compressed-tensors 0.10.3a20250721__tar.gz → 0.10.3a20250728__tar.gz

{compressed_tensors-0.10.3a20250721/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250728}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250721
+Version: 0.10.3a20250728
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -400,7 +400,10 @@ class ModelCompressor:
 
                 # in the future, support compression on same device
                 with align_module_device(module, execution_device=exec_device):
-                    state_dict = module.state_dict(prefix=f"{prefix}.")
+                    state_dict = {
+                        f"{prefix}.{name}": param
+                        for name, param in module.named_parameters(recurse=False)
+                    }
 
                 # quantization first
                 if prefix in module_to_scheme:
@@ -421,7 +424,7 @@ class ModelCompressor:
 
                 # remove any existing parameters
                 offload_device = get_offloaded_device(module)
-                for name, _ in list(module.named_parameters()):
+                for name, _ in list(module.named_parameters(recurse=False)):
                     delete_offload_parameter(module, name)
 
                 # replace with compressed parameters
@@ -458,7 +461,10 @@ class ModelCompressor:
             if prefix in module_to_scheme or prefix in sparse_compression_targets:
                 # in the future, support decompression on same device
                 with align_module_device(module, execution_device="cpu"):
-                    state_dict = module.state_dict(prefix=f"{prefix}.")
+                    state_dict = {
+                        f"{prefix}.{name}": param
+                        for name, param in module.named_parameters(recurse=False)
+                    }
 
                 # sparsity first
                 if prefix in sparse_compression_targets:
@@ -483,7 +489,7 @@ class ModelCompressor:
                 # remove any existing parameters
                 exec_device = get_execution_device(module)
                 offload_device = get_offloaded_device(module)
-                for name, _ in list(module.named_parameters()):
+                for name, _ in list(module.named_parameters(recurse=False)):
                     delete_offload_parameter(module, name)
 
                 # replace with decompressed parameters
@@ -754,8 +760,8 @@ def map_module_to_scheme(model: Module) -> Dict[str, QuantizationScheme]:
         fix_fsdp_module_name(name): module.quantization_scheme
         for name, module in model.named_modules()
         if (
-            hasattr(module, "quantization_scheme") and
-            module.quantization_scheme.weights is not None
+            hasattr(module, "quantization_scheme")
+            and module.quantization_scheme.weights is not None
         )
     }
 

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/quantization/lifecycle/forward.py

@@ -112,17 +112,21 @@ def dequantize(
             if scale.shape[1] == 1:
                 args = QuantizationArgs(strategy=QuantizationStrategy.CHANNEL)
             # Scale height matches input or is 1 -> group quantization across columns
-            # 
+            #
             # Example 1: scale.shape[0] == 1
             # x_q: (4, 8), scale: (1, 4) -> 2 columns per group
             #
-            # Example 2: scale.shape[0] == x_q.shape[0] 
+            # Example 2: scale.shape[0] == x_q.shape[0]
             # x_q: (4, 8), scale: (4, 4) -> 2 elements per group (per row)
             elif (scale.shape[0] == 1) or (scale.shape[0] == x_q.shape[0]):
                 group_size = int(x_q.shape[1] / scale.shape[1])
-                args = QuantizationArgs(strategy=QuantizationStrategy.GROUP, group_size=group_size)
+                args = QuantizationArgs(
+                    strategy=QuantizationStrategy.GROUP, group_size=group_size
+                )
             else:
-                args = QuantizationArgs(strategy=QuantizationStrategy.BLOCK, block_structure=scale.shape)
+                args = QuantizationArgs(
+                    strategy=QuantizationStrategy.BLOCK, block_structure=scale.shape
+                )
         else:
             raise ValueError(
                 f"Could not infer a quantization strategy from scale with {scale.ndim} "

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/quantization/lifecycle/initialize.py

@@ -15,6 +15,7 @@
 
 import logging
 import math
+import warnings
 from enum import Enum
 from typing import List, Optional
 
