compressed-tensors 0.10.3a20250731__py3-none-any.whl → 0.10.3a20250806__py3-none-any.whl

compressed_tensors/quantization/lifecycle/forward.py

@@ -124,8 +124,13 @@ def dequantize(
                     strategy=QuantizationStrategy.GROUP, group_size=group_size
                 )
             else:
+                rows, cols = x_q.shape[-2], x_q.shape[-1]
+                block_height = rows // scale.shape[0]  # Rows per block
+                block_width = cols // scale.shape[1]  # Columns per block
+
                 args = QuantizationArgs(
-                    strategy=QuantizationStrategy.BLOCK, block_structure=scale.shape
+                    strategy=QuantizationStrategy.BLOCK,
+                    block_structure=[block_height, block_width],
                 )
         else:
             raise ValueError(

compressed_tensors/transform/factory/base.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 
 from abc import ABC, abstractmethod
-from typing import Optional
+from collections import defaultdict
+from typing import List, Optional, Tuple, Set
 
 import torch
 import torch.nn.utils.parametrize as P
@@ -49,10 +50,13 @@ class TransformFactory(RegistryMixin, ABC):
     :param seed: random seed used to transform weight randomization
     """
 
+    transforms: List["TransformBase"]
+
     def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
         self.name = name
         self.scheme = scheme
         self.generator = torch.Generator()
+        self.transforms = list()
         if seed is not None:
             self.generator.manual_seed(seed)
 
@@ -90,6 +94,8 @@ class TransformFactory(RegistryMixin, ABC):
             for _, module in match_named_modules(model, arg.targets, arg.ignore):
                 self._apply_to_module(module, arg)
 
+        self._update_tied_weights()
+
     def _apply_to_module(self, module: Module, args: TransformArgs):
         """
         Create transforms and apply them to the module
@@ -97,9 +103,17 @@ class TransformFactory(RegistryMixin, ABC):
         :param module: target module to apply transforms to
         :param args: defines how the transform will be applied to the target module
         """
+        if has_offloaded_params(module):
+            if module._hf_hook.place_submodules:
+                raise NotImplementedError(
+                    "Applying transforms to offloaded submodules with "
+                    "`place_submodules=True` is not supported"
+                )
+
         # create transform as submodule
         transform_name = f"{self.name}_{args.location}"
         transform = self.create_transform(module, args)
+        self.transforms.append(transform)
         register_offload_module(module, transform_name, transform)
 
         # register input transformation hook
@@ -128,8 +142,9 @@ class TransformFactory(RegistryMixin, ABC):
                     raise ValueError("Offloaded training is not supported")
                 P.register_parametrization(module, "weight", transform)
 
-            # transform is no longer needed (unfusing is not supported)
-            delete_offload_module(module, transform_name)
+            else:
+                # transform is no longer needed (unfusing is not supported)
+                delete_offload_module(module, transform_name)
 
         # register output transformation hook
         elif args.location == TransformLocation.OUTPUT:
@@ -143,6 +158,31 @@ class TransformFactory(RegistryMixin, ABC):
         else:
             raise NotImplementedError()
 
+    def _update_tied_weights(self):
+        """
+        Populate the `_dynamic_tied_weights_keys` attribute of transforms,
+        which is used by transformers to detect and remove shared pointers
+        during saving
+        """
+        # map from data_ptrs to keys
+        ptr_to_keys: dict[int, List[Tuple[TransformBase, str]]] = defaultdict(list)
+        for transform in self.transforms:
+            for name, param in transform.named_parameters(recurse=False):
+                # NOTE: previously asserted that parent._hf_hook.place_submodules=False
+                if has_offloaded_params(transform):
+                    param = transform._hf_hook.weights_map[name]
+                ptr_to_keys[param.data_ptr()].append((transform, name))
+
+        # populate `_dynamic_tied_weights_keys` if there is more than one key
+        # and ensure that they share tensors
+        for shared_keys in ptr_to_keys.values():
+            if len(shared_keys) > 1:
+                tensor = getattr(shared_keys[0][0], shared_keys[0][1])
+
+                for transform, name in shared_keys:
+                    transform._dynamic_tied_weights_keys.add(name)
+                    setattr(transform, name, tensor)
+
 
 class TransformBase(InternalModule, ABC):
     """
@@ -151,6 +191,11 @@ class TransformBase(InternalModule, ABC):
 
     args: TransformArgs
     weight: Parameter
+    _dynamic_tied_weights_keys: Set[str]
+
+    def __init__(self):
+        super().__init__()
+        self._dynamic_tied_weights_keys = set()
 
