compressed-tensors 0.10.2a20250612__py3-none-any.whl → 0.10.2a20250616__py3-none-any.whl

compressed_tensors/transform/factory/hadamard.py

@@ -59,7 +59,7 @@ class HadamardFactory(TransformFactory):
         return HadamardTransform(weight, args)
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = deterministic_hadamard_matrix(size)
+        data = deterministic_hadamard_matrix(size, dtype, device)
         data = data.to(dtype=dtype, device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 

compressed_tensors/transform/factory/random_hadamard.py

@@ -29,6 +29,6 @@ class RandomHadamardFactory(HadamardFactory):
     """
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = random_hadamard_matrix(size, self.generator)
+        data = random_hadamard_matrix(size, dtype, device, self.generator)
         data = data.to(dtype=dtype, device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)

compressed_tensors/transform/utils/hadamard.py

@@ -13,95 +13,133 @@
 # limitations under the License.
 
 import math
-from typing import Optional, Tuple
+from pathlib import Path
+from typing import Optional
 
-import numpy
 import torch
+from safetensors import safe_open
 
 
-__all__ = ["random_hadamard_matrix", "deterministic_hadamard_matrix"]
+REPO_PATH = Path(__file__).parent / "hadamards.safetensors"
 
-# adapted from:
-# https://github.com/scipy/scipy/blob/v1.15.2/scipy/linalg/_special_matrices.py
-def deterministic_hadamard_matrix(size: int) -> torch.Tensor:
+
+__all__ = ["random_hadamard_matrix", "deterministic_hadamard_matrix", "is_pow2"]
+
+
+# note that hadamard matrix multiplication can be accelerated using a library such as
+# https://github.com/Dao-AILab/fast-hadamard-transform/tree/master
+
+
+def deterministic_hadamard_matrix(
+    size: int,
+    dtype: torch.dtype = torch.bfloat16,
+    device: torch.device = torch.device("cpu"),
+) -> torch.Tensor:
     """
     Construct an n-by-n Hadamard matrix, using Sylvester's construction.
     `n` must be a power of 2.
 
+    Adapated from https://github.com/scipy/scipy/blob/v1.15.2/scipy/linalg/_special_matrices.py  # noqa: E501
+
     :param size: order of the matrix, must be a power of 2
+    :param dtype: data type of matrix
+    :param device: device to construct matrix on
     :return: hadamard matrix of size `size`
     """
     if size <= 0:
         raise ValueError("Cannot construct deterministic hadamard of size <= 0")
 
-    log2 = int(math.log(size, 2))
+    log2 = int(math.log2(size))
     if size != 2**log2:
         raise ValueError("Cannot construct deterministic hadamard of size != 2^n")
 
-    H = numpy.array([[1]], dtype=int)
+    H = torch.tensor([[1]], dtype=dtype, device=device)
 
     # Sylvester's construction
-    for i in range(0, log2):
-        H = numpy.vstack((numpy.hstack((H, H)), numpy.hstack((H, -H))))
-
-    return torch.from_numpy(H / math.sqrt(size))
+    for _ in range(log2):
+        H = torch.vstack((torch.hstack((H, H)), torch.hstack((H, -H))))
 
-
-# adapted from:
-# https://github.com/facebookresearch/SpinQuant/blob/main/utils/hadamard_utils.py
-
-# TODO: the following library exists for online rotations and should be considered
-# in the future:
-# https://github.com/Dao-AILab/fast-hadamard-transform/tree/master
+    return H / math.sqrt(size)
 
 
 def random_hadamard_matrix(
-    size: int, gen: Optional[torch.Generator] = None
+    size: int,
+    dtype: torch.dtype = torch.bfloat16,
+    device: torch.device = torch.device("cpu"),
+    gen: Optional[torch.Generator] = None,
 ) -> torch.Tensor:
     """
-    Produces a randomly generated Hadamard matrix.
-    See https://cornell-relaxml.github.io/quip-sharp/ ,
-    Section "Randomized Hadamard Transformation"
+    Produces a randomly generated Hadamard matrix. Differs from
+    `deterministic_hadamard_matrix` in that this function supports non powers of 2
+    and randomization using a seeded generator
+
+    Adapated from https://github.com/facebookresearch/SpinQuant/blob/main/utils/hadamard_utils.py  # noqa: E501
+    Known matrices were retrieved from N. J. A. Sloane's Library of Hadamard Matrices http://www.neilsloane.com/hadamard/  # noqa: E501
 
