sparse_flash_attn_developer_vid_reduce.py

import tilelang
from tilelang import language as T
import torch

torch.set_default_device("npu")
torch.manual_seed(0)

tilelang.disable_cache()

pass_configs = {
    tilelang.PassConfigKey.TL_ASCEND_AUTO_CV_COMBINE: True,
    tilelang.PassConfigKey.TL_ASCEND_AUTO_CV_SYNC: True,
    tilelang.PassConfigKey.TL_ASCEND_AUTO_SYNC: True,
    tilelang.PassConfigKey.TL_ASCEND_MEMORY_PLANNING: True,
}


@tilelang.jit(out_idx=[3], pass_configs=pass_configs)
def sparse_attention_fwd(
    heads,
    dim,
    tail_dim,
    topk,
    kv_stride,
    kv_group=1,
    sm_scale=None,
    is_causal=True,
    block_I=64,
):
    assert dim == tilelang.math.next_power_of_2(dim), f"haven't check padding correctness yet, dim={dim}"
    assert tail_dim == tilelang.math.next_power_of_2(tail_dim), f"haven't check padding correctness yet, dim={tail_dim}"
    assert is_causal, "non-casual is not supported"
    assert topk % block_I == 0, "otherwise will load some index=0 thus causing wrong kv to be loaded"

    # NOTE: ascend only support exp interface instead of exp2
    sm_scale = (1.0 / (dim + tail_dim)) ** 0.5 if sm_scale is None else sm_scale

    batch = 1  # T.symbolic("batch")
    seq_len = 128  # T.symbolic("seq_len")

    seq_len_kv = 32768  # T.symbolic("seq_len_kv")
    head_kv = heads // kv_group
    q_shape = [batch, seq_len, heads, dim + tail_dim]
    kv_shape = [batch, seq_len_kv, kv_group, dim + tail_dim]
    o_shape = [batch, seq_len, heads, dim]
    indices_shape = [batch, seq_len, kv_group, topk]
    # lse_shape = [batch, seq_len, heads]
    indices_dtype = "int32"
    dtype = "float16"
    accum_dtype = "float"

    H = head_kv

    padded_H = max(tilelang.math.next_power_of_2(head_kv), 16)
    if padded_H != H:
        assert kv_group == 1, (
            "here we solve the H padding automatically, other wise you should handle Q copy and Output copy with your mask (when kv_group == 1, use g_i * padded_H:(g_i+1) * padded_H would be handled automatically)"
        )

    BI = block_I
    NI = tilelang.cdiv(topk, block_I)
    D = dim
    D_tail = tail_dim

    if head_kv > 64:
        assert head_kv % 64 == 0, "head_kv should be a multiple of 64"
        REPLICATE_H = head_kv // 64
    else:
        REPLICATE_H = 1

    H_per_block = padded_H if REPLICATE_H == 1 else 64

    v_block = H_per_block

    block_num = seq_len * REPLICATE_H * batch * kv_group

    @T.prim_func
    def main(
        Q: T.Tensor(q_shape, dtype),  # type: ignore
        KV: T.Tensor(kv_shape, dtype),  # type: ignore
        Indices: T.Tensor(indices_shape, indices_dtype),  # type: ignore
        Output: T.Tensor(o_shape, dtype),  # type: ignore
    ):
        with T.Kernel(block_num, threads=2, is_npu=True) as (cid):
            bx = cid % (seq_len * REPLICATE_H)
            by = cid // (seq_len * REPLICATE_H) % batch
            bz = cid // (seq_len * REPLICATE_H) // batch % kv_group

            q_l1 = T.alloc_shared([H_per_block, D], dtype)
            q_tail_l1 = T.alloc_shared([H_per_block, D_tail], dtype)
            kv_l1 = T.alloc_shared([BI, D], dtype)
            kv_tail_l1 = T.alloc_shared([BI, D_tail], dtype)
            acc_s_l1 = T.alloc_shared([H_per_block, BI], dtype)

            acc_s_l0c = T.alloc_fragment([H_per_block, BI], accum_dtype)
            acc_o_l0c = T.alloc_fragment([H_per_block, D], accum_dtype)

