MGQA

The open source implementation of the multi grouped query attention by the paper "GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints"

Paper Link

Appreciation

• Lucidrains
• Agorians

Install

pip install mgqa

Usage

import torch
from mgqa.transformer import MGQATransformer, Decoder

x = torch.randint(0, 20000, (1, 1024))

model = MGQATransformer(e
    num_tokens = 20000,
    max_seq_len = 1024,
    attn_layers = Decoder(
        dim = 512,
        depth = 12,
        heads = 8,
        attn_kv_heads = 2 # say you want 4 query heads to attend to 1 key / value head
    )
)

result = model(x)
print(result)

Triton

• A potential triton implementation that may or may not work, I don't have gpus to test this out. If it doesn't work and you fix please let me know so we can provide this useful attn

# !pip3 install -U --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/Triton-Nightly/pypi/simple/ triton-nightly
# !pip3 install torch

import torch
import triton
import triton.language as tl


@triton.jit
def max_fn(x, y):
    return tl.math.max(x, y)


@triton.jit
def _fwd_kernel(
    Q, K, V, sm_scale,
    L,
    Out,
    stride_qz, stride_qh, stride_qm, stride_qk,
    stride_kz, stride_kh, stride_kn, stride_kk,
    stride_vz, stride_vh, stride_vk, stride_vn,
    stride_oz, stride_oh, stride_om, stride_on,
    Z, H, N_CTX,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
    IS_CAUSAL: tl.constexpr,
):
    start_m = tl.program_id(0)
    off_hz = tl.program_id(1)
    qvk_offset = off_hz * stride_qh
    Q_block_ptr = tl.make_block_ptr(
        base=Q + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_qm, stride_qk),
        offsets=(start_m * BLOCK_M, 0),
        block_shape=(BLOCK_M, BLOCK_DMODEL),
        order=(1, 0)
    )
    K_block_ptr = tl.make_block_ptr(
        base=K + qvk_offset,
        shape=(BLOCK_DMODEL, N_CTX),
        strides=(stride_kk, stride_kn),
        offsets=(0, 0),
        block_shape=(BLOCK_DMODEL, BLOCK_N),
        order=(0, 1)
    )
    V_block_ptr = tl.make_block_ptr(
        base=V + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_vk, stride_vn),
        offsets=(0, 0),
        block_shape=(BLOCK_N, BLOCK_DMODEL),
        order=(1, 0)
    )
    # initialize offsets
    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
    offs_n = tl.arange(0, BLOCK_N)
    # initialize pointer to m and l
    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
    l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
    # scale sm_scale by log_2(e) and use
    # 2^x instead of exp in the loop because CSE and LICM
    # don't work as expected with `exp` in the loop
    qk_scale = sm_scale * 1.44269504
    # load q: it will stay in SRAM throughout
    q = tl.load(Q_block_ptr)
    q = (q * qk_scale).to(tl.float16)
    # loop over k, v and update accumulator
    lo = 0
    hi = (start_m + 1) * BLOCK_M if IS_CAUSAL else N_CTX
    for start_n in range(lo, hi, BLOCK_N):
        # -- load k, v --
        k = tl.load(K_block_ptr)
        v = tl.load(V_block_ptr)
        # -- compute qk ---
        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
        if IS_CAUSAL:
            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))
        qk += tl.dot(q, k)
        # -- compute scaling constant ---
        m_i_new = tl.maximum(m_i, tl.max(qk, 1))
        alpha = tl.math.exp2(m_i - m_i_new)
        p = tl.math.exp2(qk - m_i_new[:, None])
        # -- scale and update acc --
        acc_scale = l_i * 0 + alpha  # workaround some compiler bug
        acc *= acc_scale[:, None]
        acc += tl.dot(p.to(tl.float16), v)
        # -- update m_i and l_i --
        l_i = l_i * alpha + tl.sum(p, 1)
        m_i = m_i_new
        # update pointers
        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
    # write back l and m
    acc = acc / l_i[:, None]
    l_ptrs = L + off_hz * N_CTX + offs_m
    tl.store(l_ptrs, m_i + tl.math.log2(l_i))
    # write back O
    O_block_ptr = tl.make_block_ptr(
        base=Out + qvk_offset,
        shape=(N_CTX, BLOCK_DMODEL),
        strides=(stride_om, stride_on),
        offsets=(start_m * BLOCK_M, 0),
        block_shape=(BLOCK_M, BLOCK_DMODEL),
        order=(1, 0)
    )
    tl.store(O_block_ptr, acc.to(tl.float16))


@triton.jit
def _bwd_preprocess(
    Out, DO,
    Delta,
    BLOCK_M: tl.constexpr, D_HEAD: tl.constexpr,
):
    off_m = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
    off_n = tl.arange(0, D_HEAD)
    # load
    o = tl.load(Out + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
    do = tl.load(DO + off_m[:, None] * D_HEAD + off_n[None, :]).to(tl.float32)
    # compute
    delta = tl.sum(o * do, axis=1)
    # write-back
    tl.store(Delta + off_m, delta)


