be ignored
+    """
+    return {
+        name
+        for name, module in iter_named_leaf_modules(model)
+        if is_sparse_target(name, module, targets, ignore)
+    }
+
+
+def is_sparse_target(
+    name: str, module: Module, targets: Iterable[str], ignore: Iterable[str]
+) -> bool:
+    """
+    Determines if a module should be included in the targets based on the
+    targets and ignore lists.
+
+    Note: Targets must be regexes, layer types, or full layer names.
+
+    :param name: name of the module
+    :param module: the module itself
+    :param targets: list of targets to search for
+    :param ignore: list of targets to ignore
+    :return: True if the module is a target and not ignored, False otherwise
+    """
+    return bool(
+        find_name_or_class_matches(name, module, targets)
+        and not find_name_or_class_matches(name, module, ignore or [])
+    )
+
+
 def find_name_or_class_matches(
     name: str, module: Module, targets: Iterable[str], check_contains: bool = False
 ) -> List[str]:

compressed_tensors/quantization/lifecycle/forward.py

@@ -82,8 +82,8 @@ def quantize(
 def dequantize(
     x_q: torch.Tensor,
     scale: torch.Tensor,
-    zero_point: torch.Tensor = None,
-    args: QuantizationArgs = None,
+    zero_point: Optional[torch.Tensor] = None,
+    args: Optional[QuantizationArgs] = None,
     dtype: Optional[torch.dtype] = None,
     g_idx: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:

compressed_tensors/quantization/lifecycle/initialize.py

@@ -29,7 +29,11 @@ from compressed_tensors.quantization.quant_args import (
 from compressed_tensors.quantization.quant_config import QuantizationStatus
 from compressed_tensors.quantization.quant_scheme import QuantizationScheme
 from compressed_tensors.quantization.utils import is_kv_cache_quant_scheme
-from compressed_tensors.utils import get_execution_device, is_module_offloaded
+from compressed_tensors.utils import (
+    disable_hf_hook,
+    has_offloaded_params,
+    register_offload_parameter,
+)
 from torch.nn import Module, Parameter
 
 
@@ -112,43 +116,10 @@ def initialize_module_for_quantization(
         module.quantization_scheme = scheme
         module.quantization_status = QuantizationStatus.INITIALIZED
 
-        offloaded = False
-        # What is this doing/why isn't this in the attn case?
-        if is_module_offloaded(module):
-            try:
-                from accelerate.hooks import add_hook_to_module, remove_hook_from_module
-                from accelerate.utils import PrefixedDataset
-            except ModuleNotFoundError:
-                raise ModuleNotFoundError(
-                    "Offloaded model detected. To use CPU offloading with "
-                    "compressed-tensors the `accelerate` package must be installed, "
-                    "run `pip install compressed-tensors[accelerate]`"
-                )
-
-            offloaded = True
-            hook = module._hf_hook
-            prefix_dict = module._hf_hook.weights_map
-            new_prefix = {}
-
-            # recreate the prefix dict (since it is immutable)
-            # and add quantization parameters
-            for key, data in module.named_parameters():
-                if key not in prefix_dict:
-                    new_prefix[f"{prefix_dict.prefix}{key}"] = data
-                else:
-                    new_prefix[f"{prefix_dict.prefix}{key}"] = prefix_dict[key]
-            new_prefix_dict = PrefixedDataset(new_prefix, prefix_dict.prefix)
-            remove_hook_from_module(module)
-
-        # wrap forward call of module to perform
-        # quantized actions based on calltime status
-        wrap_module_forward_quantized(module, scheme)
-
-        if offloaded:
-            # we need to re-add the hook for offloading now that we've wrapped forward
-            add_hook_to_module(module, hook)
-            if prefix_dict is not None:
-                module._hf_hook.weights_map = new_prefix_dict
+        with disable_hf_hook(module):
+            # wrap forward call of module to perform
+            # quantized actions based on calltime status
+            wrap_module_forward_quantized(module, scheme)
 
 
 def is_attention_module(module: Module):
@@ -169,12 +140,17 @@ def _initialize_scale_zero_point(
     if quantization_args.dynamic:
         return
 