            ## 2. Vector
            acc_o = T.alloc_shared([v_block, D], accum_dtype)
            sumexp = T.alloc_shared([v_block], accum_dtype)
            m_i = T.alloc_shared([v_block], accum_dtype)
            indices_ub_ = T.alloc_shared([BI], indices_dtype)
            kv_ub = T.alloc_shared([D], dtype)
            kv_tail_ub = T.alloc_shared([D_tail], dtype)
            acc_s_ub = T.alloc_shared([v_block, BI], accum_dtype)
            m_i_prev = T.alloc_shared([v_block], accum_dtype)
            acc_s_ub_ = T.alloc_shared([v_block, BI], accum_dtype)
            sumexp_i_ub = T.alloc_shared([v_block], accum_dtype)
            acc_s_half = T.alloc_shared([v_block, BI], dtype)
            acc_o_ub = T.alloc_shared([v_block, D], accum_dtype)
            acc_o_half = T.alloc_shared([v_block, D], dtype)

            b_i = by
            g_i = bz

            s_i = bx // REPLICATE_H

            H0 = g_i * padded_H + (0 if REPLICATE_H == 1 else (bx % REPLICATE_H) * 64)
            H1 = H0 + H_per_block

            T.copy(Q[b_i, s_i, H0:H1, :D], q_l1)
            T.copy(Q[b_i, s_i, H0:H1, D:], q_tail_l1)
            T.tile.fill(acc_o, 0.0)
            T.tile.fill(sumexp, 0.0)
            T.tile.fill(m_i, -(2.0**30))
            for i_i in T.serial(NI):
                T.gemm_v0(q_l1, kv_l1, acc_s_l0c, transpose_B=True, init=True)
                T.gemm_v0(q_tail_l1, kv_tail_l1, acc_s_l0c, transpose_B=True)

                T.copy(acc_s_l0c, acc_s_ub_)

                T.gemm_v0(acc_s_l1, kv_l1, acc_o_l0c, init=True)

                T.copy(acc_o_l0c, acc_o_ub)

                T.copy(Indices[b_i, s_i, g_i, i_i * BI : i_i * BI + BI], indices_ub_)

                for bi_i in range(BI):
                    T.copy(KV[b_i, indices_ub_[bi_i], g_i, :D], kv_ub)
                    T.copy(KV[b_i, indices_ub_[bi_i], g_i, D:], kv_tail_ub)
                    T.copy(kv_ub, kv_l1[bi_i, :])
                    T.copy(kv_tail_ub, kv_tail_l1[bi_i, :])

                T.tile.fill(acc_s_ub, 0.0)

                T.copy(m_i, m_i_prev)

                for i, j in T.Parallel(v_block, BI):
                    acc_s_ub[i, j] = acc_s_ub[i, j] + acc_s_ub_[i, j]

                for i, j in T.Parallel(v_block, BI):
                    acc_s_ub[i, j] = acc_s_ub[i, j] * sm_scale

                T.reduce_max(acc_s_ub, m_i, dim=-1)

                for i in T.Parallel(v_block):
                    m_i[i] = T.max(m_i[i], m_i_prev[i])
                    m_i_prev[i] = m_i_prev[i] - m_i[i]
                    m_i_prev[i] = T.exp(m_i_prev[i])

                for h_i, j in T.Parallel(v_block, BI):
                    acc_s_ub[h_i, j] = acc_s_ub[h_i, j] - m_i[h_i]
                    acc_s_ub[h_i, j] = T.exp(acc_s_ub[h_i, j])

