compressed-tensors-nightly 0.7.1.20241031__py3-none-any.whl → 0.7.1.20241102__py3-none-any.whl

compressed_tensors/quantization/observers/helpers.py

@@ -1,149 +0,0 @@
-# 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.
-
-from collections import Counter
-from typing import Tuple
-
-import torch
-from compressed_tensors.quantization.quant_args import (
-    FP8_DTYPE,
-    QuantizationArgs,
-    QuantizationStrategy,
-    QuantizationType,
-)
-from torch import FloatTensor, IntTensor, Tensor
-
-
-__all__ = [
-    "calculate_qparams",
-    "get_observer_token_count",
-    "calculate_range",
-    "compute_dynamic_scales_and_zp",
-]
-
-
-def compute_dynamic_scales_and_zp(value: Tensor, args: QuantizationArgs):
-    """
-    Returns the computed scales and zero points for dynamic activation
-    qunatization.
-
-    :param value: tensor to calculate quantization parameters for
-    :param args: quantization args
-    :param reduce_dims: optional tuple of dimensions to reduce along,
-        returned scale and zero point will be shaped (1,) along the
-        reduced dimensions
-    :return: tuple of scale and zero point derived from the observed tensor
-    """
-    if args.strategy == QuantizationStrategy.TOKEN:
-        dim = {1, 2}
-        reduce_dims = tuple(idx for idx in range(value.ndim) if idx not in dim)
-    elif args.strategy == QuantizationStrategy.TENSOR:
-        reduce_dims = None
-    else:
-        raise ValueError(
-            f"One of {QuantizationStrategy.TOKEN} or {QuantizationStrategy.TENSOR} ",
-            "must be used for dynamic quantization",
-        )
-
-    if not reduce_dims:
-        min_val, max_val = torch.aminmax(value)
-    else:
-        min_val = torch.amin(value, dim=reduce_dims, keepdims=True)
-        max_val = torch.amax(value, dim=reduce_dims, keepdims=True)
-
-    return calculate_qparams(min_val, max_val, args)
-
-
-def get_observer_token_count(module: torch.nn.Module) -> Counter:
-    """
-    Parse the module and return the number of tokens observed by
-    each module's observer.
-
-    :param module: module to parse
-    :return: counter with the number of tokens observed by each observer
-    """
-    token_counts = Counter()
-    for name, module in module.named_modules():
-        if name.endswith(".input_observer"):
-            token_counts[
-                name.replace(".input_observer", "")
-            ] = module._num_observed_tokens
-    return token_counts
-
-
-def calculate_qparams(
-    min_vals: Tensor, max_vals: Tensor, quantization_args: QuantizationArgs
-) -> Tuple[FloatTensor, IntTensor]:
-    """
-    :param min_vals: tensor of min value(s) to calculate scale(s) and zero point(s)
-        from
-    :param max_vals: tensor of max value(s) to calculate scale(s) and zero point(s)
-        from
-    :param quantization_args: settings to quantization
-    :return: tuple of the calculated scale(s) and zero point(s)
-    """
-    min_vals = torch.min(min_vals, torch.zeros_like(min_vals))
-    max_vals = torch.max(max_vals, torch.zeros_like(max_vals))
-    device = min_vals.device
-
-    bit_min, bit_max = calculate_range(quantization_args, device)
-    bit_range = bit_max - bit_min
-    zp_dtype = quantization_args.pytorch_dtype()
-
-    if quantization_args.symmetric:
-        max_val_pos = torch.max(torch.abs(min_vals), torch.abs(max_vals))
-        scales = max_val_pos / (float(bit_range) / 2)
-        scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
-        zero_points = torch.zeros(scales.shape, device=device, dtype=min_vals.dtype)
-    else:
-        scales = (max_vals - min_vals) / float(bit_range)
-        scales = torch.clamp(scales, min=torch.finfo(torch.float32).eps)
-        zero_points = bit_min - (min_vals / scales)
-        zero_points = torch.clamp(zero_points, bit_min, bit_max)
-
-    # match zero-points to quantized type
-    zero_points = zero_points.to(zp_dtype)
-
-    if scales.ndim == 0:
-        scales = scales.reshape(1)
-        zero_points = zero_points.reshape(1)
-
-    return scales, zero_points
-
-
-def calculate_range(quantization_args: QuantizationArgs, device: str) -> Tuple:
-    """
-    Calculated the effective quantization range for the given Quantization Args
-
-    :param quantization_args: quantization args to get range of
-    :param device: device to store the range to
-    :return: tuple endpoints for the given quantization range
-    """
-    if quantization_args.type == QuantizationType.INT:
-        bit_range = 2**quantization_args.num_bits
-        q_max = torch.tensor(bit_range / 2 - 1, device=device)
-        q_min = torch.tensor(-bit_range / 2, device=device)
-    elif quantization_args.type == QuantizationType.FLOAT:
-        if quantization_args.num_bits != 8:
-            raise ValueError(
-                "Floating point quantization is only supported for 8 bits,"
-                f"got {quantization_args.num_bits}"
-            )
-        fp_range_info = torch.finfo(FP8_DTYPE)
-        q_max = torch.tensor(fp_range_info.max, device=device)
-        q_min = torch.tensor(fp_range_info.min, device=device)
-    else:
-        raise ValueError(f"Invalid quantization type {quantization_args.type}")
-
-    return q_min, q_max