     @abstractmethod
     def forward(self, value: Tensor) -> Tensor:

compressed_tensors/transform/factory/matrix_multiply.py

@@ -70,6 +70,7 @@ class RandomMatrixFactory(TransformFactory):
 
     def _create_inverse(self, weight: Parameter) -> Parameter:
         data = high_precision_invert(weight.data)
+        data = data.contiguous()  # ensure proper serialization
         return Parameter(data, requires_grad=False)
 
 

compressed_tensors/utils/match.py

@@ -15,7 +15,7 @@
 import logging
 import re
 from collections.abc import Generator
-from typing import Iterable, Tuple
+from typing import Iterable, Mapping, Optional, Tuple
 
 import torch
 from compressed_tensors.utils.internal import InternalModule
@@ -32,10 +32,14 @@ __all__ = [
 ]
 
 
+FusedMappping = Mapping[str, Iterable[str]]
+
+
 def match_named_modules(
     model: torch.nn.Module,
     targets: Iterable[str],
     ignore: Iterable[str] = tuple(),
+    fused: Optional[FusedMappping] = None,
     warn_on_fail: bool = False,
 ) -> Generator[Tuple[str, torch.nn.Module]]:
     """
@@ -45,16 +49,18 @@ def match_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
+    :fused: optional mapping from suffixes of fused modules to the suffixes of their
+        corresponding shards. See `compressed_tensors.utils.match.is_match`
     :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):
+            if is_match(name, module, target, fused):
                 unmatched_targets -= {target}
 
-                if not any(is_match(name, module, ign) for ign in ignore):
+                if not any(is_match(name, module, ign, fused) for ign in ignore):
                     yield name, module
 
     if warn_on_fail:
@@ -68,6 +74,7 @@ def match_named_parameters(
     model: torch.nn.Module,
     targets: Iterable[str],
     ignore: Iterable[str] = tuple(),
+    fused: Optional[FusedMappping] = None,
     warn_on_fail: bool = False,
 ) -> Generator[Tuple[str, torch.nn.Module, torch.nn.Parameter]]:
     """
@@ -77,6 +84,8 @@ def match_named_parameters(
     :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
+    :fused: optional mapping from suffixes of fused modules to the suffixes of their
+        corresponding shards. See `compressed_tensors.utils.match.is_match`
     :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
     """
@@ -88,10 +97,10 @@ def match_named_parameters(
         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):
+                if _match_name(param_fqn, target, fused):
                     unmatched_targets -= {target}
 
-                    if not any(_match_name(param_fqn, ign) for ign in ignore):
+                    if not any(_match_name(param_fqn, ign, fused) for ign in ignore):
                         yield param_fqn, module, param
 
     if warn_on_fail:
@@ -164,21 +173,56 @@ def match_modules_set(
         raise ValueError(f"Unable to match targets into set: {unmatched_keys}")
 
 
-def is_match(name: str, module: torch.nn.Module, target: str) -> bool:
+def is_match(
+    name: str,
+    module: torch.nn.Module,
+    target: str,
+    fused: Optional[FusedMappping] = None,
+) -> bool:
     """
     Returns true if either module name or module parent classes match against target
-    and the module is not an internal module
+    and the module is not an internal module. The name and module may refer to a fused
+    module defined by vLLM. In these cases, a `fused` mapping must be provided.
+
+    For example, in `vllm/model_executor/models/llama.py`:
+    ```python
+    packed_modules_mapping = {
+        "qkv_proj": ["q_proj", "k_proj", "v_proj"],
+        "gate_up_proj": ["gate_proj", "up_proj"]
+    }
+    ```
+
+    :param name: name of module
+    :param module: module to match
+    :param target: target which matches name or module, potentially contains regex
+    :fused: optional mapping from suffixes of fused modules to the suffixes of their
+        corresponding shards
     """
     return not isinstance(module, InternalModule) and (
-        _match_name(name, target) or _match_class(module, target)
+        _match_name(name, target, fused) or _match_class(module, target)
     )
 
 
-def _match_name(name: str, target: str) -> bool:
+def _match_name(name: str, target: str, fused: Optional[FusedMappping] = None) -> bool:
     """
-    Returns true if target string begins with "re:" and
-    regex matches or if target string exactly matches name
+    Returns true if target string begins with "re:" and regex matches or if target
+    string exactly matches name. If the name refers to a fused module defined by vLLM,
+    a `fused` mapping must be provided.
+
+    :param name: name of module
+    :param target: target name, potentially contains regex
+    :fused: optional mapping from suffixes of fused modules to the suffixes of their
+        corresponding shards
     """
+    if fused is not None:
+        for fused_suffix in fused:
+            if name.endswith(fused_suffix):
+                name_stripped = name.removesuffix(fused_suffix)
+                return any(
+                    _match_name(name_stripped + shard_suffix, target)
+                    for shard_suffix in fused[fused_suffix]
+                )
+
     if target.startswith("re:"):
         return re.match(target.removeprefix("re:"), name) is not None
     else:
@@ -187,10 +231,20 @@ def _match_name(name: str, target: str) -> bool:
 