-    device = next(module.parameters()).device
-    if is_module_offloaded(module):
-        device = get_execution_device(module)
+    # begin on the same device as other parameters or cpu if offloaded.
+    # in the offloaded case, there's no point moving tensors to the execution device
+    # if they're going to be immediately offloaded by `register_offload_parameter`
+    params_device = next(module.parameters()).device
+    device = "cpu" if has_offloaded_params(module) else params_device
 
     # infer expected scale/zero point shape
-    expected_shape = 1  # per tensor
+    if quantization_args.strategy == QuantizationStrategy.TOKEN:
+        expected_shape = (1, 1)
+    else:
+        expected_shape = 1
 
     if base_name == "weight" and weight_shape is not None:
         if quantization_args.strategy == QuantizationStrategy.CHANNEL:
@@ -193,7 +169,7 @@ def _initialize_scale_zero_point(
         torch.empty(expected_shape, dtype=scale_dtype, device=device),
         requires_grad=False,
     )
-    module.register_parameter(f"{base_name}_scale", init_scale)
+    register_offload_parameter(module, f"{base_name}_scale", init_scale)
 
     if force_zero_point or not quantization_args.symmetric:
         zp_dtype = quantization_args.pytorch_dtype()
@@ -201,7 +177,7 @@ def _initialize_scale_zero_point(
             torch.zeros(expected_shape, device=device, dtype=zp_dtype),
             requires_grad=False,
         )
-        module.register_parameter(f"{base_name}_zero_point", init_zero_point)
+        register_offload_parameter(module, f"{base_name}_zero_point", init_zero_point)
 
     # only grouped activation ordering has g_idx
     if quantization_args.actorder == ActivationOrdering.GROUP:
@@ -211,7 +187,7 @@ def _initialize_scale_zero_point(
             torch.full(g_idx_shape, -1, device=device, dtype=g_idx_dtype),
             requires_grad=False,
         )
-        module.register_parameter(f"{base_name}_g_idx", init_g_idx)
+        register_offload_parameter(module, f"{base_name}_g_idx", init_g_idx)
 
 
 def _initialize_attn_scales(module: Module) -> None:

compressed_tensors/quantization/quant_config.py

@@ -160,7 +160,7 @@ class QuantizationConfig(BaseModel):
 
     def to_dict(self):
         # for compatibility with HFQuantizer
-        return self.dict()
+        return self.model_dump()
 
     @staticmethod
     def from_pretrained(

compressed_tensors/utils/helpers.py

@@ -12,8 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Dict, Optional
+import warnings
+from functools import wraps
+from typing import Any, Callable, Dict, List, Optional
 
+import numpy
 import torch
 from transformers import AutoConfig
 
@@ -24,7 +27,13 @@ __all__ = [
     "tensor_follows_mask_structure",
     "replace_module",
     "is_compressed_tensors_config",
+    "getattr_chain",
+    "deprecated",
     "Aliasable",
+    "combine_shards",
+    "shard_tensor",
+    "pack_bitmasks",
+    "unpack_bitmasks",
 ]
 
 FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
@@ -122,6 +131,65 @@ def is_compressed_tensors_config(compression_config: Any) -> bool:
         return False
 