@@ -172,14 +173,43 @@ def _initialize_scale_zero_point(
 
     if base_name == "weight" and weight_shape is not None:
         if quantization_args.strategy == QuantizationStrategy.CHANNEL:
-            # (output_channels, 1)
+            # (output_channels, 1) - only for weights
             expected_shape = (weight_shape[0], 1)
         elif quantization_args.strategy in (
             QuantizationStrategy.TENSOR_GROUP,
             QuantizationStrategy.GROUP,
         ):
+            # GROUP/TENSOR_GROUP for both weights and activations
             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
+            if quantization_args.block_structure is None:
+                raise ValueError(
+                    "Block quantization requires block_structure to be specified"
+                )
+            block_height, block_width = quantization_args.block_structure
+            rows, cols = weight_shape[-2], weight_shape[-1]
+            num_rows_blocks = math.ceil(rows / block_height)
+            num_cols_blocks = math.ceil(cols / block_width)
+
+            # 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.",
+                    UserWarning,
+                )
+
+            expected_shape = (num_rows_blocks, num_cols_blocks)
+    elif quantization_args.strategy == QuantizationStrategy.BLOCK:
+        warnings.warn(
+            f"BLOCK quantization not supported for {base_name} activations. "
+            f"Falling back to tensor-level quantization.",
+            UserWarning,
+        )
+        expected_shape = 1
 
     # 3. Identify quantization scale and zp dtype
     scale_dtype = scale_dtype if scale_dtype is not None else module.weight.dtype
@@ -189,7 +219,12 @@ def _initialize_scale_zero_point(
     else:
         # TODO: consider erroring out in the future as if the dtype if not one of these,
         # there is likely bug
-        if scale_dtype not in [torch.float16, torch.bfloat16, torch.float32, torch.float64]:
+        if scale_dtype not in [
+            torch.float16,
+            torch.bfloat16,
+            torch.float32,
+            torch.float64,
+        ]:
             scale_dtype = torch.float16
         zp_dtype = quantization_args.pytorch_dtype()
 

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/quantization/quant_scheme.py

@@ -64,8 +64,9 @@ class QuantizationScheme(BaseModel):
                 raise ValueError("Cannot apply actorder to output activations")
 
         if (
-            inputs and weights
-            and weights.strategy == QuantizationStrategy.GROUP 
+            inputs
+            and weights
+            and weights.strategy == QuantizationStrategy.GROUP
             and inputs.strategy == QuantizationStrategy.GROUP
             and weights.group_size != inputs.group_size
         ):
@@ -75,7 +76,7 @@ class QuantizationScheme(BaseModel):
                 "may complicate fused kernel implementations. Consider using "
                 "TENSOR_GROUP strategy for both or matching group sizes.",
                 UserWarning,
-                stacklevel=2
+                stacklevel=2,
             )
 
         return model

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/transform/factory/base.py

@@ -18,7 +18,6 @@ from typing import Optional
 import torch
 import torch.nn.utils.parametrize as P
 from compressed_tensors import InternalModule
-from compressed_tensors.quantization.lifecycle import is_target  # TODO: move to utils
 from compressed_tensors.registry.registry import RegistryMixin, T
 from compressed_tensors.transform import (
     TransformArgs,
@@ -29,6 +28,7 @@ from compressed_tensors.utils import (
     align_module_device,
     delete_offload_module,
     has_offloaded_params,
+    match_named_modules,
     patch_attr,
     register_offload_module,
     update_offload_parameter,
@@ -87,9 +87,8 @@ class TransformFactory(RegistryMixin, ABC):
         :param model: module to apply transforms to
         """
         for arg in self.scheme.apply:
-            for name, module in list(model.named_modules()):
-                if is_target(name, module, arg.targets, arg.ignore):
-                    self._apply_to_module(module, arg)
+            for _, module in match_named_modules(model, arg.targets, arg.ignore):
+                self._apply_to_module(module, arg)
 
     def _apply_to_module(self, module: Module, args: TransformArgs):
         """

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/transform/factory/hadamard.py

@@ -12,9 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import math
 from typing import Optional, Union
 
-import math
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
 from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
@@ -103,7 +103,8 @@ class HadamardTransform(TransformBase):
 
         if self.args.inverse:
             weight = weight.T
- 
-        return apply_transform_weight(
-            weight, value, self.args.location, self.module_type
-        ) / self._scale
+
+        return (
+            apply_transform_weight(weight, value, self.args.location, self.module_type)
+            / self._scale
+        )

