qlinear_exllama.py

# Adapted from turboderp exllama: https://github.com/turboderp/exllama

import math
from logging import getLogger

import numpy as np
import torch
import torch.nn as nn
import transformers


logger = getLogger(__name__)

try:
    from exllama_kernels import make_q4, q4_matmul
except ImportError as e:
    exllama_import_exception = e

    def error_raiser_exllama(*args, **kwargs):
        raise ValueError(
            f"Trying to use the exllama backend, but could not import the C++/CUDA dependencies with the following error: {exllama_import_exception}"
        )

    make_q4 = error_raiser_exllama
    q4_matmul = error_raiser_exllama

# Dummy tensor to pass instead of g_idx since there is no way to pass "None" to a C++ extension
none_tensor = torch.empty((1, 1), device="meta")


def ext_make_q4(qweight, qzeros, scales, g_idx, device):
    """Construct Q4Matrix, return handle"""
    return make_q4(qweight, qzeros, scales, g_idx if g_idx is not None else none_tensor, device)


def ext_q4_matmul(x, q4, q4_width):
    """Matrix multiplication, returns x @ q4"""
    outshape = x.shape[:-1] + (q4_width,)
    x = x.view(-1, x.shape[-1])
    output = torch.empty((x.shape[0], q4_width), dtype=torch.float16, device=x.device)

    q4_matmul(x, q4, output)

    return output.view(outshape)


class QuantLinear(nn.Module):
    QUANT_TYPE = "exllama"

    """Linear layer implementation with per-group 4-bit quantization of the weights"""

    def __init__(self, bits, group_size, infeatures, outfeatures, bias, trainable=False, **kwargs):
        super().__init__()
        if bits != 4:
            raise ValueError(
                f"Exllama kernel supports only bits=4, requested bits={bits}. Something is wrong in the model initialization."
            )
        if trainable:
            raise NotImplementedError("Exllama kernel does not support training.")

        self.padding = -outfeatures % 32
        self.outfeatures = outfeatures + self.padding
        outfeatures = self.outfeatures

        self.infeatures = infeatures
        self.bits = bits
        self.group_size = group_size if group_size != -1 else infeatures
        self.trainable = trainable
        self.maxq = 2**self.bits - 1

        assert infeatures % 32 == 0
        assert infeatures % self.group_size == 0
        assert outfeatures % 32 == 0

        self.register_buffer(
            "qweight",
            torch.zeros((infeatures // 32 * self.bits, outfeatures), dtype=torch.int32),
        )
        self.register_buffer(
            "qzeros",
            torch.zeros(
                (
                    math.ceil(infeatures / self.group_size),
                    outfeatures // 32 * self.bits,
                ),
                dtype=torch.int32,
            ),
        )
        self.register_buffer(
            "scales",
            torch.zeros(
                (math.ceil(infeatures / self.group_size), outfeatures),
                dtype=torch.float16,
            ),
        )
        self.register_buffer(
            "g_idx",
            torch.tensor([i // self.group_size for i in range(infeatures)], dtype=torch.int32),
        )

        if bias:
            self.register_buffer("bias", torch.zeros((outfeatures), dtype=torch.float16))
        else:
            self.bias = None

    def post_init(self):
        assert self.qweight.device.type == "cuda"
        assert self.qweight.device.index is not None

        self.width = self.qweight.shape[1]

        # make_q4 segfaults if g_idx is not on cpu in the act-order case. In the non act-order case, None needs to be passed for g_idx.
        self.q4 = ext_make_q4(
            self.qweight,
            self.qzeros,
            self.scales,
            self.g_idx.to("cpu") if self._use_act_order else None,
            self.qweight.device.index,
        )

    def pack(self, linear, scales, zeros, g_idx=None):
        W = linear.weight.data.clone()
        if isinstance(linear, nn.Conv2d):
            W = W.flatten(1)
        if isinstance(linear, transformers.pytorch_utils.Conv1D):
            W = W.t()

        self.g_idx = g_idx.clone() if g_idx is not None else self.g_idx

        scales = scales.t().contiguous()
        zeros = zeros.t().contiguous()
        scale_zeros = zeros * scales
        self.scales = scales.clone().half()
        if linear.bias is not None:
            self.bias = linear.bias.clone().half()

        intweight = []
        for idx in range(self.infeatures):
            intweight.append(
                torch.round((W[:, idx] + scale_zeros[self.g_idx[idx]]) / self.scales[self.g_idx[idx]]).to(torch.int)[
                    :, None
                ]
            )
        intweight = torch.cat(intweight, dim=1)
        intweight = intweight.t().contiguous()
        intweight = intweight.numpy().astype(np.uint32)

        i = 0
        row = 0
        qweight = np.zeros((intweight.shape[0] // 32 * self.bits, intweight.shape[1]), dtype=np.uint32)
        while row < qweight.shape[0]:
            if self.bits in [4]:
                for j in range(i, i + (32 // self.bits)):
                    qweight[row] |= intweight[j] << (self.bits * (j - i))
                i += 32 // self.bits
                row += 1
            else:
                raise NotImplementedError("Only 4 bits are supported.")

        qweight = qweight.astype(np.int32)
        self.qweight = torch.from_numpy(qweight)

        zeros -= 1
        zeros = zeros.numpy().astype(np.uint32)
        qzeros = np.zeros((zeros.shape[0], zeros.shape[1] // 32 * self.bits), dtype=np.uint32)
        i = 0
        col = 0
        while col < qzeros.shape[1]:
            if self.bits in [4]:
                for j in range(i, i + (32 // self.bits)):
                    qzeros[:, col] |= zeros[:, j] << (self.bits * (j - i))
                i += 32 // self.bits
                col += 1
            else:
                raise NotImplementedError("Only 4 bits are supported.")

        qzeros = qzeros.astype(np.int32)
        self.qzeros = torch.from_numpy(qzeros)

    def forward(self, x):
        if x.dtype != torch.float16:
            logger.warning_once(
                f"The exllama kernel for GPTQ requires a float16 input activation, while {x.dtype} was passed. Casting to float16.\nMake sure you loaded your model with torch_dtype=torch.float16, that the model definition does not inadvertently cast to float32, or disable AMP Autocast that may produce float32 intermediate activations in the model."
            )

            x = x.half()

        out = ext_q4_matmul(x, self.q4, self.width)

        if self.bias is not None:
            out.add_(self.bias)
        return out