     :param size: The dimension of the hamadard matrix
+    :param dtype: data type of matrix
+    :param device: device to construct matrix on
     :param gen: Optional generator random values
     :return: randomly generated hadamard matrix
     """
-    # Benefits: support other shapes / non powers of 2, support randomization
-    Q = torch.randint(low=0, high=2, size=(size,), generator=gen, dtype=torch.float64)
+    Q = torch.randint(low=0, high=2, size=(size,), generator=gen, dtype=dtype)  # cpu
+    Q = Q.to(device=device)
     Q = Q * 2 - 1
     Q = torch.diag(Q)
     return _matmul_hadU(Q) / math.sqrt(size)
 
 
-def _get_hadK(n: int, transpose: bool = False) -> Tuple[torch.Tensor, int]:
-    # NOTE: we can easily extend the list of supported shapes/sizes
-    # by adding to these methods
-    hadK, K = None, None
-    if n % 20 == 0:
-        assert _is_pow2(n // 20)
-        K = 20
-        hadK = _get_had20().T if transpose else _get_had20()
-    elif n % 12 == 0:
-        assert _is_pow2(n // 12)
-        K = 12
-        hadK = _get_had12().T if transpose else _get_had12()
-    else:
-        assert _is_pow2(n)
-        K = 1
+def is_pow2(n: int) -> bool:
+    """
+    Check if a number is a power of 2
 
-    return hadK, K
+    :param n: number to check
+    :return: True iff `n` is a power of 2
+    """
+    return n > 0 and (n & (n - 1) == 0)
+
+
+def _fetch_hadamard_divisor(
+    n: int,
+    dtype: torch.dtype,
+    device: torch.device = torch.device("cpu"),
+    file_path: str = REPO_PATH,
+) -> Optional[torch.Tensor]:
+    """
+    Fetch a known hadamard matrix from the given file path. The returned matrix will
+    be of of size `k` such that `n / k` is a power of two. Return None if no such
+    matrix exists.
 
+    Note: This function reopens the safetensors file every time it is called.
+    This is technically inefficient, but a very small runtime cost and simpler
+    than forcing callers to manage the file open context
+
+    :param n: size of known hadamard matrix
+    :return: a known hadamard matrix of size `n` if one exists, else None
+    """
+    with safe_open(file_path, framework="pt", device=str(device)) as file:
+        divisors = sorted((int(key) for key in file.keys()), reverse=True)
+        for divisor in divisors:
+            if n % divisor == 0 and is_pow2(n // divisor):
+                return file.get_tensor(str(divisor)).to(dtype=dtype)
+
+    return None
+
+
+def _matmul_hadU(X: torch.Tensor) -> torch.Tensor:
+    size = X.size(0)
+    dtype = X.dtype
+    device = X.device
 
-def _matmul_hadU(X, transpose=False) -> torch.Tensor:
-    n = X.shape[-1]
     # Check if we have the determined hadamard matrix
-    hadK, K = _get_hadK(n, transpose)
+    hadK = _fetch_hadamard_divisor(size, dtype, device=device)
+    if hadK is None:
+        raise ValueError(f"Cannot construct random hadamard matrix of size {size}")
+    K = hadK.size(0)
+
     # Reshape diag matrix with randomized -1/+1
-    input = X.clone().view(-1, n, 1)
+    input = X.clone().view(-1, size, 1)
     output = input.clone()
-
-    # for cases when hadK is not predetermined, determine hadamard matrix
     while input.shape[1] > K:
         input = input.view(input.shape[0], input.shape[1] // 2, 2, input.shape[2])
         output = output.view(input.shape)
@@ -109,53 +147,14 @@ def _matmul_hadU(X, transpose=False) -> torch.Tensor:
         output[:, :, 1, :] = input[:, :, 0, :] - input[:, :, 1, :]
         output = output.view(input.shape[0], input.shape[1], -1)
         (input, output) = (output, input)
+    assert input.shape[1] == K
     del output
 