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/utils/__init__.py

@@ -15,6 +15,7 @@
 
 from .helpers import *
 from .internal import *
+from .match import *
 from .offload import *
 from .permutations_24 import *
 from .permute import *

compressed_tensors-0.10.3a20250728/src/compressed_tensors/utils/match.py

@@ -0,0 +1,196 @@
+# 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
+import re
+from collections.abc import Generator
+from typing import Iterable, Tuple
+
+import torch
+from compressed_tensors.utils.internal import InternalModule
+
+
+_LOGGER: logging.Logger = logging.getLogger(__name__)
+
+
+__all__ = [
+    "match_named_modules",
+    "match_named_parameters",
+    "match_modules_set",
+    "is_match",
+]
+
+
+def match_named_modules(
+    model: torch.nn.Module,
+    targets: Iterable[str],
+    ignore: Iterable[str] = tuple(),
+    warn_on_fail: bool = False,
+) -> Generator[Tuple[str, torch.nn.Module]]:
+    """
+    Yields names and modules which match `targets` but do not match `ignore`.
+    Values are returned in order of `model.named_modules()`
+
+    :param model: model containing submodules to match against
+    :param targets: target strings, potentially containing "re:" prefixes
+    :param ignore: targets to ignore, potentially containing "re:" prefixes
+    :param warn_on_fail: if True, warns if any targets do not match any modules in model
+    :return: generator of module names and modules
+    """
+    unmatched_targets = set(targets)
+    for name, module in model.named_modules():
+        for target in targets:
+            if is_match(name, module, target):
+                unmatched_targets -= {target}
+
+                if not any(is_match(name, module, ign) for ign in ignore):
+                    yield name, module
+
+    if warn_on_fail:
+        for target in unmatched_targets:
+            _LOGGER.warning(
+                f"Could not match `{target}` in instance of {model.__class__.__name__}"
+            )
+
+
+def match_named_parameters(
+    model: torch.nn.Module,
+    targets: Iterable[str],
+    ignore: Iterable[str] = tuple(),
+    warn_on_fail: bool = False,
+) -> Generator[Tuple[str, torch.nn.Module, torch.nn.Parameter]]:
+    """
+    Yields parameters which match `targets` but do not match `ignore`.
+    Values are returned in order of `model.named_modules()`
+
+    :param model: model containing params to match against
+    :param targets: target strings, potentially containing "re:" prefixes
+    :param ignore: targets to ignore, potentially containing "re:" prefixes
+    :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
+    """
+    unmatched_targets = set(targets)
+    for module_name, module in model.named_modules():
+        if isinstance(module, InternalModule):
+            continue
+
+        for param_name, param in module.named_parameters(recurse=False):
+            param_fqn = f"{module_name}.{param_name}"
+            for target in targets:
+                if _match_name(param_fqn, target):
+                    unmatched_targets -= {target}
+
+                    if not any(_match_name(param_fqn, ign) for ign in ignore):
+                        yield param_fqn, module, param
+
+    if warn_on_fail:
+        for target in unmatched_targets:
+            _LOGGER.warning(
+                f"Could not match `{target}` in instance of {model.__class__.__name__}"
+            )
+
+
+def match_modules_set(
+    model: torch.nn.Module,
+    targets: Iterable[str],
+    ignore: Iterable[str] = tuple(),
+) -> 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:
+    ```python3
+    match_modules_set(model, ["q_proj", "k_proj", "v_proj"]) == (
+        (
+            `model.layers.0.self_attn.q_proj`,
+            `model.layers.0.self_attn.k_proj`,
+            `model.layers.0.self_attn.v_proj`,
+        ),
+        (
+            `model.layers.1.self_attn.q_proj`,
+            `model.layers.1.self_attn.k_proj`,
+            `model.layers.1.self_attn.v_proj`,
+        ),
+        ...
+        (
+            `model.layers.32.self_attn.q_proj`,
+            `model.layers.32.self_attn.k_proj`,
+            `model.layers.32.self_attn.v_proj`,
+        ),
+    )
+    ```
+
+    This can be used to match layers to their corresponding downstream counterparts.
+    For example, matching layer norms to their subsequent linear layers
+    ```python3
+    for norm, q, k, v in match_modules_set(model, (norm_tgt, q_tgt, k_tgt, v_tgt)):
+        fuse_norm_linears(norm, [q, k, v])
+
+    :param model: model containing modules to match against
+    :param targets: target strings, potentially containing "re:" prefixes
