compressed-tensors 0.11.1a20250820__py3-none-any.whl → 0.11.1a20250828__py3-none-any.whl

compressed_tensors/quantization/lifecycle/apply.py

@@ -13,12 +13,11 @@
 # limitations under the License.
 
 import logging
-import re
-from collections import OrderedDict, defaultdict
+from collections import OrderedDict
 from copy import deepcopy
 from typing import Dict, Iterable, List, Optional
 from typing import OrderedDict as OrderedDictType
-from typing import Set, Union
+from typing import Union
 
 import torch
 from compressed_tensors.config import CompressionFormat
@@ -39,7 +38,8 @@ from compressed_tensors.quantization.utils import (
     infer_quantization_status,
     is_kv_cache_quant_scheme,
 )
-from compressed_tensors.utils.helpers import fix_fsdp_module_name, replace_module
+from compressed_tensors.utils.helpers import deprecated, replace_module
+from compressed_tensors.utils.match import match_named_modules, match_targets
 from compressed_tensors.utils.offload import update_parameter_data
 from compressed_tensors.utils.safetensors_load import get_safetensors_folder
 from safetensors import safe_open
@@ -51,8 +51,6 @@ __all__ = [
     "apply_quantization_config",
     "apply_quantization_status",
     "find_name_or_class_matches",
-    "expand_target_names",
-    "is_target",
 ]
 
 from compressed_tensors.quantization.utils.helpers import is_module_quantized
@@ -73,14 +71,14 @@ def load_pretrained_quantization_parameters(
     Loads the quantization parameters (scale and zero point) from model_name_or_path to
     a model that has already been initialized with a quantization config.
 
-    NOTE: Will always load inputs/output parameters.
-    Will conditioanlly load weight parameters, if load_weight_quantization is set to True.
+    NOTE: Will always load inputs/output parameters. Will conditioanlly load weight
+    parameters, if load_weight_quantization is set to True.
 
     :param model: model to load pretrained quantization parameters to
     :param model_name_or_path: Hugging Face stub or local folder containing a quantized
         model, which is used to load quantization parameters
-    :param load_weight_quantization: whether or not the weight quantization parameters shoud
-        be laoded
+    :param load_weight_quantization: whether or not the weight quantization parameters
+        should be loaded
     """
     model_path = get_safetensors_folder(model_name_or_path)
     mapping = get_quantization_parameter_to_path_mapping(model_path)
@@ -117,7 +115,7 @@ def load_pretrained_quantization_parameters(
 
 def apply_quantization_config(
     model: Module, config: Union[QuantizationConfig, None], run_compressed: bool = False
-) -> Dict[str, QuantizationScheme]:
+):
     """
     Initializes the model for quantization in-place based on the given config.
     Optionally coverts quantizable modules to compressed_linear modules
@@ -127,71 +125,49 @@ def apply_quantization_config(
     :param run_compressed: Whether the model will be run in compressed mode or
         decompressed fully on load
     """
-    # Workaround for when HF Quantizer passes None, see PR #180
-    if config is None:
-        return dict()
+    from compressed_tensors.linear.compressed_linear import CompressedLinear
 
-    # remove reference to the original `config`
-    # argument. This function can mutate it, and we'd
-    # like to keep the original `config` as it is.
     config = deepcopy(config)
+    if config is None:  # see PR #180
+        return dict()
+
+    # preprocess to support kv cache scheme
+    config = process_quantization_config(config)
+
     # build mapping of targets to schemes for easier matching
     # use ordered dict to preserve target ordering in config
     target_to_scheme = OrderedDict()
-    config = process_quantization_config(config)
-    names_to_scheme = dict()
     for scheme in config.config_groups.values():
         for target in scheme.targets:
             target_to_scheme[target] = scheme
 