                T.reduce_sum(acc_s_ub, sumexp_i_ub, dim=-1)

                for i in T.Parallel(v_block):
                    sumexp[i] *= m_i_prev[i]
                    sumexp[i] += sumexp_i_ub[i]

                for h_i, j in T.Parallel(v_block, D):
                    acc_o[h_i, j] = acc_o[h_i, j] * m_i_prev[h_i]

                T.copy(acc_s_ub, acc_s_half)

                T.copy(acc_s_half, acc_s_l1)

                for i, j in T.Parallel(v_block, D):
                    acc_o[i, j] += acc_o_ub[i, j]

            for h_i, j in T.Parallel(v_block, D):
                acc_o[h_i, j] = acc_o[h_i, j] / sumexp[h_i]

            T.copy(acc_o, acc_o_half)
            T.copy(acc_o_half, Output[b_i, s_i, H0 : H0 + v_block, :])

    return main


func = sparse_attention_fwd(
    heads=128,
    dim=512,
    tail_dim=64,
    topk=2048,
    kv_stride=1,
)


def ref_sparse_attention_fwd_interface(q, kv, indices, q_start_index_s, kv_stride=4, sm_scale=None, is_casual=True):
    q = q.float()
    kv = kv.float()
    indices = indices.transpose(1, 2)
    b, sq, h, dim_q = q.shape
    b, sk, g, _ = kv.shape
    if q_start_index_s is None:
        q_start_index_s = sk * kv_stride - sq

    assert kv.shape[-1] == 576, "you should assign dim otherwise"
    dim = 512
    k = kv
    v = kv[..., :dim]

    b, _, _, dim_v = v.shape
    # num_kv_per_index = 1
    g_index = g
    h_index = h // g
    compressed_casual_mask = torch.arange(q_start_index_s, sq + q_start_index_s, dtype=torch.int32).view(-1, 1) >= torch.arange(
        kv_stride - 1, sk * kv_stride, kv_stride, dtype=torch.int32
    ).view(1, -1)

    mask = q.new_zeros(b, g_index, sq, sk + 1, dtype=torch.bool).scatter(3, indices.long(), 1)
    mask = mask[..., :-1]
    mask = mask & compressed_casual_mask.view(1, 1, sq, sk)
    mask[:, :, : kv_stride - 1, 0] = True
    mask = mask.view(b, g_index, 1, sq, sk)

    q = q.view(b, sq, g, -1, dim_q)
    score = torch.einsum("bmghd,bngd->bghmn", q, k)
    sm_scale = dim_q**-0.5 if sm_scale is None else sm_scale
    score = score.masked_fill(~mask, float("-inf")).mul(sm_scale)
    p = score.softmax(dim=-1)
    p = p.view(b, g_index, h_index, -1, sq, sk)
    p = p.view(b, g, -1, sq, sk)
    o = torch.einsum("bghmn,bngd->bmghd", p.type(v.dtype), v)
    o = o.reshape(b, sq, h, dim_v)
    return o.to(torch.float16)


B, S, SKV, H, HKV, DQK, DV, topk = 1, 128, 32768, 128, 1, 576, 512, 2048
dtype = torch.float16

KV_stride = 1
q_start_s_index = 4096 * 7

q = torch.randn((B, S, H, DQK), dtype=dtype)
kv = torch.randn((B, SKV, HKV, DQK), dtype=dtype)
indices = torch.full((B, S, HKV, topk), SKV, dtype=torch.int32)
for b in range(B):
    for t in range(S):
        for h in range(HKV):
            i_i = torch.randperm(max(1, ((t + q_start_s_index) // KV_stride)))[:topk]
            indices[b, t, h, : len(i_i)] = i_i

torch.npu.synchronize()
print("init successful!")

output = func(q, kv, indices)

torch.npu.synchronize()

ref_output = ref_sparse_attention_fwd_interface(q, kv, indices, q_start_s_index, KV_stride)
torch.npu.synchronize()
torch.testing.assert_close(ref_output, output, rtol=1e-2, atol=1e-2)

print("Test Passed!")