+    :param ignore: targets to ignore, potentially containing "re:" prefixes
+    """
+    matches = dict.fromkeys(targets, None)
+    for name, module in model.named_modules():
+        # match until we get a full set
+        for target in targets:
+            if is_match(name, module, target) and not any(
+                is_match(name, module, ign) for ign in ignore
+            ):
+                if matches[target] is not None:
+                    raise ValueError(f"Matched a {target} twice before completing set")
+                matches[target] = module
+
+        # once we have a full set, yield and reset
+        if targets and all((matches[target] is not None for target in targets)):
+            yield [matches[target] for target in targets]  # ensure correct ordering
+            matches = dict.fromkeys(targets, None)
+
+    # check that none are left over
+    unmatched_keys = [match for match, value in matches.items() if value is not None]
+    if len(unmatched_keys):
+        raise ValueError(f"Unable to match targets into set: {unmatched_keys}")
+
+
+def is_match(name: str, module: torch.nn.Module, target: str) -> bool:
+    """
+    Returns true if either module name or module parent classes match against target
+    and the module is not an internal module
+    """
+    return not isinstance(module, InternalModule) and (
+        _match_name(name, target) or _match_class(module, target)
+    )
+
+
+def _match_name(name: str, target: str) -> bool:
+    """
+    Returns true if target string begins with "re:" and
+    regex matches or if target string exactly matches name
+    """
+    if target.startswith("re:"):
+        return re.match(target.removeprefix("re:"), name) is not None
+    else:
+        return target == name
+
+
+def _match_class(module: torch.nn.Module, target: str) -> bool:
+    """
+    Returns true if any torch parent class names match the target string exactly
+    """
+    # will never match against a regex pattern since `:` is not allowed in class names
+    return any(
+        issubclass(cls, torch.nn.Module) and cls.__name__ == target
+        for cls in module.__class__.__mro__
+    )

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250721'
+__version__ = version = '0.10.3.a20250728'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250721
+Version: 0.10.3a20250728
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/src/compressed_tensors.egg-info/SOURCES.txt

@@ -88,6 +88,7 @@ src/compressed_tensors/transform/utils/matrix.py
 src/compressed_tensors/utils/__init__.py
 src/compressed_tensors/utils/helpers.py
 src/compressed_tensors/utils/internal.py
+src/compressed_tensors/utils/match.py
 src/compressed_tensors/utils/offload.py
 src/compressed_tensors/utils/permutations_24.py
 src/compressed_tensors/utils/permute.py
@@ -141,6 +142,7 @@ tests/test_transform/factory/test_memory.py
 tests/test_transform/utils/test_hadamard.py
 tests/test_utils/__init__.py
 tests/test_utils/test_helpers.py
+tests/test_utils/test_match.py
 tests/test_utils/test_offload.py
 tests/test_utils/test_safetensors_load.py
 utils/copyright.py

{compressed_tensors-0.10.3a20250721 → compressed_tensors-0.10.3a20250728}/tests/test_quantization/lifecycle/test_initialize.py

@@ -174,8 +174,8 @@ def test_initialize_module_for_quantization_offloaded(
             ),
         ),
         (
-            QuantizationArgs(strategy="block"),
-            QuantizationArgs(strategy="block"),
+            QuantizationArgs(strategy="block", block_structure=[2, 4]),
+            None,
         ),
         (
             QuantizationArgs(strategy="token"),
@@ -227,7 +227,17 @@ def test_initialize_quantization_parameters(weights, input_activations):
             expected_shape = (layer.weight.shape[0], max(num_groups, 1))
 
         elif args.strategy == QuantizationStrategy.BLOCK:
-            expected_shape = (1,)
+            # For block quantization, only weights get block-level scales
+            # Activations fall back to tensor-level since shape is unknown at init
+            if q_type == "weights" and args.block_structure is not None:
+                block_height, block_width = args.block_structure
+                rows, cols = layer.weight.shape[-2], layer.weight.shape[-1]
+                num_rows_blocks = math.ceil(rows / block_height)
+                num_cols_blocks = math.ceil(cols / block_width)
+                expected_shape = (num_rows_blocks, num_cols_blocks)
+            else:
+                # For activations or when block_structure is None
+                expected_shape = (1,)
 
         elif args.strategy == QuantizationStrategy.TOKEN:
             expected_shape = (1, 1)