compressed_tensors/quantization/observers/min_max.py

@@ -1,104 +0,0 @@
-# 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.
-
-from typing import Any, Optional, Tuple
-
-import torch
-from compressed_tensors.quantization.observers.base import Observer
-from compressed_tensors.quantization.observers.helpers import calculate_qparams
-from compressed_tensors.quantization.quant_args import QuantizationArgs
-from torch import FloatTensor, IntTensor, Tensor
-
-
-__all__ = ["MovingAverageMinMaxObserver"]
-
-
-@Observer.register("minmax")
-class MovingAverageMinMaxObserver(Observer):
-    """
-    Implements a dynamic quantization observer that sets the scale and
-    zero point based on a moving average of the overall min and max observed values
-    """
-
-    def __init__(
-        self, quantization_args: QuantizationArgs, averaging_constant: float = 0.01
-    ):
-        super().__init__(quantization_args=quantization_args)
-
-        self.min_val = {}
-        self.max_val = {}
-        self.averaging_constant = averaging_constant
-
-    def calculate_qparams(
-        self,
-        observed: Tensor,
-        reduce_dims: Optional[Tuple[int]] = None,
-        tensor_id: Optional[Any] = None,
-    ) -> Tuple[FloatTensor, IntTensor]:
-        """
-        Updates the observed min and max using a moving average smoothed by the
-        averaging_constant
-
-        :param observed: observed tensor to calculate quantization parameters for
-        :param reduce_dims: optional tuple of dimensions to reduce along,
-            returned scale and zero point will be shaped (1,) along the
-            reduced dimensions
-        :param tensor_id: Optional id if different ranges of observed tensors are
-            passed, useful for sharding tensors by group_size
-        :return: tuple of scale and zero point derived from the observed tensor
-        """
-        tensor_id = tensor_id or "default"
-
-        if not reduce_dims:
-            min_val, max_val = torch.aminmax(observed)
-        else:
-            min_val = torch.amin(observed, dim=reduce_dims, keepdims=True)
-            max_val = torch.amax(observed, dim=reduce_dims, keepdims=True)
-
-        running_min_val = self.min_val.get(tensor_id, None)
-        running_max_val = self.max_val.get(tensor_id, None)
-
-        if running_min_val is None or running_max_val is None:
-            updated_min_val = min_val
-            updated_max_val = max_val
-        else:
-            updated_min_val = running_min_val + self.averaging_constant * (
-                min_val - running_min_val
-            )
-            updated_max_val = running_max_val + self.averaging_constant * (
-                max_val - running_max_val
-            )
-
-        self.min_val[tensor_id] = updated_min_val
-        self.max_val[tensor_id] = updated_max_val
-
-        return calculate_qparams(
-            updated_min_val, updated_max_val, self.quantization_args
-        )
-
-    def get_qparams_along_dim(
-        self, observed, dim: int, tensor_id: Optional[Any] = None
-    ):
-        reduce_dims = tuple(idx for idx in range(observed.ndim) if idx != dim)
-        return self.calculate_qparams(
-            observed, reduce_dims=reduce_dims, tensor_id=tensor_id
-        )
-
-    def reset(self):
-        """
-        Reset the state of the observer, including min and maximum values
-        """
-        super().reset()
-        self.min_val = {}
-        self.max_val = {}

