ColossalAI/colossalai/kernel/triton/context_attention.py

import math

import torch

try:
    import triton
    import triton.language as tl

    HAS_TRITON = True
except ImportError:
    HAS_TRITON = False
    print("please install triton from https://github.com/openai/triton")

if HAS_TRITON:
    """
    this function is modified from
    https://github.com/ModelTC/lightllm/blob/f093edc20683ac3ea1bca3fb5d8320a0dd36cf7b/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py#L10
    """

    @triton.jit
    def _context_flash_attention_kernel(
        Q,
        K,
        V,
        sm_scale,
        B_Start_Loc,
        B_Seqlen,
        TMP,
        alibi_ptr,
        Out,
        stride_qbs,
        stride_qh,
        stride_qd,
        stride_kbs,
        stride_kh,
        stride_kd,
        stride_vbs,
        stride_vh,
        stride_vd,
        stride_obs,
        stride_oh,
        stride_od,
        stride_tmp_b,
        stride_tmp_h,
        stride_tmp_s,
        # suggtest set-up 64, 128, 256, 512
        BLOCK_M: tl.constexpr,
        BLOCK_DMODEL: tl.constexpr,
        BLOCK_N: tl.constexpr,
    ):
        batch_id = tl.program_id(0)
        cur_head = tl.program_id(1)
        start_m = tl.program_id(2)

        # initialize offsets
        offs_n = tl.arange(0, BLOCK_N)
        offs_d = tl.arange(0, BLOCK_DMODEL)
        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)

        # get batch info
        cur_batch_seq_len = tl.load(B_Seqlen + batch_id)
        cur_batch_start_index = tl.load(B_Start_Loc + batch_id)
        block_start_loc = BLOCK_M * start_m

        load_p_ptrs = (
            Q
            + (cur_batch_start_index + offs_m[:, None]) * stride_qbs
            + cur_head * stride_qh
            + offs_d[None, :] * stride_qd
        )
        q = tl.load(load_p_ptrs, mask=offs_m[:, None] < cur_batch_seq_len, other=0.0)

        k_ptrs = K + offs_n[None, :] * stride_kbs + cur_head * stride_kh + offs_d[:, None] * stride_kd
        v_ptrs = V + offs_n[:, None] * stride_vbs + cur_head * stride_vh + offs_d[None, :] * stride_vd
        t_ptrs = TMP + batch_id * stride_tmp_b + cur_head * stride_tmp_h + offs_m * stride_tmp_s

        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)

        if alibi_ptr is not None:
            alibi_m = tl.load(alibi_ptr + cur_head)

        block_mask = tl.where(block_start_loc < cur_batch_seq_len, 1, 0)

        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
            start_n = tl.multiple_of(start_n, BLOCK_N)
            k = tl.load(
                k_ptrs + (cur_batch_start_index + start_n) * stride_kbs,
                mask=(start_n + offs_n[None, :]) < cur_batch_seq_len,
                other=0.0,
            )

            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
            qk += tl.dot(q, k)
            qk *= sm_scale

            if alibi_ptr is not None:
                alibi_loc = offs_m[:, None] - (start_n + offs_n[None, :])
                qk -= alibi_loc * alibi_m

            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))

            m_ij = tl.max(qk, 1)
            p = tl.exp(qk - m_ij[:, None])
            l_ij = tl.sum(p, 1)
            # -- update m_i and l_i
            m_i_new = tl.maximum(m_i, m_ij)
            alpha = tl.exp(m_i - m_i_new)
            beta = tl.exp(m_ij - m_i_new)
            l_i_new = alpha * l_i + beta * l_ij
            # -- update output accumulator --
            # scale p
            p_scale = beta / l_i_new
            p = p * p_scale[:, None]
            # scale acc
            acc_scale = l_i / l_i_new * alpha
            tl.store(t_ptrs, acc_scale)
            acc_scale = tl.load(t_ptrs)
            acc = acc * acc_scale[:, None]
            # update acc
            v = tl.load(
                v_ptrs + (cur_batch_start_index + start_n) * stride_vbs,
                mask=(start_n + offs_n[:, None]) < cur_batch_seq_len,
                other=0.0,
            )

            p = p.to(v.dtype)
            acc += tl.dot(p, v)
            # update m_i and l_i
            l_i = l_i_new
            m_i = m_i_new

        off_o = (
            (cur_batch_start_index + offs_m[:, None]) * stride_obs + cur_head * stride_oh + offs_d[None, :] * stride_od
        )
        out_ptrs = Out + off_o
        tl.store(out_ptrs, acc, mask=offs_m[:, None] < cur_batch_seq_len)
        return