-    # K == 1 when hadK is None; this happens when the size dim (n)
-    # is not comaptible with any of the maintained hadamard matrices
-
-    if K > 1:
-        # Do not explicitly repeat - OOM
-        # input = torch.bmm(
-        #     hadK.repeat(len(input), 1, 1).to(input.device).to(input.dtype), input)
-        # Use bcast instead
-
-        # for cases when hadK is pre-determined
-        input = hadK.view(1, K, K).to(input) @ input
+    # Do not explicitly repeat - OOM
+    # input = torch.bmm(
+    #     hadK.repeat(len(input), 1, 1).to(input.device).to(input.dtype), input)
+    # Use bcast instead
+    input = hadK.view(1, K, K).to(input) @ input
 
     # normalize
     return input.view(X.shape)
-
-
-def _is_pow2(n: int) -> bool:
-    return (n & (n - 1) == 0) and (n > 0)
-
-
-def _reshape_bits(packed_bits: numpy.ndarray, original_size: int) -> numpy.ndarray:
-    had_unpacked = numpy.unpackbits(packed_bits)
-    had_unpacked = [1 if x == 1 else -1 for x in had_unpacked]
-    had_unpacked = numpy.array(had_unpacked).reshape((original_size, original_size))
-    return had_unpacked
-
-
-# http://www.neilsloane.com/hadamard/index.html
-def _get_had12() -> torch.Tensor:
-    # fmt: off
-    had_12 = numpy.array([128,  13,  29, 232, 235,  71, 218,  
-        62, 209, 246, 139, 180, 157, 168, 237, 199, 106,  59], dtype=numpy.uint8)
-    # fmt: on
-    # TODO: just unpack during apply
-    had_12_unpacked = _reshape_bits(had_12, original_size=12)
-    return torch.tensor(had_12_unpacked)
-
-
-def _get_had20() -> torch.Tensor:
-    # fmt: off
-    had_20 = numpy.array([128, 0,  13, 133, 121, 236,  43, 203,  97,  94, 155,  10, 252, 
-        216, 87, 230, 194, 191,  54,  21, 249, 176, 171, 205, 133, 222, 108,  42, 243,  
-        97, 215, 155,  10, 188, 216, 149, 230, 200, 175, 54, 133, 121, 188,  43, 
-        205, 225,  94, 107,  10, 243], dtype=numpy.uint8)
-    # fmt: on
-    # TODO: just unpack during apply
-    had_20_unpacked = _reshape_bits(had_20, original_size=20)
-    return torch.tensor(had_20_unpacked)

compressed_tensors/utils/offload.py

@@ -31,9 +31,10 @@ import contextlib
 import warnings
 from functools import wraps
 from operator import attrgetter
-from typing import Any, Callable, Dict, Iterable, Literal, Optional, Union
+from typing import Any, Callable, Dict, Iterable, Literal, Optional, Tuple, Union
 
 import torch
+from compressed_tensors.utils import patch_attr
 
 
 try:
@@ -83,6 +84,7 @@ __all__ = [
     "register_offload_module",
     "delete_offload_module",
     "offloaded_dispatch",
+    "disable_offloading",
 ]
 
 
@@ -204,9 +206,24 @@ def register_offload_parameter(
     has_onload = any(p.device != torch.device("meta") for p in module.parameters())
     module.register_parameter(name, parameter)
 
+    # do everything AlignDevicesHook.init_hook does
+    # https://github.com/huggingface/accelerate/blob/main/src/accelerate/hooks.py#L281
     if has_offloaded_params(module):
-        weights_map = module._hf_hook.weights_map
-        offload_to_weights_map(weights_map, name, parameter.data, offload_device)
+        hook: AlignDevicesHook = module._hf_hook
+        assert hook.weights_map is not None
+
+        # append to original_devices
+        hook.original_devices[name] = parameter.device
+
+        # append to weights map
+        offload_to_weights_map(hook.weights_map, name, parameter.data, offload_device)
+
+        # append to tied_params_map
+        offloaded = hook.weights_map[name]
+        if hook.tied_params_map is not None:
+            hook.tied_params_map[offloaded.data_ptr()] = {}  # (1)
+
+        # perform offloading
         if not has_onload:
             set_module_tensor_to_device(module, name, "meta")
 