-    if run_compressed:
-        from compressed_tensors.linear.compressed_linear import CompressedLinear
-
-    # list of submodules to ignore
-    ignored_submodules = defaultdict(list)
     # mark appropriate layers for quantization by setting their quantization schemes
-    for name, submodule in model.named_modules():
-        # potentially fix module name to remove FSDP wrapper prefix
-        name = fix_fsdp_module_name(name)
-        if matches := find_name_or_class_matches(name, submodule, config.ignore):
-            for match in matches:
-                ignored_submodules[match].append(name)
-            continue  # layer matches ignore list, continue
-
-        targets = find_name_or_class_matches(name, submodule, target_to_scheme)
-
-        if targets:
-            # mark modules to be quantized by adding
-            # quant scheme to the matching layers
-            scheme = _scheme_from_targets(target_to_scheme, targets, name)
-            if run_compressed:
-                format = config.format
-                if format != CompressionFormat.dense.value:
-                    if isinstance(submodule, torch.nn.Linear):
-                        # TODO: expand to more module types
-                        compressed_linear = CompressedLinear.from_linear(
-                            submodule,
-                            quantization_scheme=scheme,
-                            quantization_format=format,
-                        )
-                        replace_module(model, name, compressed_linear)
-
-            # target matched - add layer and scheme to target list
-            submodule.quantization_scheme = scheme
-
-            names_to_scheme[name] = submodule.quantization_scheme
-
-    if config.ignore is not None and ignored_submodules is not None:
-        if set(config.ignore) - set(ignored_submodules):
-            _LOGGER.warning(
-                "Some layers that were to be ignored were "
-                "not found in the model: "
-                f"{set(config.ignore) - set(ignored_submodules)}"
-            )
+    for name, submodule in match_named_modules(
+        model, target_to_scheme, config.ignore, warn_on_fail=True
+    ):
+        # mark modules to be quantized by adding
+        # quant scheme to the matching layers
+        matched_targets = match_targets(name, submodule, target_to_scheme)
+        scheme = _scheme_from_targets(target_to_scheme, matched_targets, name)
+        # target matched - add layer and scheme to target list
+        submodule.quantization_scheme = scheme
+
+        # replace with run compressed if applicable
+        # FUTURE: move this to model compressor
+        if isinstance(submodule, torch.nn.Linear) and run_compressed:
+            format = config.format
+            if format != CompressionFormat.dense.value:
+                if isinstance(submodule, torch.nn.Linear):
+                    # TODO: expand to more module types
+                    compressed_linear = CompressedLinear.from_linear(
+                        submodule,
+                        quantization_scheme=scheme,
+                        quantization_format=format,
+                    )
+                    replace_module(model, name, compressed_linear)
 
     # apply current quantization status across all targeted layers
     apply_quantization_status(model, config.quantization_status)
-    return names_to_scheme
 
 
 def process_quantization_config(config: QuantizationConfig) -> QuantizationConfig:
@@ -262,54 +238,10 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
         model.apply(compress_quantized_weights)
 
 
-def expand_target_names(
-    model: Module,
-    targets: Optional[Iterable[str]] = None,
-    ignore: Optional[Iterable[str]] = None,
-) -> 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: Iterable of targets to search for
-    :param ignore: Iterable 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 model.named_modules()
-        if is_target(name, module, targets, ignore)
-    }
-
-
-def is_target(
-    name: str,
-    module: Module,
-    targets: Optional[Iterable[str]] = None,
-    ignore: Optional[Iterable[str]] = None,
-) -> 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: Iterable of targets to search for
-    :param ignore: Iterable 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 or [])
-        and not find_name_or_class_matches(name, module, ignore or [])
-    )
-
-
+@deprecated(
+    message="This function is deprecated and will be removed in a future release."
+    "Please use `match_targets` from `compressed_tensors.utils.match` instead."
+)
 def find_name_or_class_matches(
     name: str, module: Module, targets: Iterable[str], check_contains: bool = False
 ) -> List[str]:
@@ -322,38 +254,13 @@ def find_name_or_class_matches(
         2. matches on regex patterns
         3. matches on module names
     """
-    from compressed_tensors import InternalModule
-
-    if isinstance(module, InternalModule):
-        return []
-
-    targets = sorted(targets, key=lambda x: ("re:" in x, x))
-    if isinstance(targets, Iterable):
-        matches = _find_matches(name, targets) + _find_matches(
-            module.__class__.__name__, targets, check_contains
+    if check_contains:
+        raise NotImplementedError(
+            "This function is deprecated, and the check_contains=True option has been"
+            " removed."
         )
-        matches = [match for match in matches if match is not None]
-        return matches
 
