Files
FastDeploy/fastdeploy/model_executor/layers/quantization/block_wise.py
2025-06-16 00:04:48 +08:00

95 lines
3.2 KiB
Python

"""
# Copyright (c) 2025 PaddlePaddle Authors. 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 Optional
import paddle
import fastdeploy
import fastdeploy.model_executor.ops.gpu.deep_gemm as deep_gemm
from ..utils import per_block_cast_to_fp8
from .quant_base import QuantConfigBase, QuantMethodBase
QUANT_ALIGNMENT_OFFSET = 127
QUANT_BLOCK_SIZE = 128
class BlockWiseConfig(QuantConfigBase):
"""
block wise quantization config, only support fp8 quant and only supports loading weights in BF16 format.
After loading the weights, it will automatically compute quantization sparsity and dynamically perform
per-token quantization of activations during inference.
"""
def __init__(self, weight_block_size: list = [-1, -1]) -> None:
super().__init__()
self.weight_block_size = weight_block_size
def get_name(self) -> str:
return "block_wise"
@classmethod
def from_config(cls, config: dict) -> "BlockWiseConfig":
weight_block_size = config["weight_block_size"]
return cls(weight_block_size)
def get_quant_method(self, layer) -> Optional[QuantMethodBase]:
return BlockWiseLinearMethod(self)
class BlockWiseLinearMethod(QuantMethodBase):
"""
block wise quantization method for linear
"""
def __init__(
self,
quant_config: BlockWiseConfig,
) -> None:
super().__init__()
self.quant_config = quant_config
def create_weights(self, layer):
layer.linear_weight_scale = self.create_parameter(
shape=[
(layer.embed_dim + QUANT_ALIGNMENT_OFFSET) // QUANT_BLOCK_SIZE,
(layer.num_heads * layer.head_dim + QUANT_ALIGNMENT_OFFSET) //
QUANT_BLOCK_SIZE,
],
dtype="float32",
is_bias=False,
)
def process_loaded_weights(self, layer, weights) -> None:
weight_tensor = weights.transpose([1, 0])
quanted_weight_tensor, weight_block_scale_tensor = (
per_block_cast_to_fp8(weight_tensor))
layer.linear_weight.copy_(quanted_weight_tensor, False)
layer.linear_weight_scale.set_value(weight_block_scale_tensor)
def apply(self, layer, x):
x, x_scale_tensor = fastdeploy.model_executor.ops.gpu.per_token_quant_padding(
x, self.quant_config.weight_block_size[0])
linear_out = paddle.empty(
(x.shape[0], layer.llm_config.model_config.hidden_size),
dtype=paddle.bfloat16)
deep_gemm.gemm_fp8_fp8_bf16_nt(
(x, x_scale_tensor),
(layer.linear_weight, layer.linear_weight_scale),
linear_out,
)
return linear_out