7-AWQ算子

AWQ算子实现

目标

AWQ算子本质上就是实现一个矩阵乘 A×BA\times B

  • 输入 A: 正常的连续内存 [M, K]。
  • 输入 B: 高度int4压缩。水平方向(N)打包了 bit,垂直方向(K)按 group_size 进行了分块量化。
  • 输出 C: 正常的连续内存 [M, N]

如下图所示:

输出矩阵 C

Triton Kernel: awq_gemm_kernel

输入矩阵 B: 量化权重

输入矩阵 A: 激活值

逻辑视图

存储结构

形状: [M, K]
类型: float16/bfloat16
(连续内存排布)

qweight
[K, N/8]
int32 (打包 8个 4-bit)

qzeros
[K/G, N/8]
int32 (打包 8个 4-bit)

scales
[K/G, N]
float16

解包后 B': [K, N]
(用于计算)

加载 A 分块

加载 qweight/zeros/scales

位移解包 + 反量化:
W = (qweight - zero) * scale

Tensor Core 矩阵乘
tl.dot(A, W)

形状: [M, N]
类型: float16

mermaid
graph TD
    subgraph Input_A [输入矩阵 A: 激活值]
        A_data["形状: [M, K]<br/>类型: float16/bfloat16<br/>(连续内存排布)"]
    end

    subgraph Input_B [输入矩阵 B: 量化权重]
        direction TB
        subgraph B_Storage [存储结构]
            QW["qweight<br/>[K, N/8]<br/>int32 (打包 8个 4-bit)"]
            QZ["qzeros<br/>[K/G, N/8]<br/>int32 (打包 8个 4-bit)"]
            QS["scales<br/>[K/G, N]<br/>float16"]
        end

        subgraph B_Logic [逻辑视图]
            W_Logic["解包后 B': [K, N]<br/>(用于计算)"]
        end
    end

    subgraph Operation [Triton Kernel: awq_gemm_kernel]
        direction LR
        LoadA["加载 A 分块"]
        LoadB["加载 qweight/zeros/scales"]
        Dequant["位移解包 + 反量化:<br/>W = (qweight - zero) * scale"]
        Dot["Tensor Core 矩阵乘<br/>tl.dot(A, W)"]
    end

    subgraph Output_C [输出矩阵 C]
        C_data["形状: [M, N]<br/>类型: float16"]
    end

    A_data --> LoadA
    QW --> LoadB
    QZ --> LoadB
    QS --> LoadB
    LoadB --> Dequant
    LoadA --> Dot
    Dequant --> Dot
    Dot --> C_data

实现代码

代码来自最新的vllm算子实现,事实上sglang中也是adapt的vllm的算子

实现代码:

python
# Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/quantization/awq_triton.py

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project

import torch
import triton
import triton.language as tl

AWQ_TRITON_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]


@triton.jit
def awq_gemm_kernel(
    a_ptr,
    b_ptr,
    c_ptr,
    zeros_ptr,
    scales_ptr,
    M,
    N,
    K,
    group_size,
    BLOCK_SIZE_M: tl.constexpr,
    BLOCK_SIZE_N: tl.constexpr,
    BLOCK_SIZE_K: tl.constexpr,
    SPLIT_K: tl.constexpr,
):
    pid = tl.program_id(axis=0)
    pid_z = tl.program_id(1)

    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
    # num_pid_n = (N + BLOCK_SIZE_N - 1) // BLOCK_SIZE_N
    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)

    pid_m = pid // num_pid_n
    pid_n = pid % num_pid_n

    accumulator_dtype = c_ptr.type.element_ty

    # NOTE: This doesn't work in TRITON_INTERPRET=1 mode.  Use below instead.
    # accumulator = tl.arange(0, BLOCK_SIZE_N)
    # accumulator = tl.broadcast_to(accumulator[None, :],
    # (BLOCK_SIZE_M, BLOCK_SIZE_N))
    # accumulator = accumulator & 0x0
    # accumulator = accumulator.to(accumulator_dtype)
    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=accumulator_dtype)

    # Create reverse AWQ order as tensor: [0, 4, 1, 5, 2, 6, 3, 7]
    # that will map given indices to the correct order.
    reverse_awq_order_tensor = (
        (tl.arange(0, 2) * 4)[None, :] + tl.arange(0, 4)[:, None]
    ).reshape(8)

    # Create the necessary shifts to use to unpack.
    shifts = reverse_awq_order_tensor * 4
    shifts = tl.broadcast_to(shifts[None, :], (BLOCK_SIZE_K * (BLOCK_SIZE_N // 8), 8))
    shifts = tl.reshape(shifts, (BLOCK_SIZE_K, BLOCK_SIZE_N))

