compressed-tensors 0.10.3a20250721__py3-none-any.whl → 0.10.3a20250724__py3-none-any.whl

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/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 +173,41 @@ 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 +217,12 @@ 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/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/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/utils/match.py

@@ -0,0 +1,191 @@
+# 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
+
+
+_LOGGER: logging.Logger = logging.getLogger(__name__)
+
+
+__all__ = [
+    "match_named_modules",
+    "match_named_parameters",
+    "match_modules_set",
+    "is_match",
+    "match_name",
+    "match_class",
+]
+
+
+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():
+        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
+    """
+    return 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/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.a20250724'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250721.dist-info → compressed_tensors-0.10.3a20250724.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250721
+Version: 0.10.3a20250724
 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.dist-info → compressed_tensors-0.10.3a20250724.dist-info}/RECORD

@@ -1,11 +1,11 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=d45NnXPlUzem78cnmH1dUdhwbxzgKieO5KjFyfBwSZA,523
+compressed_tensors/version.py,sha256=LKiXh8O_XB2unUsk0HmC-_PgpfbOswj5PZqtDHOPnRg,523
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
 compressed_tensors/compressors/model_compressors/__init__.py,sha256=5RGGPFu4YqEt_aOdFSQYFYFDjcZFJN0CsMqRtDZz3Js,666
-compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=MIMgjLK0VJu4bF_CwyvSftFrC5tvIIMfRt6XB0yLjOE,33056
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=e-2nLkLLzbPlm3xubZPv44BIXC16xaHDZmCp4eXCjZU,33316
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=KvaFBL_Q84LxRGJOV035M8OBoCkAx8kOkfphswgkKWk,745
 compressed_tensors/compressors/quantized_compressors/base.py,sha256=YGUMzbxekj_36ChgQnVZN6T8uDjXtGG1zfMIBGBLWco,10354
 compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=0ANDcuD8aXPqTYNPY6GnX9iS6eXJw6P0TzNV_rYS2l8,5369
@@ -34,7 +34,7 @@ compressed_tensors/quantization/lifecycle/apply.py,sha256=wM8mVcbKvZjBo18pSXMp28
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=jT70Mbbu9pH10vu5ALVD7VWGoFdMEUpxmihGrf4frjM,17432
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=D7yxua1zELmsBYlQiJUTiClBOMIe2J0-IrN2d-jLFPk,8653
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=3Vuj1a-Y7f_7QXagG7BAeAPnDGtbWGFJXBATg6eT-O0,10241
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
 compressed_tensors/quantization/utils/helpers.py,sha256=Je96Wai9SOizbdE5ph0nsJ86zS96lE4fkf_9q9o2tpA,17212
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
@@ -45,7 +45,7 @@ compressed_tensors/transform/transform_args.py,sha256=jJY-Qt996w45LWQ10AHd7tUtNr
 compressed_tensors/transform/transform_config.py,sha256=A3RuLNDqBNEByQNeu40Kg7sItwE6kWgnX18Umg1uONI,2128
 compressed_tensors/transform/transform_scheme.py,sha256=uGLC4avdbhrVqNC3-Eo0p7WzNRQK92Fpg0N9hWiuCRQ,1752
 compressed_tensors/transform/factory/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
-compressed_tensors/transform/factory/base.py,sha256=w9ic5eSxfSNn2Xju-xZvG4_iXAIsJCU56qik8w---aI,5994
+compressed_tensors/transform/factory/base.py,sha256=Zplf8QO-mFqGwDEhLdYL_afSu7v4nMa79oNhidRNPvY,5880
 compressed_tensors/transform/factory/hadamard.py,sha256=iJ2OyKitR2Duw0z5Jqj69GTih2C1WtHRXQCTtATaTtw,4180
 compressed_tensors/transform/factory/matrix_multiply.py,sha256=LdoV2E12HTucmUWcw7UKOpRNnL8QhOOIUnNVlpOpGiI,3925
 compressed_tensors/transform/factory/random_hadamard.py,sha256=nUhTlFa4ikSpcl4Umme71pnjMPgwYoGlwjKlU27UHZ4,1634
@@ -53,16 +53,17 @@ compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh
 compressed_tensors/transform/utils/hadamard.py,sha256=hDJZC0Gw2fKdxqa3f8TmFc5J0eJqxHtFRxswLU_yVJc,5548
 compressed_tensors/transform/utils/hadamards.safetensors,sha256=mFd1GzNodGG-ifA1IoH-0nHYzfraCOvrq_dX2zFI1B4,1436901
 compressed_tensors/transform/utils/matrix.py,sha256=FIHCUlpWVIIhdr3c6EbQec41JeiPAAjCM9Ejz77wb-w,6181
-compressed_tensors/utils/__init__.py,sha256=QFQzF6MpV3yStajPzYktZkmvZsxvfpKUZq2oGbd1Cvw,832
+compressed_tensors/utils/__init__.py,sha256=KZctuotCmX4byXhwDvSeXgp-Ny_awpziAX-WUkZfodI,853
 compressed_tensors/utils/helpers.py,sha256=Q3iRAa2XSdmmn4vSpUplnvKOmWwn4Clao9ZkPBHXtpI,12604
 compressed_tensors/utils/internal.py,sha256=7SSWgDoNFRnlfadwkoFhLW-T2jOc7Po_WzWv5h32Sa8,982
+compressed_tensors/utils/match.py,sha256=DjqTH-J9-E7ULVXPLV-HBRhdi07JhK-H90PbFK-DRAY,7017
 compressed_tensors/utils/offload.py,sha256=3XiBuWbUkBAt8v1t5i57qDcbB3VJQs_FDeayi-JzIWg,23896
 compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.10.3a20250721.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.3a20250721.dist-info/METADATA,sha256=SZk9s5bqtGXwCSxZgsD1tg5GrTH4M8x0d8A3WQkrtlw,7031
-compressed_tensors-0.10.3a20250721.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.3a20250721.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.3a20250721.dist-info/RECORD,,
+compressed_tensors-0.10.3a20250724.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.3a20250724.dist-info/METADATA,sha256=ZH66sWeKBfvuLUe-ArnII1LYXG3UAEYUt6D6YPQ_W-M,7031
+compressed_tensors-0.10.3a20250724.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.3a20250724.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.3a20250724.dist-info/RECORD,,