compressed-tensors 0.10.3a20250728__py3-none-any.whl → 0.10.3a20250805__py3-none-any.whl

compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py

@@ -61,6 +61,27 @@ class NVFP4PackedCompressor(BaseQuantizationCompressor):
             "weight_global_scale",
         )
 
+    def compression_param_info(
+        self,
+        weight_shape: torch.Size,
+        quantization_args: Optional[QuantizationArgs] = None,
+    ) -> Dict[str, Tuple[torch.Size, torch.dtype]]:
+        """
+        Creates a dictionary of expected shapes and dtypes for each compression
+            parameter used by the compressor
+
+        :param weight_shape: uncompressed weight shape
+        :param quantization_args: quantization parameters for the weight
+        :return: dictionary mapping compressed parameter names to shape and dtype
+        """
+        output = {
+            "weight_packed": (
+                torch.Size((weight_shape[0], weight_shape[1] // 2)),
+                torch.uint8,
+            ),
+        }
+        return output
+
     def compress_weight(
         self,
         weight: Tensor,

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(
@@ -257,13 +262,10 @@ def _process_quantization(
         QuantizationStrategy.GROUP,
         QuantizationStrategy.TENSOR_GROUP,
     ):
-        n_dims = x.shape
-        if len(n_dims) > 2:
-            x = x.squeeze(0)
 
         output_dtype = dtype if dtype is not None else x.dtype
         output = torch.zeros_like(x).to(output_dtype)
-        columns = output.shape[1]
+        columns = output.shape[-1]
 
         # TODO: make validation step for inputs
 
@@ -293,14 +295,12 @@ def _process_quantization(
             perm = torch.argsort(g_idx)
             x = safe_permute(x, perm, dim=1)
 
-        x = torch.reshape(
-            x,
-            (
-                x.shape[0],
-                ceil(x.shape[1] / group_size),
-                group_size,
-            ),
+        # Maintain all dimensions apart from the last dim, which is divided by the group_size
+        reshaped_dims = (
+            ceil(x.shape[-1] / group_size),
+            group_size,
         )
+        x = x.unflatten(-1, reshaped_dims)
 
         if do_quantize:
             output = _quantize(
@@ -323,19 +323,12 @@ def _process_quantization(
                 global_scale=global_scale,
             )
 
-        output = torch.reshape(
-            output,
-            (output.shape[0], output.shape[1] * output.shape[2]),
-        )
-
+        output = output.flatten(start_dim=-2)
         output = output.to(output_dtype)
 
         if not is_column_order:
             output = safe_permute(output, torch.argsort(perm), dim=1)
 
-        if len(n_dims) > 2:
-            output = output.unsqueeze(0)
-
     else:  # covers channel, token and tensor strategies
         if do_quantize:
             output = _quantize(

compressed_tensors/quantization/utils/helpers.py

@@ -175,20 +175,16 @@ def compute_dynamic_scales_and_zp(
         QuantizationStrategy.TENSOR_GROUP,
         QuantizationStrategy.GROUP,
     ):
-        if len(value.shape) > 2:
-            value = value.squeeze(0)
 
-        dim = {0, 1}
-        reduce_dims = tuple(idx for idx in range(3) if idx not in dim)
+        reduce_dims = -1
         keep_dims = False
-        value = torch.reshape(
-            value,
-            (
-                value.shape[0],
-                math.ceil(value.shape[1] / args.group_size),
-                args.group_size,
-            ),
+
+        reshaped_dims = (
+            math.ceil(value.shape[-1] / args.group_size),
+            args.group_size,
         )
+        value = value.unflatten(-1, reshaped_dims)
+
     else:
         supported_strategies = (
             QuantizationStrategy.TOKEN,

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.a20250728'
+__version__ = version = '0.10.3.a20250805'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250728.dist-info → compressed_tensors-0.10.3a20250805.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250728
+Version: 0.10.3a20250805
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250728.dist-info → compressed_tensors-0.10.3a20250805.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=EY3NpvLIsm31BPA-e32djbQIUYdm3sP8W28lHH72d0Y,523
+compressed_tensors/version.py,sha256=UcH3DkUtSV6xgd1l5QTWXLV_iWa7GzNrCWIOpZvkzkE,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
@@ -9,7 +9,7 @@ compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=e-2n
 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
-compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=Gw-lVzk5jrKUlM5UTCiJBmhM5gHzB9mn8r298MVUbDI,6395
+compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=tKEaYom4SdMwZWg4MDMMMLNGTLgcVT20lPzewboVpMM,7145
 compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=47W1hFTi5YHVNKEWptzztsSutwI1kxy2Troh-NW1y14,11244
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
 compressed_tensors/compressors/sparse_compressors/base.py,sha256=YNZWcHjDleAlqbgRZQ6oJf44MQb_UDNvJGOqhl26uFA,8098
@@ -32,11 +32,11 @@ 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=V98jWzb3rfV91EC6kfzAyXtmnbLjNF01Rd_EHU2bLo8,17506
+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
-compressed_tensors/quantization/utils/helpers.py,sha256=Je96Wai9SOizbdE5ph0nsJ86zS96lE4fkf_9q9o2tpA,17212
+compressed_tensors/quantization/utils/helpers.py,sha256=7a89X0kg6xDGplw6trOrkRQzMRPu-txY_qvEt07Vcgc,17036
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=0s15BxdGgzBv8RL4kUJCYcuDOFUh_KZYvNvLEeRqWTc,11956
 compressed_tensors/transform/__init__.py,sha256=v2wfl4CMfA6KbD7Hxx_MbRev63y_6QLDlccZq-WTtdw,907
@@ -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-0.10.3a20250728.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.3a20250728.dist-info/METADATA,sha256=rQbbrFahVspKPEfY86EpebdjgoYAtSyyH7JLOPTPcrg,7031
-compressed_tensors-0.10.3a20250728.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.3a20250728.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.3a20250728.dist-info/RECORD,,
+compressed_tensors-0.10.3a20250805.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.3a20250805.dist-info/METADATA,sha256=8SpvZ9SNB_DGL6L4I8QrtLczHtxI17ezOlwf6Ew_4R8,7031
+compressed_tensors-0.10.3a20250805.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.3a20250805.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.3a20250805.dist-info/RECORD,,