compressed-tensors 0.3.3__py3-none-any.whl → 0.5.0__py3-none-any.whl

compressed_tensors/utils/helpers.py

@@ -12,21 +12,24 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 from typing import Optional
 
-from compressed_tensors.base import SPARSITY_CONFIG_NAME
-from compressed_tensors.compressors import ModelCompressor
-from compressed_tensors.config import CompressionConfig
+import torch
 from transformers import AutoConfig
 
 
-__all__ = ["infer_compressor_from_model_config"]
+__all__ = [
+    "infer_compressor_from_model_config",
+    "fix_fsdp_module_name",
+    "tensor_follows_mask_structure",
+]
+
+FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
 
 
 def infer_compressor_from_model_config(
     pretrained_model_name_or_path: str,
-) -> Optional[ModelCompressor]:
+) -> Optional["ModelCompressor"]:  # noqa: F821
     """
     Given a path to a model config, extract a sparsity config if it exists and return
     the associated ModelCompressor
@@ -34,8 +37,11 @@ def infer_compressor_from_model_config(
     :param pretrained_model_name_or_path: path to model config on disk or HF hub
     :return: matching compressor if config contains a sparsity config
     """
+    from compressed_tensors.compressors import ModelCompressor
+    from compressed_tensors.config import CompressionConfig
+
     config = AutoConfig.from_pretrained(pretrained_model_name_or_path)
-    sparsity_config = getattr(config, SPARSITY_CONFIG_NAME, None)
+    sparsity_config = ModelCompressor.parse_sparsity_config(config)
     if sparsity_config is None:
         return None
 
@@ -43,3 +49,44 @@ def infer_compressor_from_model_config(
     sparsity_config = CompressionConfig.load_from_registry(format, **sparsity_config)
     compressor = ModelCompressor.load_from_registry(format, config=sparsity_config)
     return compressor
+
+
+# TODO: There is already the same function in
+# SparseML, should be moved to a shared location
+# in the future
+def fix_fsdp_module_name(name: str) -> str:
+    """
+    Remove FSDP wrapper prefixes from a module name
+    Accounts for scenario where FSDP_WRAPPER_NAME is
+    at the end of the name, as well as in the middle.
+    :param name: name to strip
+    :return: stripped name
+    """
+    return name.replace(FSDP_WRAPPER_NAME + ".", "").replace(
+        "." + FSDP_WRAPPER_NAME, ""
+    )
+
+
+def tensor_follows_mask_structure(tensor, mask: str = "2:4") -> bool:
+    """
+    :param tensor: tensor to check
+    :param mask: mask structure to check for, in the format "n:m"
+    :return: True if the tensor follows the mask structure, False otherwise.
+        Note, some weights can incidentally be zero, so we check for
+        atleast n zeros in each chunk of size m
+    """
+
+    n, m = tuple(map(int, mask.split(":")))
+    # Reshape the tensor into chunks of size m
+    tensor = tensor.view(-1, m)
+
+    # Count the number of zeros in each chunk
+    zero_counts = (tensor == 0).sum(dim=1)
+
+    # Check if the number of zeros in each chunk atleast n
+    # Greater than sign is needed as some weights can incidentally
+    # be zero
+    if not torch.all(zero_counts >= n).item():
+        raise ValueError()
+
+    return True