    # Offsets and masks.
    offsets_am = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    masks_am = offsets_am < M

    offsets_bn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
    masks_bn = offsets_bn < N // 8

    offsets_zn = pid_n * (BLOCK_SIZE_N // 8) + tl.arange(0, BLOCK_SIZE_N // 8)
    masks_zn = offsets_zn < N // 8

    offsets_sn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    masks_sn = offsets_sn < N

    offsets_k = pid_z * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
    offsets_a = K * offsets_am[:, None] + offsets_k[None, :]
    offsets_b = (N // 8) * offsets_k[:, None] + offsets_bn[None, :]

    a_ptrs = a_ptr + offsets_a
    b_ptrs = b_ptr + offsets_b

    # NOTE: Use this in TRITON_INTERPRET=1 mode instead of tl.cdiv
    # block_offset = BLOCK_SIZE_K * SPLIT_K
    # for k in range(0, (K + block_offset - 1) // (block_offset)):
    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K * SPLIT_K)):
        masks_k = offsets_k < K
        masks_a = masks_am[:, None] & masks_k[None, :]
        a = tl.load(a_ptrs, mask=masks_a, other=0.0)

        masks_b = masks_k[:, None] & masks_bn[None, :]
        b = tl.load(b_ptrs, mask=masks_b, other=0.0)
        b = tl.interleave(b, b)
        b = tl.interleave(b, b)
        b = tl.interleave(b, b)

        # Dequantize b.
        offsets_szk = (
            BLOCK_SIZE_K * SPLIT_K * k + pid_z * BLOCK_SIZE_K
        ) // group_size + tl.arange(0, 1)
        offsets_z = (N // 8) * offsets_szk[:, None] + offsets_zn[None, :]
        masks_zk = offsets_szk < K // group_size
        masks_z = masks_zk[:, None] & masks_zn[None, :]
        zeros_ptrs = zeros_ptr + offsets_z
        zeros = tl.load(zeros_ptrs, mask=masks_z, other=0.0)
        zeros = tl.interleave(zeros, zeros)
        zeros = tl.interleave(zeros, zeros)
        zeros = tl.interleave(zeros, zeros)
        zeros = tl.broadcast_to(zeros, (BLOCK_SIZE_K, BLOCK_SIZE_N))

        offsets_s = N * offsets_szk[:, None] + offsets_sn[None, :]
        masks_sk = offsets_szk < K // group_size
        masks_s = masks_sk[:, None] & masks_sn[None, :]
        scales_ptrs = scales_ptr + offsets_s
        scales = tl.load(scales_ptrs, mask=masks_s, other=0.0)
        scales = tl.broadcast_to(scales, (BLOCK_SIZE_K, BLOCK_SIZE_N))

        b = (b >> shifts) & 0xF
        zeros = (zeros >> shifts) & 0xF
        b = (b - zeros) * scales
        b = b.to(c_ptr.type.element_ty)

        # Accumulate results.
        accumulator = tl.dot(a, b, accumulator, out_dtype=accumulator_dtype)

        offsets_k += BLOCK_SIZE_K * SPLIT_K
        a_ptrs += BLOCK_SIZE_K * SPLIT_K
        b_ptrs += BLOCK_SIZE_K * SPLIT_K * (N // 8)

    c = accumulator.to(c_ptr.type.element_ty)
    offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
    c_ptrs = c_ptr + pid_z * N * M + N * offs_cm[:, None] + offs_cn[None, :]
    c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
    tl.store(c_ptrs, c, mask=c_mask)


# input   - [M, K]
# qweight - [K, N // 8]
# qzeros  - [K // G, N // 8]
# scales  - [K // G, N]
# split_k_iters - parallelism along K-dimension, int, power of 2.
def awq_gemm_triton(
    input: torch.Tensor,
    qweight: torch.Tensor,
    scales: torch.Tensor,
    qzeros: torch.Tensor,
    split_k_iters: int,
    block_size_m: int = 32,
    block_size_n: int = 32,
    block_size_k: int = 32,
) -> torch.Tensor:
    M, K = input.shape
    N = qweight.shape[1] * 8
    group_size = qweight.shape[0] // qzeros.shape[0]

    assert N > 0 and K > 0 and M > 0
    assert qweight.shape[0] == K and qweight.shape[1] == N // 8
    assert qzeros.shape[0] == K // group_size and qzeros.shape[1] == N // 8
    assert scales.shape[0] == K // group_size and scales.shape[1] == N
    assert split_k_iters & (split_k_iters - 1) == 0 and split_k_iters != 0
    assert split_k_iters <= 32
    assert group_size <= K
    assert group_size in AWQ_TRITON_SUPPORTED_GROUP_SIZES or group_size == K

    grid = lambda META: (
        triton.cdiv(M, META["BLOCK_SIZE_M"]) * triton.cdiv(N, META["BLOCK_SIZE_N"]),
        split_k_iters,
    )

    result = torch.zeros((split_k_iters, M, N), dtype=scales.dtype, device=input.device)