    @torch.no_grad()
    def bloom_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len, alibi=None):
        BLOCK = 128
        # shape constraints
        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
        assert Lq == Lk, "context process only supports equal query, key, value length"
        assert Lk == Lv, "context process only supports equal query, key, value length"
        assert Lk in {16, 32, 64, 128}

        sm_scale = 1.0 / math.sqrt(Lk)
        batch, head = b_seq_len.shape[0], q.shape[1]

        grid = (batch, head, triton.cdiv(max_input_len, BLOCK))

        num_warps = 4 if Lk <= 64 else 8

        tmp = torch.empty((batch, head, max_input_len + 256), device=q.device, dtype=torch.float32)

        _context_flash_attention_kernel[grid](
            q,
            k,
            v,
            sm_scale,
            b_start_loc,
            b_seq_len,
            tmp,
            alibi,
            o,
            q.stride(0),
            q.stride(1),
            q.stride(2),
            k.stride(0),
            k.stride(1),
            k.stride(2),
            v.stride(0),
            v.stride(1),
            v.stride(2),
            o.stride(0),
            o.stride(1),
            o.stride(2),
            tmp.stride(0),
            tmp.stride(1),
            tmp.stride(2),
            # manually setting this blcok num, we can use tuning config to futher speed-up
            BLOCK_M=BLOCK,
            BLOCK_DMODEL=Lk,
            BLOCK_N=BLOCK,
            num_warps=num_warps,
            num_stages=1,
        )
        return

    @torch.no_grad()
    def llama_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len):
        BLOCK = 128
        # shape constraints
        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
        assert Lq == Lk, "context process only supports equal query, key, value length"
        assert Lk == Lv, "context process only supports equal query, key, value length"
        assert Lk in {16, 32, 64, 128}

        sm_scale = 1.0 / math.sqrt(Lk)
        batch, head = b_seq_len.shape[0], q.shape[1]

        grid = (batch, head, triton.cdiv(max_input_len, BLOCK))

        tmp = torch.empty((batch, head, max_input_len + 256), device=q.device, dtype=torch.float32)
        num_warps = 4 if Lk <= 64 else 8
        # num_warps = 4
        _context_flash_attention_kernel[grid](
            q,
            k,
            v,
            sm_scale,
            b_start_loc,
            b_seq_len,
            tmp,
            None,
            o,
            q.stride(0),
            q.stride(1),
            q.stride(2),
            k.stride(0),
            k.stride(1),
            k.stride(2),
            v.stride(0),
            v.stride(1),
            v.stride(2),
            o.stride(0),
            o.stride(1),
            o.stride(2),
            tmp.stride(0),
            tmp.stride(1),
            tmp.stride(2),
            BLOCK_M=BLOCK,
            BLOCK_DMODEL=Lk,
            BLOCK_N=BLOCK,
            num_warps=num_warps,
            num_stages=1,
        )
        return

    @triton.jit
    def _fwd_kernel_latest(
        Q,
        K,
        V,
        sm_scale,
        B_Start_Loc,
        B_Seqlen,
        Out,
        stride_qbs,
        stride_qh,
        stride_qd,
        stride_kbs,
        stride_kh,
        stride_kd,
        stride_vbs,
        stride_vh,
        stride_vd,
        stride_obs,
        stride_oh,
        stride_od,
        kv_group_num,
        BLOCK_M: tl.constexpr,
        BLOCK_DMODEL: tl.constexpr,
        BLOCK_N: tl.constexpr,
    ):
        cur_batch = tl.program_id(0)
        cur_head = tl.program_id(1)
        start_m = tl.program_id(2)

        cur_kv_head = cur_head // kv_group_num

        cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
        cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)

        block_start_loc = BLOCK_M * start_m

        # initialize offsets
        offs_n = tl.arange(0, BLOCK_N)
        offs_d = tl.arange(0, BLOCK_DMODEL)
        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
        off_q = (
            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs
            + cur_head * stride_qh
            + offs_d[None, :] * stride_qd
        )
        off_k = offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh + offs_d[:, None] * stride_kd
        off_v = offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh + offs_d[None, :] * stride_vd

        q = tl.load(Q + off_q, mask=offs_m[:, None] < cur_batch_seq_len, other=0.0)

        k_ptrs = K + off_k
        v_ptrs = V + off_v

        # 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)

        block_mask = tl.where(block_start_loc < cur_batch_seq_len, 1, 0)