compressed_tensors/utils/offload.py

@@ -0,0 +1,104 @@
+# 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 torch
+from torch.nn import Module
+
+
+__all__ = [
+    "is_module_offloaded",
+    "get_execution_device",
+    "get_offloaded_device",
+    "update_prefix_dict",
+    "update_parameter_data",
+]
+
+
+def is_module_offloaded(module: Module) -> bool:
+    """
+    :param module: layer to check
+    :return: True if layer is offloaded from GPU, False otherwise
+    """
+    return hasattr(module, "_hf_hook") and module._hf_hook.offload
+
+
+def get_execution_device(module: Module) -> torch.device:
+    """
+    :param module: layer to check
+    :return: device layer is loaded onto during forward pass
+    """
+    if is_module_offloaded(module):
+        return module._hf_hook.execution_device
+    return next(module.parameters()).device
+
+
+def get_offloaded_device(module: Module) -> torch.device:
+    """
+    :param module: layer to check
+    :return: device layer is offloaded to onto after forward pass
+    """
+    if is_module_offloaded(module):
+        first_key = list(module._hf_hook.weights_map.keys())[0]
+        prefix_dataset = module._hf_hook.weights_map.dataset
+        return prefix_dataset[first_key].device
+    return next(module.parameters()).device
+
+
+def update_prefix_dict(module: Module, key: str, data: torch.Tensor):
+    """
+    Updates the offloaded state dict for a given module. Parameter named key is replaced
+    by data. This is neccesary because parameter updates for offloaded modules do not
+    persist automatically between loads. This function only affects the offloaded
+    state dict and not the current state of the loaded module.
+
+    :param module: layer containing the parameter to update
+    :param key: name of parameter to update
+    :param data: tensor to update parameter with in the offloaded state dict
+    """
+    if not is_module_offloaded(module):
+        raise ValueError("Prefix dict is only applicable to offloaded modules")
+    prefix_dict = module._hf_hook.weights_map
+    prefix_dict.dataset[f"{prefix_dict.prefix}{key}"] = data
+
+
+def update_parameter_data(
+    module: Module, new_param_data: torch.Tensor, param_name: str
+):
+    """
+    Updates the paramter value named param_name for a given module. This function
+    updates both the current loaded module state and the offloaded state dict if
+    the module is offloaded. This is neccesary because parameter updates for offloaded
+    modules do not persist automatically between loads.
+
+    :param module: layer containing the parameter to update
+    :param new_param_data: tensor to update parameter with
+    :param param_name:
+    """
+    device = next(module.parameters()).device
+
+    offloaded = False
+    if is_module_offloaded(module):
+        offload_device = get_offloaded_device(module)
+        offloaded = True
+
+    parameter = getattr(module, param_name, None)
+    dtype = parameter.dtype
+    parameter.data = new_param_data.to(device).to(dtype)
+
+    if offloaded:
+        prefix_dict = module._hf_hook.weights_map.dataset
+        prefix = module._hf_hook.weights_map.prefix
+        prefix_dict[f"{prefix}{param_name}"] = new_param_data.to(offload_device).to(
+            dtype
+        )

compressed_tensors/utils/permutations_24.py

@@ -0,0 +1,65 @@
+# 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 numpy
+import torch
+
+
+__all__ = ["get_permutations_24"]
+
+
+# Precompute permutations for Marlin24 weight and scale shuffling
+# Originally implemented in nm-vllm/vllm/model_executor/layers/quantization/utils/marlin_24_perms.py # noqa: E501
+#
+# Marlin works on [16*2,64] tiles. The goal of the permutations is to reorder the weight
+# data so that it is compatible with the tensor-core format that is described here:
+# https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#matrix-fragments-for-mma-m16n8k16-with-floating-point-type # noqa: E501
+#
+# As a result of this reordering, the vector loads inside the kernel will get the data
+# as it is needed for tensor-core (without the need to use ldmatrix instructions)
+def get_permutations_24(num_bits):
+    perm_list = []
+    for i in range(32):
+        perm1 = []
+        col = i // 4
+        col_o = col // 2
+        for block in [0, 1]:
+            for row in [
+                2 * (i % 4),
+                2 * (i % 4) + 1,
+                2 * (i % 4 + 4),
+                2 * (i % 4 + 4) + 1,
+            ]:
+                perm1.append(16 * row + col_o * 256 + 8 * (col % 2) + 4 * block)
+        for j in range(4):
+            perm_list.extend([p + 1 * j for p in perm1])
+    perm = numpy.array(perm_list)
+
+    if num_bits == 4:
+        interleave = numpy.array([0, 2, 4, 6, 1, 3, 5, 7])
+    elif num_bits == 8:
+        interleave = numpy.array([0, 2, 1, 3])
+    else:
+        raise ValueError("num_bits must be 4 or 8, got {}".format(num_bits))
+
+    perm = perm.reshape((-1, len(interleave)))[:, interleave].ravel()
+    perm = torch.from_numpy(perm)
+    scale_perm = []
+    for i in range(8):
+        scale_perm.extend([i * 8 + j for j in [0, 4, 1, 5, 2, 6, 3, 7]])
+    scale_perm_single = []
+    for i in range(8):
+        scale_perm_single.extend([8 * i + j for j in [0, 1, 2, 3, 4, 5, 6, 7]])
+    return perm, scale_perm, scale_perm_single

compressed_tensors/utils/safetensors_load.py

@@ -31,6 +31,7 @@ __all__ = [
     "get_weight_mappings",
     "get_nested_weight_mappings",
     "get_quantization_state_dict",
+    "is_quantization_param",
 ]
 
