|
| 1 | +# SPDX-FileCopyrightText: Copyright (c) 2023-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. |
| 2 | +# SPDX-License-Identifier: Apache-2.0 |
| 3 | +# |
| 4 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | +# you may not use this file except in compliance with the License. |
| 6 | +# You may obtain a copy of the License at |
| 7 | +# |
| 8 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | +# |
| 10 | +# Unless required by applicable law or agreed to in writing, software |
| 11 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | +# See the License for the specific language governing permissions and |
| 14 | +# limitations under the License. |
| 15 | +"""Implements INT8 quantization for efficient tensor storage and computation.""" |
| 16 | + |
| 17 | +from typing import Union |
| 18 | + |
| 19 | +import torch |
| 20 | + |
| 21 | +from ..qtensor.base_qtensor import BaseQuantizedTensor |
| 22 | +from ..utils import ( |
| 23 | + convert_quantization_axis_to_reduce_axis, |
| 24 | + reduce_amax, |
| 25 | + reduce_block_amax, |
| 26 | + reduce_block_padding, |
| 27 | +) |
| 28 | + |
| 29 | + |
| 30 | +class INT8QTensor(BaseQuantizedTensor): |
| 31 | + """Implements the INT8 quantization on tensors for more efficient storage or computation. |
| 32 | +
|
| 33 | + Attributes: |
| 34 | + quantized_data (torch.Tensor): The quantized data stored as an INT8 tensor. |
| 35 | + """ |
| 36 | + |
| 37 | + @classmethod |
| 38 | + def quantize( |
| 39 | + cls, |
| 40 | + input: torch.Tensor, |
| 41 | + scales: torch.Tensor = None, |
| 42 | + axis: Union[tuple, int, None] = None, |
| 43 | + block_sizes: dict = None, |
| 44 | + ) -> tuple: |
| 45 | + """Converting a tensor to a quantized format based on INT8 quantization. |
| 46 | +
|
| 47 | + Args: |
| 48 | + input (torch.Tensor): The input tensor to be quantized. |
| 49 | + scales (torch.Tensor): The scales for quantization. |
| 50 | + axis: The dimensions to reduce for quantization. None or int or tuple of ints. |
| 51 | + block_sizes (dict): A dictionary specifying the block size for each dimension. |
| 52 | + Note: One can only provide axis or block_sizes for INT8 quantization. |
| 53 | +
|
| 54 | + Returns: |
| 55 | + tuple: INT8QTensor, scales |
| 56 | + """ |
| 57 | + original_input = input |
| 58 | + if scales is None: |
| 59 | + if block_sizes: |
| 60 | + input = reduce_block_padding(input, block_sizes) |
| 61 | + amax = reduce_block_amax(input, block_sizes) |
| 62 | + else: |
| 63 | + reduce_axis = convert_quantization_axis_to_reduce_axis(input, axis) |
| 64 | + amax = reduce_amax(input, axis=reduce_axis) |
| 65 | + scales = amax / 127.0 |
| 66 | + |
| 67 | + # Calculate the scale shape and make sure it aligns with input and block_sizes |
| 68 | + expected_shape = list(input.shape) |
| 69 | + expanded_scales = scales.clone() |
| 70 | + if block_sizes: |
| 71 | + for dim, block_size in block_sizes.items(): |
| 72 | + dim = dim if dim >= 0 else len(input.shape) + dim # Convert negative index |
| 73 | + assert input.shape[dim] % block_size == 0, ( |
| 74 | + f"Tensor dimension {dim}, {input.shape[dim]} is not divisible by {block_size}." |
| 75 | + ) |
| 76 | + expected_shape[dim] = ( |
| 77 | + input.shape[dim] // block_size |
| 78 | + ) # Adjust expected shape for blocks |
| 79 | + |
| 80 | + # Assert the shape of `scales` matches expected reduced dimensions |
| 81 | + assert scales.shape == tuple(expected_shape), ( |
| 82 | + f"Mismatch in expected scale shape: {scales.shape} vs {tuple(expected_shape)}" |
| 83 | + ) |
| 84 | + |
| 85 | + # Expand scales for broadcasting |
| 86 | + for dim, block_size in block_sizes.items(): |
| 87 | + expanded_scales = expanded_scales.repeat_interleave(block_size, dim=dim) |
| 88 | + |
| 89 | + # Quantization |
| 90 | + quantized_data = (input / expanded_scales).round().clamp(-128, 127).to(torch.int8) |
| 91 | + |
| 92 | + return cls(original_input.shape, original_input.dtype, quantized_data), scales |
| 93 | + |
| 94 | + def dequantize(self, dtype: torch.dtype = None, **kwarg): |
| 95 | + """Dequantize INT8 packed tensor to a target dtype.""" |
| 96 | + if dtype is None: |
| 97 | + dtype = self.metadata["dtype"] |
| 98 | + assert "scale" in kwarg, "Require scale for INT8 dequantization." |
| 99 | + |
| 100 | + # Get args |
| 101 | + scales = kwarg["scale"] |
| 102 | + block_sizes = kwarg.get("block_sizes", None) |
| 103 | + |
| 104 | + shape = self._quantized_data.shape |
| 105 | + if block_sizes: |
| 106 | + # Compute expanded shape for broadcasting scales |
| 107 | + expanded_shape = list(shape) |
| 108 | + for dim, block_size in block_sizes.items(): |
| 109 | + assert shape[dim] % block_size == 0, ( |
| 110 | + f"Dimension {shape[dim]} is not divisible by {block_size}." |
| 111 | + ) |
| 112 | + expanded_shape[dim] //= block_size # Reduce the dimension size for blocks |
| 113 | + |
| 114 | + assert tuple(expanded_shape) == scales.shape, ( |
| 115 | + f"Scales shape {scales.shape} must match expected {tuple(expanded_shape)}." |
| 116 | + ) |
| 117 | + |
| 118 | + # Expand scales for broadcasting |
| 119 | + for dim, block_size in block_sizes.items(): |
| 120 | + scales = scales.repeat_interleave(block_size, dim=dim) |
| 121 | + |
| 122 | + # Handle padded tensors |
| 123 | + slices = tuple(slice(0, dim) for dim in self.metadata["shape"]) |
| 124 | + |
| 125 | + return (self._quantized_data.view(torch.int8).to(dtype) * scales.to(dtype))[slices] |
0 commit comments