        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
            start_n = tl.multiple_of(start_n, BLOCK_N)
            # -- compute qk ----
            k = tl.load(
                k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs,
                mask=(start_n + offs_n[None, :]) < cur_batch_seq_len,
                other=0.0,
            )
            # mask = tl.load(mask_ptrs + start_n, mask=start_n + offs_n < cur_batch_end_loc, other=0.0)

            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
            qk += tl.dot(q, k)
            qk *= sm_scale
            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))

            # -- compute m_ij, p, l_ij
            m_ij = tl.max(qk, 1)
            p = tl.exp(qk - m_ij[:, None])
            l_ij = tl.sum(p, 1)
            # -- update m_i and l_i
            m_i_new = tl.maximum(m_i, m_ij)
            alpha = tl.exp(m_i - m_i_new)
            beta = tl.exp(m_ij - m_i_new)
            l_i_new = alpha * l_i + beta * l_ij
            # -- update output accumulator --
            # scale p
            p_scale = beta / l_i_new
            p = p * p_scale[:, None]
            # scale acc
            acc_scale = l_i / l_i_new * alpha
            acc = acc * acc_scale[:, None]
            # update acc
            v = tl.load(
                v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs,
                mask=(start_n + offs_n[:, None]) < cur_batch_seq_len,
                other=0.0,
            )

            p = p.to(v.dtype)
            acc += tl.dot(p, v)
            # update m_i and l_i
            l_i = l_i_new
            m_i = m_i_new
        # initialize pointers to output
        off_o = (
            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs
            + cur_head * stride_oh
            + offs_d[None, :] * stride_od
        )
        out_ptrs = Out + off_o
        tl.store(out_ptrs, acc, mask=offs_m[:, None] < cur_batch_seq_len)
        return

    @triton.jit
    def _fwd_kernel_old(
        Q,
        K,
        V,
        sm_scale,
        B_Start_Loc,
        B_Seqlen,
        TMP,  # NOTE: TMP is a scratchpad buffer to workaround a compiler bug
        Out,
        stride_qbs,
        stride_qh,
        stride_qd,
        stride_kbs,
        stride_kh,
        stride_kd,
        stride_vbs,
        stride_vh,
        stride_vd,
        stride_obs,
        stride_oh,
        stride_od,
        stride_tmp_b,
        stride_tmp_h,
        stride_tmp_s,
        kv_group_num,
        BLOCK_M: tl.constexpr,
        BLOCK_DMODEL: tl.constexpr,
        BLOCK_N: tl.constexpr,
    ):
        cur_batch = tl.program_id(0)
        cur_head = tl.program_id(1)
        start_m = tl.program_id(2)

        cur_kv_head = cur_head // kv_group_num

        cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
        cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)

        block_start_loc = BLOCK_M * start_m

        # initialize offsets
        offs_n = tl.arange(0, BLOCK_N)
        offs_d = tl.arange(0, BLOCK_DMODEL)
        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
        off_q = (
            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs
            + cur_head * stride_qh
            + offs_d[None, :] * stride_qd
        )
        off_k = offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh + offs_d[:, None] * stride_kd
        off_v = offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh + offs_d[None, :] * stride_vd
        q = tl.load(Q + off_q, mask=offs_m[:, None] < cur_batch_seq_len, other=0.0)

        k_ptrs = K + off_k
        v_ptrs = V + off_v

        t_ptrs = TMP + cur_batch * stride_tmp_b + cur_head * stride_tmp_h + offs_m * stride_tmp_s
        # t_ptrs = TMP + offs_m
        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)

        block_mask = tl.where(block_start_loc < cur_batch_seq_len, 1, 0)