 
@@ -214,7 +215,7 @@ def get_quantization_state_dict(model_path: str) -> Dict[str, Tensor]:
     weight_mappings = get_weight_mappings(model_path)
     state_dict = {}
     for weight_name, safe_path in weight_mappings.items():
-        if not _is_quantization_weight(weight_name):
+        if not is_quantization_param(weight_name):
             continue
         with safe_open(safe_path, framework="pt", device="cpu") as f:
             state_dict[weight_name] = f.get_tensor(weight_name)
@@ -222,7 +223,7 @@ def get_quantization_state_dict(model_path: str) -> Dict[str, Tensor]:
     return state_dict
 
 
-def _is_quantization_weight(name: str) -> bool:
+def is_quantization_param(name: str) -> bool:
     """
     Checks is a parameter name is associated with a quantization parameter
 

compressed_tensors/utils/semi_structured_conversions.py

@@ -0,0 +1,341 @@
+#
+# Modified by Roberto Lopez Castro (roberto.lopez.castro@udc.es).
+# Pulled from nm-vllm/vllm/model_executor/layers/quantization/utils/format_24.py
+#
+# flake8: noqa
+# isort: skip_file
+
+# 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 torch
+
+
+__all__ = [
+    "sparse_semi_structured_from_dense_cutlass",
+    "sparse_semi_structured_to_dense_cutlass",
+    "mask_creator",
+]
+
+# This is PyTorch implementation of main part of reorder_meta()
+# function, from tools/util/include/cutlass/util/host_reorder.h file
+# of CUTLASS source tree.  Furthermore, CUTLASS template for sparse
+# GEMM decides upon layout of this matrix, and at the moment for the
+# sparse GEMM executed on tensor cores, this is layout described by
+# ColumnMajorInterleaved<2> data structure, in
+# include/cutlass/layout/matrix.h of CUTLASS source tree.  The
+# reordering of meta matrix into meta_reordered matrix calculated
+# according to these segments of CUTLASS code is re-implemented here.
+# Note that this calculation produces offsets for scattering metadata
+# matrix elements into reordered metadata matrix elements (or,
+# equivalently, for gathering reordered metadata matrix element back
+# into metadata matrix elements).
+def _calculate_meta_reordering_scatter_offsets(m, meta_ncols, meta_dtype, device):
+    dst_rows = torch.arange(0, m, device=device)[:, None].repeat(1, meta_ncols)
+    dst_cols = torch.arange(0, meta_ncols, device=device).repeat(m, 1)
+
+    # Reorder the rows, then swizzle the 2x2 blocks.
+    group_x = 64
+    group_y = 32 if meta_dtype.itemsize == 2 else 16
+
+    dst_rows = (
+        dst_rows // group_x * group_x
+        + (dst_rows % 2) * 2
+        + (dst_rows % 8) // 4
+        + ((dst_rows % group_y) % 4) // 2 * 32
+        + ((dst_rows % group_x) // 8) * 4
+    )
+
+    topright = ((dst_rows % 2 == 0) & (dst_cols % 2 == 1)).to(torch.int8)
+    bottomleft = ((dst_rows % 2 == 1) & (dst_cols % 2 == 0)).to(torch.int8)
+    dst_rows += topright - bottomleft
+    dst_cols -= topright - bottomleft
+
+    # Assumed that meta tensor is to be stored in CUTLASS
+    # InterleavedColumnMajor layout, and reverse engineered
+    # corresponding code to store values into this tensor.
+    interleave = 2
+    cols_maj = dst_cols // interleave
+    cols_min = dst_cols % interleave
+    return (cols_maj * m * interleave + dst_rows * interleave + cols_min).view(-1)
+
+
+# This function converts dense matrix into sparse semi-structured
+# representation, producing "compressed" matrix, in the layout used by
+# CUTLASS backend, and corresponding metadata matrix.
+def sparse_semi_structured_from_dense_cutlass(dense):
+    if dense.dim() != 2:
+        raise RuntimeError(
+            f"Expected 2-dimensional dense tensor, got {dense.dim()}-dimensional tensor"  # noqa: E501
+        )
+
+    m, k = dense.shape
+    device = dense.device
+
+    meta_dtype = torch.int8
+    if dense.dtype == torch.int8:
+        meta_dtype = torch.int32
+    elif dense.dtype in [torch.half, torch.bfloat16, torch.float, torch.int32]:
+        meta_dtype = torch.int16
+    else:
+        raise RuntimeError(f"Invalid datatype {dense.dtype} of dense matrix")
+    quadbits_per_meta_elem = meta_dtype.itemsize * 8 // 4
+    if quadbits_per_meta_elem not in (4, 8):
+        raise RuntimeError("Invalid number of elements per meta element calculated")
+
+    if meta_dtype == torch.int32:
+        if m % 16 != 0:
+            raise RuntimeError(
+                f"Number of rows of dense matrix {m} must be divisible by 16"
+            )
+    else:
+        if m % 32 != 0:
+            raise RuntimeError(
+                f"Number of rows of dense matrix {m} must be divisible by 32"
+            )
+    if k % (4 * quadbits_per_meta_elem) != 0:
+        raise RuntimeError(
+            f"Number of columns of dense matrix {k} must be divisible by {4 * quadbits_per_meta_elem}"  # noqa: E501
+        )
+
+    if dense.dtype != torch.float:
+        ksparse = 4
+        dense_4 = dense.view(-1, k // ksparse, ksparse)
+        m0, m1, m2, m3 = (dense_4 != 0).unbind(-1)
+    else:
+        ksparse = 2
+        dense_2 = dense.view(-1, k // ksparse, ksparse)
+        m0, m2 = m1, m3 = (dense_2 != 0).unbind(-1)
+    meta_ncols = k // (ksparse * quadbits_per_meta_elem)
+
+    # Encoding quadruples of True/False values as follows:
+    #     [True,  True,  False, False] -> 0b0100
+    #     [True,  False, True,  False] -> 0b1000
+    #     [False, True,  True,  False] -> 0b1001
+    #     [True,  False, False, True ] -> 0b1100
+    #     [False, True,  False, True ] -> 0b1101
+    #     [False, False, True,  True ] -> 0b1110
+    # Thus, lower two bits in the encoding are index of the True value
+    # at the lowest index in the quadruple, and the higher two bits in
+    # the encoding are index of the other True value in the quadruple.
+    # In case there are less than two True values, than False value or
+    # values at some index or indices are considered True for the
+    # encoding.  In case there are more than two True values, then the
+    # excess True value(s) at some indices are considered False for
+    # the encoding.  The exact encodings used for these cases are as
+    # follows:
+    #     [False, False, False, False] -> 0b1110
+    #     [False, False, False, True ] -> 0b1110
+    #     [False, False, True,  False] -> 0b1110
+    #     [False, True,  False, False] -> 0b1001
+    #     [False, True,  True,  True ] -> 0b1101
+    #     [True,  False, False, False] -> 0b1000
+    #     [True,  False, True,  True ] -> 0b1100
+    #     [True,  True,  False, True ] -> 0b0100
+    #     [True,  True,  True,  False] -> 0b0100
+    #     [True,  True,  True,  True ] -> 0b0100
+    # These particular encodings are chosen, with the help of Espresso
+    # logic minimizer software, for the purpose of minimization of
+    # corresponding Boolean functions, that translate non-zero flags
+    # into encoding bits.  Note also possible choices for the first
+    # and last of these encodings were limited only to (0b0100,
+    # 0b1110), in order to produce valid encodings for 1:2 sparsity
+    # case.
+
+    expr0 = m0 & m1
+    expr1 = ~m0 & m1
+    expr2 = ~m0 & ~m1
+    bit0 = expr1
+    bit1 = expr2
+    bit2 = expr0 | expr2 | m3
+    bit3 = expr1 | ~m1
+    idxs0 = bit0 | (bit1.to(torch.int64) << 1)
+    idxs1 = bit2 | (bit3.to(torch.int64) << 1)
+
+    if dense.dtype != torch.float:
+        sparse0 = dense_4.gather(
+            -1, idxs0.unsqueeze(-1)
+        )  # type: ignore[possibly-undefined]
+        sparse1 = dense_4.gather(-1, idxs1.unsqueeze(-1))
+        sparse = torch.stack((sparse0, sparse1), dim=-1).view(m, k // 2)
+    else:
+        sparse = dense_2.gather(-1, idxs0.unsqueeze(-1) // 2).view(
+            m, k // 2
+        )  # type: ignore[possibly-undefined]
+
+    meta_4 = idxs0 | (idxs1 << 2)
+    meta_n = meta_4.view((-1, meta_ncols, quadbits_per_meta_elem)).to(meta_dtype)
+
+    if quadbits_per_meta_elem == 4:
+        meta = (
+            meta_n[:, :, 0]
+            | (meta_n[:, :, 1] << 4)
+            | (meta_n[:, :, 2] << 8)
+            | (meta_n[:, :, 3] << 12)
+        )
+    elif quadbits_per_meta_elem == 8:
+        meta = (
+            meta_n[:, :, 0]
+            | (meta_n[:, :, 1] << 4)
+            | (meta_n[:, :, 2] << 8)
+            | (meta_n[:, :, 3] << 12)
+            | (meta_n[:, :, 4] << 16)
+            | (meta_n[:, :, 5] << 20)
+            | (meta_n[:, :, 6] << 24)
+            | (meta_n[:, :, 7] << 28)
+        )
+
+    # Reorder meta tensor elements.
+    meta_reordered = meta.new_empty(
+        (m * meta_ncols,)
+    )  # type: ignore[possibly-undefined]
+    meta_offsets = _calculate_meta_reordering_scatter_offsets(
+        m, meta_ncols, meta_dtype, device
+    )
+    meta_reordered.scatter_(0, meta_offsets, meta.view(-1))
+
+    return (sparse, meta_reordered.view(m, meta_ncols))
+
+
+# This function performs reverse of the function above - it
+# reconstructs dense matrix from a pair of "compressed" matrix, given
+# in the layout used by CUTLASS backend, and accompanying metadata
+# matrix.
+def sparse_semi_structured_to_dense_cutlass(sparse, meta_reordered):
+    if sparse.dim() != 2:
+        raise RuntimeError(
+            f"Expected 2-dimensional sparse tensor, got {sparse.dim()}-dimensional tensor"  # noqa: E501
+        )
+
+    m, k = sparse.shape
+    device = sparse.device
+
+    if meta_reordered.dim() != 2:
+        raise RuntimeError(
+            f"Expected 2-dimensional meta tensor, got {meta_reordered.dim()}-dimensional tensor"  # noqa: E501
+        )
+    if meta_reordered.device != device:
+        raise RuntimeError(
+            f"Expected meta matrix to be on {device} device, got matrix on {meta_reordered.device} device"  # noqa: E501
+        )
+
+    meta_dtype = meta_reordered.dtype
+    if meta_dtype not in (torch.int16, torch.int32):
+        raise RuntimeError(f"Invalid datatype {meta_dtype} of meta matrix")
+    quadbits_per_meta_elem = meta_dtype.itemsize * 8 // 4
+
+    ksparse = 4 if sparse.dtype != torch.float else 2
+
+    meta_nrows, meta_ncols = meta_reordered.shape
+    if meta_nrows != m:
+        raise RuntimeError(
+            f"Number of rows of meta matrix {meta_nrows} must be equal to number of columns of spase matrix {m}"  # noqa: E501
+        )
+    if meta_ncols * ksparse * quadbits_per_meta_elem != 2 * k:
+        raise RuntimeError(
+            f"Number of columns of sparse matrix {k} different from the {meta_ncols * ksparse * quadbits_per_meta_elem // 2}, "  # noqa: E501
+            "expected according to the number of columns of meta matrix"
+        )
+
+    # Undo meta tensor elements reordering.
+    meta_offsets = _calculate_meta_reordering_scatter_offsets(
+        m, meta_ncols, meta_dtype, device
+    )
+    meta = torch.gather(meta_reordered.view(-1), 0, meta_offsets).view(m, meta_ncols)
+
+    # Unpack sparse tensor back to original dense tensor, using
+    # information provided by meta tensor.  Note that torch.float
+    # datatype is handled pretty much the same as
+    # torch.half/torch.bfloat16, as metadata for a pair of torch.float
+    # value is encoded as if underlying 8 bytes contain four
+    # torch.half/torch.bfloat16 values, where either first two or last
+    # two are zeros.
+    meta_2 = torch.empty(
+        (m, meta_ncols, 2 * quadbits_per_meta_elem),
+        dtype=meta_dtype,
+        device=device,
+    )
+    if quadbits_per_meta_elem == 4:
+        meta_2[:, :, 0] = meta & 0b11
+        meta_2[:, :, 1] = (meta >> 2) & 0b11
+        meta_2[:, :, 2] = (meta >> 4) & 0b11
+        meta_2[:, :, 3] = (meta >> 6) & 0b11
+        meta_2[:, :, 4] = (meta >> 8) & 0b11
+        meta_2[:, :, 5] = (meta >> 10) & 0b11
+        meta_2[:, :, 6] = (meta >> 12) & 0b11
+        meta_2[:, :, 7] = (meta >> 14) & 0b11
+    elif quadbits_per_meta_elem == 8:
+        meta_2[:, :, 0] = meta & 0b11
+        meta_2[:, :, 1] = (meta >> 2) & 0b11
+        meta_2[:, :, 2] = (meta >> 4) & 0b11
+        meta_2[:, :, 3] = (meta >> 6) & 0b11
+        meta_2[:, :, 4] = (meta >> 8) & 0b11
+        meta_2[:, :, 5] = (meta >> 10) & 0b11
+        meta_2[:, :, 6] = (meta >> 12) & 0b11
+        meta_2[:, :, 7] = (meta >> 14) & 0b11
+        meta_2[:, :, 8] = (meta >> 16) & 0b11
+        meta_2[:, :, 9] = (meta >> 18) & 0b11
+        meta_2[:, :, 10] = (meta >> 20) & 0b11
+        meta_2[:, :, 11] = (meta >> 22) & 0b11
+        meta_2[:, :, 12] = (meta >> 24) & 0b11
+        meta_2[:, :, 13] = (meta >> 26) & 0b11
+        meta_2[:, :, 14] = (meta >> 28) & 0b11
+        meta_2[:, :, 15] = (meta >> 30) & 0b11
+
+    dense_offsets = meta_2.view(-1) + (
+        torch.arange(0, 2 * m * k // ksparse, device=device) * 4
+    ).view(-1, 1).repeat(1, 2).view(-1)
+
+    dense = torch.zeros((m * 2 * k,), dtype=sparse.dtype, device=device)
+    if sparse.dtype != torch.float:
+        # dense.scatter_(0, dense_offsets, sparse.view(-1))
+        dense.scatter_(0, dense_offsets, sparse.reshape(-1))
+    else:
+        dense.view(torch.half).scatter_(
+            0, dense_offsets, sparse.view(torch.half).view(-1)
+        )
+
+    return dense.view(m, 2 * k)
+
+
+def mask_creator(tensor):
+    """
+    Class for creating N:M sparsity masks.
+    Masks will be created using the N:M ratio, where for every block of
+    M weights, N will be pruned based on ranked weight value. Each mask
+    will correspond to the given tensor.
+
+    :param N: The number of weights in a group to keep
+    :param M: The size of a weight group
+    """
+    N = 2
+    M = 4
+
+    mask = None
+    # for i, tensor in enumerate(tensors):
+    if tensor.numel() % M != 0:
+        raise ValueError(
+            f"Tensor of size {tensor.shape} can't be evenly divided into " f"{M} groups"
+        )
+
+    num_groups = tensor.numel() // M
+
+    # N:M sparsity for linear layers
+    tensor_temp = tensor.detach().abs().reshape(num_groups, M)
+    index = torch.argsort(tensor_temp, dim=1)[:, : int(M - N)]
+
+    w_b = torch.ones(tensor_temp.shape, device=tensor_temp.device)
+    mask = w_b.scatter_(dim=1, index=index, value=0).reshape(tensor.shape)
+
+    return mask

compressed_tensors/version.py

@@ -0,0 +1,53 @@
+# 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.
+
+"""
+Functionality for storing and setting the version info for SparseML
+"""
+
+
+version_base = "0.5.0"
+is_release = True  # change to True to set the generated version as a release version
+
+
+def _generate_version(
+    is_release: bool,
+    version_base: str,
+):
+    from datetime import date
+
+    if is_release:
+        return version_base
+    else:
+        return f"{version_base}.{date.today().strftime('%Y%m%d')}"
+
+
+__all__ = [
+    "__version__",
+    "version_base",
+    "is_release",
+    "version",
+    "version_major",
+    "version_minor",
+    "version_patch",
+    "version_build",
+    "version_major_minor",
+]
+__version__ = _generate_version(is_release, version_base)
+
+version = __version__
+version_major, version_minor, version_patch, version_build = version.split(".") + (
+    [None] if len(version.split(".")) < 4 else []
+)  # handle conditional for version being 3 parts or 4 (4 containing build date)
+version_major_minor = f"{version_major}.{version_minor}"