-
-def _find_matches(
-    value: str, targets: Iterable[str], check_contains: bool = False
-) -> List[str]:
-    # returns all the targets that match value either
-    # exactly or as a regex after 're:'. if check_contains is set to True,
-    # additionally checks if the target string is contained with value.
-    matches = []
-    for target in targets:
-        if target.startswith("re:"):
-            pattern = target[3:]
-            if re.match(pattern, value):
-                matches.append(target)
-        elif check_contains:
-            if target.lower() in value.lower():
-                matches.append(target)
-        elif target == value:
-            matches.append(target)
-    return matches
+    return match_targets(name, module, targets)
 
 
 def _infer_status(model: Module) -> Optional[QuantizationStatus]:
@@ -429,7 +336,6 @@ def _scheme_from_targets(
 def _merge_schemes(
     schemes_to_merge: List[QuantizationScheme], name: str
 ) -> QuantizationScheme:
-
     kv_cache_quantization_scheme = [
         scheme for scheme in schemes_to_merge if is_kv_cache_quant_scheme(scheme)
     ]

compressed_tensors/quantization/lifecycle/forward.py

@@ -205,7 +205,8 @@ def _process_quantization(
     q_min, q_max = calculate_range(args, x.device)
     group_size = args.group_size
 
-    # blockwise FP8: quantize per 2D block, supports block_structure for static block quant
+    # blockwise FP8: quantize per 2D block, supports block_structure for static block
+    # quantization
     if args.strategy == QuantizationStrategy.BLOCK:
         original_shape = x.shape
         rows, cols = x.shape[-2], x.shape[-1]
@@ -214,8 +215,8 @@ def _process_quantization(
         # Ensure exact division (tensor dimensions must be divisible by block size)
         if rows % block_height != 0:
             raise ValueError(
-                f"Tensor height {rows} is not divisible by block_height {block_height}. "
-                f"Block quantization requires exact division."
+                f"Tensor height {rows} is not divisible by block_height {block_height}."
+                f" Block quantization requires exact division."
             )
         if cols % block_width != 0:
             raise ValueError(
@@ -295,7 +296,7 @@ def _process_quantization(
             perm = torch.argsort(g_idx)
             x = safe_permute(x, perm, dim=1)
 
-        # Maintain all dimensions apart from the last dim, which is divided by the group_size
+        # Maintain all dimensions except the last dim, which is divided by group_size
         reshaped_dims = (
             ceil(x.shape[-1] / group_size),
             group_size,

compressed_tensors/quantization/lifecycle/initialize.py

@@ -17,7 +17,7 @@ import logging
 import math
 import warnings
 from enum import Enum
-from typing import List, Optional
+from typing import Optional
 
 import torch
 from compressed_tensors.quantization.lifecycle.forward import (
@@ -87,7 +87,6 @@ def initialize_module_for_quantization(
         _initialize_attn_scales(module)
 
     else:
-
         if scheme.input_activations is not None:
             _initialize_scale_zero_point(
                 module,
@@ -183,7 +182,8 @@ def _initialize_scale_zero_point(
             num_groups = math.ceil(weight_shape[1] / quantization_args.group_size)
             expected_shape = (weight_shape[0], max(num_groups, 1))
         elif quantization_args.strategy == QuantizationStrategy.BLOCK:
-            # For block quantization, scale shape should match number of blocks - only for weights
+            # For block quantization, scale shape should match number of blocks - only
+            # for weights
             if quantization_args.block_structure is None:
                 raise ValueError(
                     "Block quantization requires block_structure to be specified"
@@ -196,9 +196,10 @@ def _initialize_scale_zero_point(
             # Warn if dimensions don't divide evenly
             if rows % block_height != 0 or cols % block_width != 0:
                 warnings.warn(
-                    f"Block quantization: tensor shape {weight_shape} does not divide evenly "
-                    f"by block structure {quantization_args.block_structure}. "
-                    f"Some blocks will be incomplete which may affect quantization quality.",
+                    f"Block quantization: tensor shape {weight_shape} does not divide"
+                    f"evenly by block structure {quantization_args.block_structure}. "
+                    f"Some blocks will be incomplete which may affect quantization"
+                    "quality.",
                     UserWarning,
                 )
 

compressed_tensors/quantization/quant_args.py

@@ -217,16 +217,18 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                 return [int(x) for x in value.split("x")]
             except Exception:
                 raise ValueError(
-                    f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+                    f"Invalid block_structure '{value}'. Must be a list of ints "
+                    "[rows, cols]."
                 )
         if isinstance(value, (list, tuple)):
             if len(value) != 2 or not all(isinstance(v, int) for v in value):
                 raise ValueError(
-                    f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+                    f"Invalid block_structure '{value}'. Must be a list of ints "
+                    "[rows, cols]."
                 )
             return list(value)
         raise ValueError(
-            f"Invalid block_structure '{value}'. Must be a list of two ints [rows, cols]."
+            f"Invalid block_structure '{value}'. Must be a list of ints [rows, cols]."
         )
 