@@ -214,7 +231,7 @@ def register_offload_parameter(
 def update_offload_parameter(
     module: torch.nn.Module,
     name: str,
-    data: Optional[torch.Tensor],
+    data: torch.Tensor,
     offload_device: Optional[Union[torch.device, Literal["disk"]]] = None,
 ):
     """
@@ -227,7 +244,7 @@ def update_offload_parameter(
     :param offload_device: device on which weight will be offloaded to. If None is
         provided, then infer device from parameters on module
     """
-    param = getattr(module, name)
+    param: torch.nn.Parameter = getattr(module, name)
     if param.data.shape != data.shape:
         warnings.warn(
             f"Shape of parameter being updated {param.data.shape} does not match shape "
@@ -235,7 +252,7 @@ def update_offload_parameter(
         )
 
     # copy data into onloaded parameter if applicable
-    if param.device != torch.device("meta"):
+    if param.device != torch.device("meta") and data is not param.data:
         param.data.copy_(data)
 
     # update offload dict
@@ -420,7 +437,6 @@ def register_offload_module(base: torch.nn.Module, name: str, module: torch.nn.M
         hook: AlignDevicesHook = base._hf_hook
         assert hook.offload
         assert hook.weights_map is not None
-        assert hook.tied_params_map is not None
 
         # offloading kwargs for submodule
         place_submodules = False
@@ -435,7 +451,8 @@ def register_offload_module(base: torch.nn.Module, name: str, module: torch.nn.M
             module, include_buffers=offload_buffers, recurse=place_submodules
         ):
             offloaded = param.to(offload_device)
-            hook.tied_params_map[offloaded.data_ptr()] = {}  # (1)
+            if hook.tied_params_map is not None:
+                hook.tied_params_map[offloaded.data_ptr()] = {}  # (1)
             offload_to_weights_map(hook.weights_map, f"{name}.{param_name}", offloaded)
 
             # if the parent places submodules, offload here
@@ -463,9 +480,6 @@ def register_offload_module(base: torch.nn.Module, name: str, module: torch.nn.M
 
     base.register_module(name, module)
 
-    # (1): Since we cannot know which pointers are shared when we add parameters in an
-    # online way, assume that all pointers are shared. This comes at no runtime cost
-
 
 def delete_offload_module(base: torch.nn.Module, name: str):
     """
@@ -501,7 +515,9 @@ def offloaded_dispatch(
         raise NotImplementedError("Disk offloading is not currently supported")
 
     # create weights map
-    weights_map = OffloadedWeightsLoader(state_dict=module.state_dict(), device="cpu")
+    state_dict = module.state_dict()
+    state_dict = {key: val.to(offload_device) for key, val in state_dict.items()}
+    weights_map = OffloadedWeightsLoader(state_dict=state_dict, device=offload_device)
 
     # create tied params map
     tied_params = find_tied_parameters(module)
@@ -522,6 +538,36 @@ def offloaded_dispatch(
     return module
 
 
+@contextlib.contextmanager
+def disable_offloading():
+    """
+    Keep modules onloaded and disable offloading until this context exits.
+    Affects modules which have been hooked with accelerate's `AlignDevicesHook`
+    """
+    original_pre_forward = AlignDevicesHook.pre_forward
+    onloaded_modules: Dict[torch.nn.Module, Tuple[AlignDevicesHook, bool]] = dict()
+
+    # onload once and disable any future onloading/offloading steps
+    def keep_onload_pre_forward(self: AlignDevicesHook, module, *args, **kwargs):
+        ret = original_pre_forward(self, module, *args, **kwargs)
+        if module not in onloaded_modules:
+            onloaded_modules[module] = (self, self.offload)
+            self.offload = False
+        return ret
+
+    # use the patched pre_forward function within the context
+    with patch_attr(AlignDevicesHook, "pre_forward", keep_onload_pre_forward):
+        yield
+
+    # manually offload all modules that were onloaded
+    # update any parameters which may have changed
+    for module, (hook, offload) in onloaded_modules.items():
+        hook.offload = offload
+        for name, param in module.named_parameters():
+            update_offload_parameter(module, name, param.data)
+        hook.post_forward(module, None)
+
+
 """ Upstreamed Functions """
 