@triton.jit
def _bwd_kernel(
    Q, K, V, sm_scale, Out, DO,
    DQ, DK, DV,
    L,
    D,
    stride_qz, stride_qh, stride_qm, stride_qk,
    stride_kz, stride_kh, stride_kn, stride_kk,
    stride_vz, stride_vh, stride_vk, stride_vn,
    Z, H, N_CTX,
    num_block,
    BLOCK_M: tl.constexpr, BLOCK_DMODEL: tl.constexpr,
    BLOCK_N: tl.constexpr,
    CAUSAL: tl.constexpr,
):
    off_hz = tl.program_id(0)
    off_z = off_hz // H
    off_h = off_hz % H
    qk_scale = sm_scale * 1.44269504
    # offset pointers for batch/head
    Q += off_z * stride_qz + off_h * stride_qh
    K += off_z * stride_qz + off_h * stride_qh
    V += off_z * stride_qz + off_h * stride_qh
    DO += off_z * stride_qz + off_h * stride_qh
    DQ += off_z * stride_qz + off_h * stride_qh
    DK += off_z * stride_qz + off_h * stride_qh
    DV += off_z * stride_qz + off_h * stride_qh
    for start_n in range(0, num_block):
        if CAUSAL:
            lo = start_n * BLOCK_M
        else:
            lo = 0
        # initialize row/col offsets
        offs_qm = lo + tl.arange(0, BLOCK_M)
        offs_n = start_n * BLOCK_M + tl.arange(0, BLOCK_M)
        offs_m = tl.arange(0, BLOCK_N)
        offs_k = tl.arange(0, BLOCK_DMODEL)
        # initialize pointers to value-like data
        q_ptrs = Q + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
        k_ptrs = K + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
        v_ptrs = V + (offs_n[:, None] * stride_qm + offs_k[None, :] * stride_qk)
        do_ptrs = DO + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
        dq_ptrs = DQ + (offs_qm[:, None] * stride_qm + offs_k[None, :] * stride_qk)
        # pointer to row-wise quantities in value-like data
        D_ptrs = D + off_hz * N_CTX
        l_ptrs = L + off_hz * N_CTX
        # initialize dv amd dk
        dv = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
        dk = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
        # k and v stay in SRAM throughout
        k = tl.load(k_ptrs)
        v = tl.load(v_ptrs)
        # loop over rows
        for start_m in range(lo, num_block * BLOCK_M, BLOCK_M):
            offs_m_curr = start_m + offs_m
            # load q, k, v, do on-chip
            q = tl.load(q_ptrs)
            # recompute p = softmax(qk, dim=-1).T
            if CAUSAL:
                qk = tl.where(offs_m_curr[:, None] >= (offs_n[None, :]), float(0.), float("-inf"))
            else:
                qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
            qk += tl.dot(q, tl.trans(k))
            qk *= qk_scale
            l_i = tl.load(l_ptrs + offs_m_curr)
            p = tl.math.exp2(qk - l_i[:, None])
            # compute dv
            do = tl.load(do_ptrs)
            dv += tl.dot(tl.trans(p.to(Q.dtype.element_ty)), do)
            # compute dp = dot(v, do)
            Di = tl.load(D_ptrs + offs_m_curr)
            dp = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32) - Di[:, None]
            dp += tl.dot(do, tl.trans(v))
            # compute ds = p * (dp - delta[:, None])
            ds = p * dp * sm_scale
            # compute dk = dot(ds.T, q)
            dk += tl.dot(tl.trans(ds.to(Q.dtype.element_ty)), q)
            # compute dq
            dq = tl.load(dq_ptrs)
            dq += tl.dot(ds.to(Q.dtype.element_ty), k)
            tl.store(dq_ptrs, dq)
            # increment pointers
            dq_ptrs += BLOCK_M * stride_qm
            q_ptrs += BLOCK_M * stride_qm
            do_ptrs += BLOCK_M * stride_qm
        # write-back
        dv_ptrs = DV + (offs_n[:, None] * stride_qm + offs_k[None, :] * stride_qk)
        dk_ptrs = DK + (offs_n[:, None] * stride_kn + offs_k[None, :] * stride_kk)
        tl.store(dv_ptrs, dv)
        tl.store(dk_ptrs, dk)


empty = torch.empty(128, device="cuda")


class _mgqa_attention(torch.autograd.Function):

    @staticmethod
    def forward(
        ctx, 
        q, 
        k, 
        v, 
        causal, 
        sm_scale, 
        num_groups
    ):
        # shape constraints
        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
        assert Lq == Lk and Lk == Lv
        assert Lk in {16, 32, 64, 128}
        o = torch.empty_like(q)
        BLOCK_M = 128
        BLOCK_N = 64
        grid = (triton.cdiv(q.shape[2], BLOCK_M), q.shape[0] * q.shape[1], 1)
        L = torch.empty((q.shape[0] * q.shape[1], q.shape[2]), device=q.device, dtype=torch.float32)

        num_warps = 4 if Lk <= 64 else 8
        
        #divide query heads into G groups
        q_groups = torch.chunk(q, num_groups, dim=1)
        k_groups = torch.chunk(k, num_groups, dim=1)
        v_groups = torch.chunk(v, num_groups, dim=1)

        for i in range(num_groups):    
            _fwd_kernel[grid](
                q_groups[i], k_groups[i], v_groups[i], sm_scale,
                L,
                o,