 
+def getattr_chain(obj: Any, chain_str: str, *args, **kwargs) -> Any:
+    """
+    Chain multiple getattr calls, separated by `.`
+
+    :param obj: base object whose attributes are being retrieved
+    :param chain_str: attribute names separated by `.`
+    :param default: default value, throw error otherwise
+    """
+    if len(args) >= 1:
+        has_default = True
+        default = args[0]
+    elif "default" in kwargs:
+        has_default = True
+        default = kwargs["default"]
+    else:
+        has_default = False
+
+    attr_names = chain_str.split(".")
+
+    res = obj
+    for attr_name in attr_names:
+        if not hasattr(res, attr_name):
+            if has_default:
+                return default
+            else:
+                raise AttributeError(f"{res} object has no attribute {attr_name}")
+        res = getattr(res, attr_name)
+
+    return res
+
+
+def deprecated(future_name: Optional[str] = None, message: Optional[str] = None):
+    """
+    Decorator to mark functions as deprecated
+
+    :param new_function: Function called in place of depreciated function
+    :param message: Depreciation message, replaces default depreciation message
+    """
+
+    def decorator(func: Callable[[Any], Any]):
+        nonlocal message
+
+        if message is None:
+            message = (
+                f"{func.__name__} is deprecated and will be removed in a future release"
+            )
+            if future_name is not None:
+                message += f". Please use {future_name} instead."
+
+        @wraps(func)
+        def wrapped(*args, **kwargs):
+            warnings.warn(message, DeprecationWarning, stacklevel=2)
+            return func(*args, **kwargs)
+
+        return wrapped
+
+    return decorator
+
+
 class Aliasable:
     """
     A mixin for enums to allow aliasing of enum members
@@ -151,3 +219,108 @@ class Aliasable:
     def __hash__(self):
         canonical_value = self.aliases.get(self.value, self.value)
         return hash(canonical_value)
+
+
+def shard_tensor(
+    tensor: torch.Tensor, shard_sizes: List[int], dim: int = 0
+) -> List[torch.Tensor]:
+    """
+    Shards a tensor into a list of tensors along a given dimension.
+
+    raises: ValueError: If the sum of shard_sizes does not match the
+        size of the tensor along the given dimension.
+
+    :param tensor: The input tensor to shard.
+    :param shard_sizes : List of sizes for each shard along the specified dimension.
+    :param dim : The dimension along which to shard the tensor.
+    :returns: A list of tensors sharded along the specified dimension.
+    """
+    if sum(shard_sizes) != tensor.size(dim):
+        raise ValueError(
+            "Sum of shard_sizes must equal the size of the tensor "
+            "along the specified dimension."
+        )
+
+    shards = []
+    start_idx = 0
+
+    for size in shard_sizes:
+        end_idx = start_idx + size
+        shard = tensor.narrow(dim, start_idx, size)
+        shards.append(shard)
+        start_idx = end_idx
+
+    return shards
+
+
+def combine_shards(shards, dim=0):
+    """
+    Combine decompressed shards along a given dimension using `narrow`.
+
+    :param shards: List of decompressed shard tensors.
+    :param dim: Dimension to combine along (default: 0).
+    :return: Combined decompressed tensor.
+    """
+    if not shards:
+        raise ValueError("The list of shards is empty.")
+
+    # Assert that all shards have the same dtype
+    shard_dtypes = {shard.dtype for shard in shards}
+    if len(shard_dtypes) > 1:
+        raise ValueError("All shards must have the same dtype.")
+
+    # Determine the total shape of the combined tensor
+    total_shape = list(shards[0].shape)
+    total_shape[dim] = sum(shard.shape[dim] for shard in shards)
+
+    # Create the combined tensor
+    combined = torch.zeros(total_shape, dtype=shards[0].dtype, device=shards[0].device)
+
+    # Fill the combined tensor using narrow
+    shard_offset = 0
+    for shard in shards:
+        shard_size = shard.shape[dim]
+        combined.narrow(dim, shard_offset, shard_size).copy_(shard)
+        shard_offset += shard_size
+
+    return combined
+
+
+def pack_bitmasks(bytemasks: torch.Tensor) -> torch.Tensor:
+    """
+    Converts a bytemask tensor to a bitmask tensor to reduce memory. Shape RxC will be
+    compressed to R x ceil(C/8)
+
+    :param bytemasks: mask tensor where each byte corresponds to a weight
+    :return: mask tensor where each bit corresounds to a weight
+    """
+    packed_bits_numpy = numpy.packbits(bytemasks.numpy(), axis=-1, bitorder="little")
+    packed_bits_torch = torch.from_numpy(packed_bits_numpy)
+
+    return packed_bits_torch
+
+
+def unpack_bitmasks(
+    packed_bitmasks: torch.Tensor, original_shape: torch.Size
+) -> torch.Tensor:
+    """
+    Converts a bitmask tensor back to a bytemask tensor for use during decompression
+
+    :param packed_bitmasks: mask tensor where each bit corresponds to a weight
+    :param original_shape: dense shape to decompress to
+    :return: boolean mask of weights in the original dense shape
+    """
+    # Unpack the bits
+    unpacked_bits = numpy.unpackbits(
+        packed_bitmasks.cpu().numpy(),
+        axis=-1,
+        count=original_shape[-1],
+        bitorder="little",
+    )
+
+    # Reshape to match the original shape
+    unpacked_bitmasks_torch = torch.from_numpy(
+        unpacked_bits.reshape(original_shape).astype(bool)
+    )
+
+    return unpacked_bitmasks_torch