 def _match_class(module: torch.nn.Module, target: str) -> bool:
     """
-    Returns true if any torch parent class names match the target string exactly
+    Returns true if any torch parent class names match the target string exactly.
+    A special exception is made for vllm's `LinearBase` class which matches `Linear`
+
+    :param module: module to match
+    :param target: target which matches name or module
     """
     # 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
+        (
+            issubclass(cls, torch.nn.Module)
+            and (
+                cls.__name__ == target
+                or (cls.__name__ == "LinearBase" and target == "Linear")
+            )
+        )
         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.a20250731'
+__version__ = version = '0.10.3.a20250806'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250731.dist-info → compressed_tensors-0.10.3a20250806.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250731
+Version: 0.10.3a20250806
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250731.dist-info → compressed_tensors-0.10.3a20250806.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=cuOuj6FL5GE-iPKjLVFuRjlwW0_6uDC3tDxFkkHyXFg,523
+compressed_tensors/version.py,sha256=AuoKIjSgjjAcZIPZe3HN5zhNJ7enhDAjwQrqUHPg76o,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
@@ -32,7 +32,7 @@ compressed_tensors/quantization/quant_scheme.py,sha256=xk2LPn18tjS1PEOyf0WKvavBq
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
 compressed_tensors/quantization/lifecycle/apply.py,sha256=wM8mVcbKvZjBo18pSXMp28i30YWwUXJPSS7_HCakH9U,17892
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
-compressed_tensors/quantization/lifecycle/forward.py,sha256=lQwibkDGroJqONhP9ATZWwaZF9suPmCZMQEagFlFc94,17329
+compressed_tensors/quantization/lifecycle/forward.py,sha256=HzfoRkK3CkEHuCqRWatq0kyu5sFx8ULZHNmmjRNIpWI,17571
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
 compressed_tensors/quantization/lifecycle/initialize.py,sha256=BM7bR_uNa-Ex4T-roHonWiRaxCi5sFysXyl0cFh1ZVs,10257
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
@@ -45,9 +45,9 @@ 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=Zplf8QO-mFqGwDEhLdYL_afSu7v4nMa79oNhidRNPvY,5880
+compressed_tensors/transform/factory/base.py,sha256=NJ3lI95tJk6gHOeZEVheQ_Ae7NHhhUG_9FHXu613x30,7740
 compressed_tensors/transform/factory/hadamard.py,sha256=B0BVjbF3y707MO6L2XfEoZJTQU965vU9dUPLOiUSXII,4193
-compressed_tensors/transform/factory/matrix_multiply.py,sha256=LdoV2E12HTucmUWcw7UKOpRNnL8QhOOIUnNVlpOpGiI,3925
+compressed_tensors/transform/factory/matrix_multiply.py,sha256=kCB7cfM_PCgJDyyhg2d1rKTEiyuscwzhprXY7VfIx6E,3989
 compressed_tensors/transform/factory/random_hadamard.py,sha256=nUhTlFa4ikSpcl4Umme71pnjMPgwYoGlwjKlU27UHZ4,1634
 compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/transform/utils/hadamard.py,sha256=hDJZC0Gw2fKdxqa3f8TmFc5J0eJqxHtFRxswLU_yVJc,5548
@@ -56,14 +56,14 @@ compressed_tensors/transform/utils/matrix.py,sha256=FIHCUlpWVIIhdr3c6EbQec41JeiP
 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=ZVBPzrGYExq7-6RRUlU5XeCjl0ooLaNUoDO6Cgnn9cY,7220
+compressed_tensors/utils/match.py,sha256=9x-yZIlq7ndSLf2aQwNT7IpBQDe-8H6utiJkji8wPrQ,9397
 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-0.10.3a20250731.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.3a20250731.dist-info/METADATA,sha256=1NCpfVbLTf6aGJ38rJz3Lmu9DptHpuYm5vTRxIB9PB8,7031
-compressed_tensors-0.10.3a20250731.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.3a20250731.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.3a20250731.dist-info/RECORD,,
+compressed_tensors-0.10.3a20250806.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.3a20250806.dist-info/METADATA,sha256=e8DIx-6UDn2Wj7fGLEBgVru2k9Tme9dOPgxS_ciZDcw,7031
+compressed_tensors-0.10.3a20250806.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.3a20250806.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.3a20250806.dist-info/RECORD,,