     @field_validator("strategy", mode="before")
@@ -307,7 +309,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
             )
             if strategy not in supported_strategies:
                 raise ValueError(
-                    f"One of {supported_strategies} must be used for dynamic quantization"
+                    f"One of {supported_strategies} must be used for dynamic quant."
                 )
 
             if (
@@ -322,7 +324,7 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                         observer != "memoryless"
                     ):  # avoid annoying users with old configs
                         warnings.warn(
-                            "No observer is used for dynamic quantization, setting to None"
+                            "No observer is used for dynamic quant., setting to None"
                         )
                     observer = None
             else:

compressed_tensors/quantization/quant_scheme.py

@@ -81,9 +81,10 @@ class QuantizationScheme(BaseModel):
         ):
             warnings.warn(
                 "Using GROUP strategy for both weights and input_activations "
-                f"with different group sizes ({weights.group_size} vs {inputs.group_size}) "
-                "may complicate fused kernel implementations. Consider using "
-                "TENSOR_GROUP strategy for both or matching group sizes.",
+                f"with different group sizes ({weights.group_size} vs "
+                f"{inputs.group_size}) may complicate fused kernel implementations. "
+                "Consider using TENSOR_GROUP strategy for both or matching group"
+                " sizes.",
                 UserWarning,
                 stacklevel=2,
             )

compressed_tensors/quantization/utils/helpers.py

@@ -29,7 +29,6 @@ from compressed_tensors.quantization.quant_scheme import QuantizationScheme
 from compressed_tensors.utils import deprecated
 from torch import FloatTensor, IntTensor, Tensor
 from torch.nn import Module
-from tqdm import tqdm
 
 
 __all__ = [

compressed_tensors/registry/registry.py

@@ -55,7 +55,7 @@ def standardize_lookup_name(name: str) -> str:
 
 
 def standardize_alias_name(
-    name: Union[None, str, List[str]]
+    name: Union[None, str, List[str]],
 ) -> Union[None, str, List[str]]:
     if name is None:
         return None

compressed_tensors/transform/transform_config.py

@@ -14,7 +14,7 @@
 
 from typing import Dict
 
-from compressed_tensors.transform import TransformArgs, TransformScheme
+from compressed_tensors.transform import TransformScheme
 from pydantic import BaseModel, ConfigDict
 
 

compressed_tensors/transform/utils/matrix.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Callable, Optional, Tuple
+from typing import Optional
 
 import torch
 from compressed_tensors.transform import TransformLocation

compressed_tensors/utils/match.py

@@ -27,6 +27,7 @@ _LOGGER: logging.Logger = logging.getLogger(__name__)
 __all__ = [
     "match_named_modules",
     "match_named_parameters",
+    "match_targets",
     "match_modules_set",
     "is_match",
 ]
@@ -37,8 +38,8 @@ FusedMappping = Mapping[str, Iterable[str]]
 
 def match_named_modules(
     model: torch.nn.Module,
-    targets: Iterable[str],
-    ignore: Iterable[str] = tuple(),
+    targets: Optional[Iterable[str]],
+    ignore: Optional[Iterable[str]] = None,
     fused: Optional[FusedMappping] = None,
     warn_on_fail: bool = False,
 ) -> Generator[Tuple[str, torch.nn.Module]]:
@@ -54,14 +55,18 @@ def match_named_modules(
     :param warn_on_fail: if True, warns if any targets do not match any modules in model
     :return: generator of module names and modules
     """
+    targets = targets or []
+    ignore = ignore or []
+
     unmatched_targets = set(targets)
+
     for name, module in model.named_modules():
         for target in targets:
             if is_match(name, module, target, fused=fused):
                 unmatched_targets -= {target}
-
                 if not is_match(name, module, ignore, fused=fused):
                     yield name, module
+                break
 