compressed_tensors/quantization/observers/mse.py

@@ -1,164 +0,0 @@
-# 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.
-
-from typing import Any, Optional, Tuple
-
-import torch
-from compressed_tensors.quantization.observers.base import Observer
-from compressed_tensors.quantization.observers.helpers import calculate_qparams
-from compressed_tensors.quantization.quant_args import QuantizationArgs
-from torch import FloatTensor, IntTensor, Tensor
-
-
-__all__ = ["MovingAverageMSEObserver"]
-
-
-@Observer.register("mse")
-class MovingAverageMSEObserver(Observer):
-    """
-    Implements a dynamic quantization observer that sets the scale and
-    zero point based on a moving average of the mse-clipped min and max observed values
-    """
-
-    def __init__(
-        self,
-        quantization_args: QuantizationArgs,
-        averaging_constant: float = 0.01,
-        grid: float = 100.0,
-        maxshrink: float = 0.80,
-        norm: float = 2.4,
-    ):
-        super().__init__(quantization_args=quantization_args)
-
-        self.min_val = {}
-        self.max_val = {}
-        self.averaging_constant = averaging_constant
-        self.grid = grid
-        self.maxshrink = maxshrink
-        self.norm = norm
-
-    def calculate_mse_min_max(
-        self,
-        observed: Tensor,
-        reduce_dims: Optional[Tuple[int]] = None,
-    ):
-        """
-        Computes the mse-clipped min and max values of the observed tensor by
-        optimizing for quantization error
-
-        :param observed: observed tensor to calculate quantization parameters for
-        :param reduce_dims: optional tuple of dimensions to reduce along,
-            returned values will be shaped (1,) along the reduced dimensions
-        :return: tuple of min and max values derived from the observed tensor
-        """
-        from compressed_tensors.quantization.lifecycle import fake_quantize
-
-        if not reduce_dims:
-            absolute_min_val, absolute_max_val = torch.aminmax(observed)
-        else:
-            absolute_min_val = torch.amin(observed, dim=reduce_dims, keepdims=True)
-            absolute_max_val = torch.amax(observed, dim=reduce_dims, keepdims=True)
-
-        best = torch.full_like(
-            absolute_min_val, torch.finfo(absolute_min_val.dtype).max
-        )
-        min_val = torch.ones_like(absolute_min_val)
-        max_val = torch.zeros_like(absolute_max_val)
-        for i in range(int(self.maxshrink * self.grid)):
-            p = 1 - i / self.grid
-            shrinked_min_val = p * absolute_min_val
-            shrinked_max_val = p * absolute_max_val
-
-            candidate_scales, candidate_zero_points = calculate_qparams(
-                shrinked_min_val, shrinked_max_val, self.quantization_args
-            )
-            q = fake_quantize(
-                observed,
-                candidate_scales,
-                candidate_zero_points,
-                self.quantization_args,
-            )
-
-            q -= observed
-            q.abs_()
-            q.pow_(self.norm)
-            if not reduce_dims:
-                err = torch.sum(q)
-            else:
-                err = torch.sum(q, reduce_dims, keepdims=True)
-
-            tmp = err < best
-            if torch.any(tmp):
-                best[tmp] = err[tmp]
-                min_val[tmp] = shrinked_min_val[tmp]
-                max_val[tmp] = shrinked_max_val[tmp]
-        return min_val, max_val
-
-    def calculate_qparams(
-        self,
-        observed: Tensor,
-        reduce_dims: Optional[Tuple[int]] = None,
-        tensor_id: Optional[Any] = None,
-    ) -> Tuple[FloatTensor, IntTensor]:
-        """
-        Updates the mse-clipped min and max values of the observed tensor using
-        a moving average smoothed by the averaging_constant
-
-        :param observed: observed tensor to calculate quantization parameters for
-        :param reduce_dims: optional tuple of dimensions to reduce along,
-            returned scale and zero point will be shaped (1,) along the
-            reduced dimensions
-        :param tensor_id: Optional id if different ranges of observed tensors are
-            passed, useful for sharding tensors by group_size
-        :return: tuple of scale and zero point derived from the observed tensor
-        """
-        min_val, max_val = self.calculate_mse_min_max(observed, reduce_dims)
-
-        running_min_val = self.min_val.get(tensor_id, None)
-        running_max_val = self.max_val.get(tensor_id, None)
-
-        if running_min_val is None or running_max_val is None:
-            updated_min_val = min_val
-            updated_max_val = max_val
-        else:
-            updated_min_val = running_min_val + self.averaging_constant * (
-                min_val - running_min_val
-            )
-            updated_max_val = running_max_val + self.averaging_constant * (
-                max_val - running_max_val
-            )
-
-        tensor_id = tensor_id or "default"
-        self.min_val[tensor_id] = updated_min_val
-        self.max_val[tensor_id] = updated_max_val
-
-        return calculate_qparams(
-            updated_min_val, updated_max_val, self.quantization_args
-        )
-
-    def get_qparams_along_dim(
-        self, observed, dim: int, tensor_id: Optional[Any] = None
-    ):
-        reduce_dims = tuple(idx for idx in range(observed.ndim) if idx != dim)
-        return self.calculate_qparams(
-            observed, reduce_dims=reduce_dims, tensor_id=tensor_id
-        )
-
-    def reset(self):
-        """
-        Reset the state of the observer, including min and maximum values
-        """
-        super().reset()
-        self.min_val = {}
-        self.max_val = {}