awq.py

from typing import Any, Dict, List, Optional

import torch
from torch.nn.parameter import Parameter

from vllm import _custom_ops as ops
from vllm.model_executor.layers.linear import LinearBase, LinearMethodBase
from vllm.model_executor.layers.quantization.base_config import (
    QuantizationConfig)
from vllm.model_executor.utils import set_weight_attrs

class AWQShareWorkSpace:
    _instance = None
    
    def __new__(cls, *args, **kwargs):
        if cls._instance is None:
            cls._instance = super(AWQShareWorkSpace, cls).__new__(cls, *args, **kwargs)
            # 执行初始化
            cls._instance._initialize()
        return cls._instance

    def _initialize(self):
        self.awqworkshapcesize = 2 << 29
        self.awqworkshapce = torch.zeros(self.awqworkshapcesize // 2 + 1, dtype=torch.float16).cuda()
        #print("AWQShareWorkSpace _initialize\n")
        #print("self.awqworkshapce.device:",self.awqworkshapce.device)



class AWQConfig(QuantizationConfig):
    """Config class for AWQ.

    Reference: https://arxiv.org/abs/2306.00978
    """

    def __init__(
        self,
        weight_bits: int,
        group_size: int,
        zero_point: bool,
    ) -> None:
        self.weight_bits = weight_bits
        self.group_size = group_size
        self.zero_point = zero_point

        if self.weight_bits != 4:
            raise ValueError(
                "Currently, only 4-bit weight quantization is supported for "
                f"AWQ, but got {self.weight_bits} bits.")
        self.pack_factor = 32 // self.weight_bits

    def __repr__(self) -> str:
        return (f"AWQConfig(weight_bits={self.weight_bits}, "
                f"group_size={self.group_size}, "
                f"zero_point={self.zero_point})")

    def get_name(self) -> str:
        return "awq"

    def get_supported_act_dtypes(self) -> List[torch.dtype]:
        return [torch.half]

    def get_min_capability(self) -> int:
        # The AWQ kernel only supports Turing or newer GPUs.
        return 75

    @staticmethod
    def get_config_filenames() -> List[str]:
        return [
            "quant_config.json",  # E.g., casperhansen/vicuna-7b-v1.5-awq
            # E.g., abhinavkulkarni/mosaicml-mpt-7b-instruct-w4-g128-awq
            "quantize_config.json",
        ]

    @classmethod
    def from_config(cls, config: Dict[str, Any]) -> "AWQConfig":
        weight_bits = cls.get_from_keys(config, ["w_bit", "bits"])
        group_size = cls.get_from_keys(config, ["q_group_size", "group_size"])
        zero_point = cls.get_from_keys(config, ["zero_point"])
        return cls(weight_bits, group_size, zero_point)

    def get_quant_method(
            self, layer: torch.nn.Module) -> Optional["AWQLinearMethod"]:
        if isinstance(layer, LinearBase):
            return AWQLinearMethod(self)
        return None

    def get_scaled_act_names(self) -> List[str]:
        return ["gelu", "gelu_fast", "gelu_new", "gelu_pytorch_tanh"]


class AWQLinearMethod(LinearMethodBase):
    """Linear method for AWQ.

    Args:
        quant_config: The AWQ quantization config.
    """

    def __init__(self, quant_config: AWQConfig):
        self.quant_config = quant_config
        self.awqsingleton= AWQShareWorkSpace()

    def create_weights(self, layer: torch.nn.Module,
                       input_size_per_partition: int,
                       output_partition_sizes: List[int], input_size: int,
                       output_size: int, params_dtype: torch.dtype,
                       **extra_weight_attrs):
        if input_size_per_partition % self.quant_config.group_size != 0:
            raise ValueError(
                "The input size is not aligned with the quantized "
                "weight shape. This can be caused by too large "
                "tensor parallel size.")

        output_size_per_partition = sum(output_partition_sizes)
        if output_size_per_partition % self.quant_config.pack_factor != 0:
            raise ValueError(
                "The output size is not aligned with the quantized "
                "weight shape. This can be caused by too large "
                "tensor parallel size.")

        qweight = Parameter(
            torch.empty(
                input_size_per_partition,
                output_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        set_weight_attrs(
            qweight, {
                "input_dim": 0,
                "output_dim": 1,
                "packed_dim": 1,
                "pack_factor": self.quant_config.pack_factor,
            })
        qzeros = Parameter(
            torch.empty(
                input_size_per_partition // self.quant_config.group_size,
                output_size_per_partition // self.quant_config.pack_factor,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        set_weight_attrs(
            qzeros, {
                "input_dim": 0,
                "output_dim": 1,
                "packed_dim": 1,
                "pack_factor": self.quant_config.pack_factor,
            })
        scales = Parameter(
            torch.empty(
                input_size_per_partition // self.quant_config.group_size,
                output_size_per_partition,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        set_weight_attrs(scales, {
            "input_dim": 0,
            "output_dim": 1,
        })
        
        zeros_and_scales=Parameter(
            torch.empty(
                (input_size_per_partition // self.quant_config.group_size),
                output_size_per_partition,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )    
        set_weight_attrs(zeros_and_scales, {
            "input_dim": 0,
            "output_dim": 1,
        })

        layer.register_parameter("qweight", qweight)
        set_weight_attrs(qweight, extra_weight_attrs)
        layer.register_parameter("qzeros", qzeros)
        set_weight_attrs(qzeros, extra_weight_attrs)
        layer.register_parameter("scales", scales)
        set_weight_attrs(scales, extra_weight_attrs)
        layer.register_parameter("zeros_and_scales", zeros_and_scales)
        set_weight_attrs(zeros_and_scales, extra_weight_attrs)

    def apply(self,
              layer: torch.nn.Module,
              x: torch.Tensor,
              bias: Optional[torch.Tensor] = None) -> torch.Tensor:
        qweight = layer.qweight
        zeros_and_scales = layer.zeros_and_scales
        
        out_shape = (x.shape[:-1] + (qweight.shape[0] * 1, ))
        reshaped_x = x.reshape(-1, x.shape[-1])
        
        m = reshaped_x.shape[0]
        k = reshaped_x.shape[-1]
        n = qweight.shape[0]
        
        if k % 4096==0:
            padding_group=2
        else:
            padding_group=0
        
        out = ops.awq_gemm(reshaped_x,
                            qweight,
                            zeros_and_scales,
                            m,
                            n,
                            k,
                            self.quant_config.group_size,
                            padding_group,
                            self.awqsingleton.awqworkshapce,
                            self.awqsingleton.awqworkshapcesize)
        #下面是采用rocblas的做法
        # deqweight=ops.dequant_w4_gemm_colmajor(    #shape[n,k/8]--->[n,k]
        #                   qweight, 
        #                   zeros_and_scales,
        #                   k,
        #                   n,
        #                   self.quant_config.group_size)
        # output=F.linear(reshaped_x, deqweight)    
        
        if bias is not None:
            out.add_(bias)
        return out.reshape(out_shape)