     if warn_on_fail:
         for target in unmatched_targets:
@@ -72,8 +77,8 @@ def match_named_modules(
 
 def match_named_parameters(
     model: torch.nn.Module,
-    targets: Iterable[str],
-    ignore: Iterable[str] = tuple(),
+    targets: Optional[Iterable[str]],
+    ignore: Optional[Iterable[str]] = None,
     fused: Optional[FusedMappping] = None,
     warn_on_fail: bool = False,
 ) -> Generator[Tuple[str, torch.nn.Module, torch.nn.Parameter]]:
@@ -89,6 +94,9 @@ def match_named_parameters(
     :param warn_on_fail: if True, warns if any targets do not match any params in model
     :return: generator of fully-qualified param names, parent modules, and params
     """
+    targets = targets or []
+    ignore = ignore or []
+
     unmatched_targets = set(targets)
     for module_name, module in model.named_modules():
         if isinstance(module, InternalModule):
@@ -110,16 +118,54 @@ def match_named_parameters(
             )
 
 
+def match_targets(
+    name: str, module: torch.nn.Module, targets: Optional[Iterable[str]]
+) -> List[str]:
+    """
+    Returns the targets that match the given name and module.
+
+    :param name: the name of the module
+    :param module: the module to match
+    :param targets: the target strings, potentially containing "re:" prefixes
+    :return: the targets that match the given name and module
+
+    Outputs are ordered by type: exact name match, regex name match, class name match
+    """
+    targets = targets or []
+
+    if isinstance(module, InternalModule):
+        return []
+
+    # The order of the output `matches` list matters, the are arranged from most
+    # specific to least specific, and this order will be used when merging configs.
+    # The entries are sorted in the following order:
+    #     1. matches on exact strings
+    #     2. matches on regex patterns
+    #     3. matches on module names
+
+    targets = sorted(targets, key=lambda x: ("re:" in x, x))
+    matched_targets = []
+    for target in targets:
+        if _match_name(name, target):
+            matched_targets.append(target)
+
+    for target in targets:
+        if _match_class(module, target) and target not in matched_targets:
+            matched_targets.append(target)
+
+    return matched_targets
+
+
 def match_modules_set(
     model: torch.nn.Module,
-    targets: Iterable[str],
-    ignore: Iterable[str] = tuple(),
+    targets: Optional[Iterable[str]],
+    ignore: Optional[Iterable[str]] = None,
 ) -> Generator[Iterable[torch.nn.Module]]:
     """
     Yields modules grouped with the same order and size as `targets`.
     Values are returned in order of `model.named_modules()`
 
-    For example, the following targets would yield module belonging to the following layers:
+    E.g. the following targets would yield module belonging to the following layers:
     ```python3
     match_modules_set(model, ["q_proj", "k_proj", "v_proj"]) == (
         (
@@ -151,6 +197,9 @@ def match_modules_set(
     :param targets: target strings, potentially containing "re:" prefixes
     :param ignore: targets to ignore, potentially containing "re:" prefixes
     """
+    targets = targets or []
+    ignore = ignore or []
+
     matches = dict.fromkeys(targets, None)
     for name, module in model.named_modules():
         # match until we get a full set

compressed_tensors/utils/offload.py

@@ -296,7 +296,6 @@ def disable_hf_hook(module: torch.nn.Module):
     hooks = {}
 
     def collect_hooks(module):
-        nonlocal hooks
         if hasattr(module, "_hf_hook"):
             hooks[module] = module._hf_hook
             remove_hook_from_module(module)

compressed_tensors/utils/safetensors_load.py

@@ -18,7 +18,6 @@ import re
 import struct
 from typing import Dict, Iterable, Optional, Tuple, Union
 
-from safetensors import safe_open
 from torch import Tensor
 from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, SAFE_WEIGHTS_NAME, cached_file
 

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.11.1.a20250820'
+__version__ = version = '0.11.1.a20250828'
 __version_tuple__ = version_tuple = (0, 11, 1)

{compressed_tensors-0.11.1a20250820.dist-info → compressed_tensors-0.11.1a20250828.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.11.1a20250820
+Version: 0.11.1a20250828
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.