    # A = input, B = qweight, C = result
    # A = M x K, B = K x N, C = M x N
    awq_gemm_kernel[grid](
        input,
        qweight,
        result,
        qzeros,
        scales,
        M,
        N,
        K,
        group_size,
        BLOCK_SIZE_M=block_size_m,
        BLOCK_SIZE_N=block_size_n,
        BLOCK_SIZE_K=block_size_k,
        SPLIT_K=split_k_iters,
    )

    result = result.sum(0)

    return result

代码分析:

3. 结果合并: Split-K Reduction

2. Triton Kernel 内部: 每一个程序实例 pid

1. Python/Host 端: 任务分发与 Grid 配置

计算阶段

核心反量化: 寄存器级操作

加载阶段

输入 A [M, K] (fp16)
qweight [K, N/8] (int32)

计算 Grid 尺寸:
(M/BLOCK_M * N/BLOCK_N, SPLIT_K)

分配临时 Result [SPLIT_K, M, N] (fp16)

加载 A 分块 [BLOCK_M, BLOCK_K]
(fp16)

加载 qweight 分块 [BLOCK_K, BLOCK_N/8]
(int32)

tl.interleave (8x)
将 int32 广播扩展

>> shifts & 0xF
(按 [0,4,1,5,2,6,3,7] 顺序)

加载 scales [1, BLOCK_N] (fp16)
加载 qzeros [1, BLOCK_N/8] (int32)

W = (W_int4 - zero) * scale
(类型转换为 fp16)

tl.dot(A, W)
利用 Tensor Core 执行

Accumulator 累加 (K-loop)

result.sum(axis=0)
将不同 K 段的结果相加

输出 C [M, N] (fp16)

mermaid
graph TB
    subgraph Host_Side [1. Python/Host 端: 任务分发与 Grid 配置]
        Init["输入 A [M, K] (fp16)<br/>qweight [K, N/8] (int32)"]
        GridConfig["计算 Grid 尺寸:<br/>(M/BLOCK_M * N/BLOCK_N, SPLIT_K)"]
        Workspace["分配临时 Result [SPLIT_K, M, N] (fp16)"]
    end

    subgraph Triton_Kernel [2. Triton Kernel 内部: 每一个程序实例 pid]
        direction TB

        subgraph Load_Phases [加载阶段]
            LoadA["加载 A 分块 [BLOCK_M, BLOCK_K]<br/>(fp16)"]
            LoadB["加载 qweight 分块 [BLOCK_K, BLOCK_N/8]<br/>(int32)"]
        end

        subgraph Dequant_Logic [核心反量化: 寄存器级操作]
            direction LR
            Interleave["tl.interleave (8x)<br/>将 int32 广播扩展"]
            Shift[">> shifts & 0xF<br/>(按 [0,4,1,5,2,6,3,7] 顺序)"]
            LoadSZ["加载 scales [1, BLOCK_N] (fp16)<br/>加载 qzeros [1, BLOCK_N/8] (int32)"]
            ApplySZ["W = (W_int4 - zero) * scale<br/>(类型转换为 fp16)"]
        end

        subgraph Compute_Phase [计算阶段]
            Dot["tl.dot(A, W)<br/>利用 Tensor Core 执行"]
            Accumulate["Accumulator 累加 (K-loop)"]
        end
    end

    subgraph Reduction_Phase [3. 结果合并: Split-K Reduction]
        Sum["result.sum(axis=0)<br/>将不同 K 段的结果相加"]
        Final["输出 C [M, N] (fp16)"]
    end

    Init --> GridConfig
    GridConfig --> Workspace
    Workspace --> Load_Phases
    LoadA --> Dot
    LoadB --> Dequant_Logic
    LoadSZ --> Dequant_Logic
    Dequant_Logic --> Dot
    Dot --> Accumulate
    Accumulate --> Sum
    Sum --> Final

    style Dequant_Logic fill:#f9f,stroke:#333,stroke-width:2px
    style Compute_Phase fill:#bbf,stroke:#333,stroke-width:2px

计算逻辑:

在 Triton kernel 内部,每一轮循环(K 维度的 BLOCK_SIZE_K):

  1. 加载: 从 qweight 取出一个 [BLOCK_SIZE_K, BLOCK_SIZE_N / 8] 的块。
  2. 解包 (Unpack): 通过 (iweights >> shifts) & 0xF 得到 8 倍宽度的矩阵,恢复到逻辑上的 N 。
  3. 对齐 Group: 根据当前 k 索引找到对应的 scales 和 zeros。由于 scales 形状是 [K/G, N],而 k 在变动,所以每跑完 group_size 步,才会更新一次 scale。
  4. 反量化 (Dequantize):

Wfp16=(Wint4−Zeroint4)×Scalefp16W_{fp16}=\left(W_{int4}−Zero_{int4}\right)\times Scale_{fp16}(1)

  1. 累加: 与激活值 A 进行半精度矩阵乘法。
Join the discussion Powered by GitHub Discussions