        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
            start_n = tl.multiple_of(start_n, BLOCK_N)
            # -- compute qk ----
            k = tl.load(
                k_ptrs + (cur_batch_in_all_start_index + start_n) * stride_kbs,
                mask=(start_n + offs_n[None, :]) < cur_batch_seq_len,
                other=0.0,
            )

            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
            qk += tl.dot(q, k)
            qk *= sm_scale
            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk, float("-inf"))

            m_ij = tl.max(qk, 1)
            p = tl.exp(qk - m_ij[:, None])
            l_ij = tl.sum(p, 1)
            # -- update m_i and l_i
            m_i_new = tl.maximum(m_i, m_ij)
            alpha = tl.exp(m_i - m_i_new)
            beta = tl.exp(m_ij - m_i_new)
            l_i_new = alpha * l_i + beta * l_ij
            # -- update output accumulator --
            # scale p
            p_scale = beta / l_i_new
            p = p * p_scale[:, None]
            # scale acc
            acc_scale = l_i / l_i_new * alpha
            tl.store(t_ptrs, acc_scale)
            acc_scale = tl.load(t_ptrs)  # BUG: have to store and immediately load
            acc = acc * acc_scale[:, None]
            # update acc
            v = tl.load(
                v_ptrs + (cur_batch_in_all_start_index + start_n) * stride_vbs,
                mask=(start_n + offs_n[:, None]) < cur_batch_seq_len,
                other=0.0,
            )

            p = p.to(v.dtype)
            acc += tl.dot(p, v)
            # update m_i and l_i
            l_i = l_i_new
            m_i = m_i_new
        # initialize pointers to output
        off_o = (
            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs
            + cur_head * stride_oh
            + offs_d[None, :] * stride_od
        )
        out_ptrs = Out + off_o
        tl.store(out_ptrs, acc, mask=offs_m[:, None] < cur_batch_seq_len)

        return

    @torch.no_grad()
    def llama2_context_attn_fwd(q, k, v, o, b_start_loc, b_seq_len, max_input_len):
        if triton.__version__ >= "2.1.0":
            BLOCK = 128
            # 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}
            sm_scale = 1.0 / (Lq**0.5)  # 计算scale系数
            batch, head = b_seq_len.shape[0], q.shape[1]
            kv_group_num = q.shape[1] // k.shape[1]

            grid = (batch, head, triton.cdiv(max_input_len, BLOCK))  # batch, head,

            num_warps = 4 if Lk <= 64 else 8
            _fwd_kernel_latest[grid](
                q,
                k,
                v,
                sm_scale,
                b_start_loc,
                b_seq_len,
                o,
                q.stride(0),
                q.stride(1),
                q.stride(2),
                k.stride(0),
                k.stride(1),
                k.stride(2),
                v.stride(0),
                v.stride(1),
                v.stride(2),
                o.stride(0),
                o.stride(1),
                o.stride(2),
                kv_group_num=kv_group_num,
                BLOCK_M=BLOCK,
                BLOCK_DMODEL=Lk,
                BLOCK_N=BLOCK,
                num_warps=num_warps,
                num_stages=1,
            )
            return

        elif triton.__version__ == "2.0.0":
            BLOCK = 128
            # 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}

            sm_scale = 1.0 / (Lq**0.5)
            batch, head = b_seq_len.shape[0], q.shape[1]
            kv_group_num = q.shape[1] // k.shape[1]

            grid = (batch, head, triton.cdiv(max_input_len, BLOCK))
            tmp = torch.empty((batch, head, max_input_len + 256), device=q.device, dtype=torch.float32)
            num_warps = 4 if Lk <= 64 else 8
            # num_warps = 4
            _fwd_kernel_old[grid](
                q,
                k,
                v,
                sm_scale,
                b_start_loc,
                b_seq_len,
                tmp,
                o,
                q.stride(0),
                q.stride(1),
                q.stride(2),
                k.stride(0),
                k.stride(1),
                k.stride(2),
                v.stride(0),
                v.stride(1),
                v.stride(2),
                o.stride(0),
                o.stride(1),
                o.stride(2),
                tmp.stride(0),
                tmp.stride(1),
                tmp.stride(2),
                kv_group_num=kv_group_num,
                BLOCK_M=BLOCK,
                BLOCK_DMODEL=Lk,
                BLOCK_N=BLOCK,
                num_warps=num_warps,
                num_stages=1,
            )
            return