 
@@ -589,3 +635,7 @@ def align_module_device(
 
     else:
         yield
+
+
+# (1): Since we cannot know which pointers are shared when we add parameters in an
+# online way, assume that all pointers are shared. This has virtually no runtime cost

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.2.a20250612'
+__version__ = version = '0.10.2.a20250616'
 __version_tuple__ = version_tuple = (0, 10, 2)

{compressed_tensors-0.10.2a20250612.dist-info → compressed_tensors-0.10.2a20250616.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.2a20250612
+Version: 0.10.2a20250616
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.2a20250612.dist-info → compressed_tensors-0.10.2a20250616.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=F2izwCTRKbiv1mAW6qD3TbJD5cXQrz4zRmew4qZ4Ud0,523
+compressed_tensors/version.py,sha256=luzxcDdA8EA_-n05GudjzinnLHUkR5eQCjyTTNLxDnw,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
@@ -45,21 +45,22 @@ compressed_tensors/transform/transform_config.py,sha256=6JA8VFcoz4EGHOev6thj51Ou
 compressed_tensors/transform/transform_scheme.py,sha256=c7NAuLDL0itFgUfBMNShegMI9bzKL7s4LR3QJTHsXLs,1733
 compressed_tensors/transform/factory/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/transform/factory/base.py,sha256=yVrYWEnrr2RFWE5AjSNeXzO9aXc443dTNMVSxuLztz8,5940
-compressed_tensors/transform/factory/hadamard.py,sha256=tuFVpKsv__SeDj-QwKxtipLvjb993DOSIFvWcUh42Ww,3124
+compressed_tensors/transform/factory/hadamard.py,sha256=zkq6w8uJXRLokUXajAkFb2fJrH0K3SL6qrR2dARrAr8,3139
 compressed_tensors/transform/factory/matrix_multiply.py,sha256=0g4sYC_tOmCjOomae2gl54UTXiFdl0mCCkmbqIRX8yw,3613
-compressed_tensors/transform/factory/random_hadamard.py,sha256=6kqr9z6kFc-2qRNskhWRsLGTDT_NfNAkFcTLMqQJcWA,1484
+compressed_tensors/transform/factory/random_hadamard.py,sha256=TFInxbHslqREOFFiy_mpR88eEYXQnslxXmyh-ZbN-MU,1499
 compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
-compressed_tensors/transform/utils/hadamard.py,sha256=SmPZmnHtc5N36gJA5EbM1T65uf4w1_flgl7SWBeg_W8,5642
+compressed_tensors/transform/utils/hadamard.py,sha256=U27Kvo-eDebKcVt8oXTSIAaQ5DvPQj9tDv2hdXHCPPQ,5584
+compressed_tensors/transform/utils/hadamards.safetensors,sha256=mFd1GzNodGG-ifA1IoH-0nHYzfraCOvrq_dX2zFI1B4,1436901
 compressed_tensors/transform/utils/utils.py,sha256=PRPTYwPs2nnNaQMq2GEbC4QYKHFKlZwaRyPgdDhl66g,2992
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
 compressed_tensors/utils/helpers.py,sha256=cPg-ikdeA92aIGwBONg8GmPNvcGlFhozyJVwsRiXBTA,11981
-compressed_tensors/utils/offload.py,sha256=myV7iC75gA8A3BGgwR3uoeaJkIC9oigKp9CcqsHsVJc,20686
+compressed_tensors/utils/offload.py,sha256=11k8p3y_0WVbrozYsrPjJb_4Ln8YkeO_4Q8W4ibFVlc,22710
 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-0.10.2a20250612.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.2a20250612.dist-info/METADATA,sha256=541wdYU5905X69fwti-7pubCIzjsENQnbOxpJt4X2qQ,7005
-compressed_tensors-0.10.2a20250612.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.2a20250612.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.2a20250612.dist-info/RECORD,,
+compressed_tensors-0.10.2a20250616.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.2a20250616.dist-info/METADATA,sha256=V_L8-0yZM5t7FG0knLacaaIAHNTuEmTe1f7CqmDCAXg,7005
+compressed_tensors-0.10.2a20250616.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.2a20250616.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.2a20250616.dist-info/RECORD,,