example_mha_inference.py

import torch
import torch.nn.functional as F
import tilelang
from tilelang.autotuner import *
import tilelang.language as T
from functools import partial

num_split = 4


@tilelang.jit(out_idx=[5], pass_configs={tilelang.PassConfigKey.TL_DISABLE_TMA_LOWER: True})
def flashattn(batch, heads, seqlen_q, seqlen_kv, dim, is_causal, block_M, block_N):
    scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
    shape_q = [batch, seqlen_q, heads, dim]
    shape_kv = [batch, seqlen_kv, heads, dim]
    part_shape = [batch, seqlen_q, heads, num_split, dim]
    dtype = "float16"
    accum_dtype = "float"

    @T.macro
    def MMA0(
        K: T.Tensor(shape_kv, dtype),
        Q_shared: T.SharedBuffer([block_M, dim], dtype),
        K_shared: T.SharedBuffer([block_N, dim], dtype),
        acc_s: T.FragmentBuffer([block_M, block_N], accum_dtype),
        k: T.int32,
        mid: T.int32,
        hid: T.int32,
        bid: T.int32,
        sid: T.int32,
    ):
        T.copy(
            K[bid, (seqlen_kv // num_split) * sid + k * block_N:(seqlen_kv // num_split) * sid +
              (k + 1) * block_N, hid, :], K_shared)
        # TODO: Handle causal split case
        if is_causal:
            for i, j in T.Parallel(block_M, block_N):
                acc_s[i, j] = T.if_then_else(mid * block_M + i >= k * block_N + j, 0,
                                             -T.infinity(acc_s.dtype))
        else:
            T.clear(acc_s)
        T.gemm(Q_shared, K_shared, acc_s, transpose_B=True, policy=T.GemmWarpPolicy.FullRow)

    @T.macro
    def MMA1(
        V: T.Tensor(shape_kv, dtype),
        V_shared: T.SharedBuffer([block_M, dim], dtype),
        acc_s_cast: T.FragmentBuffer([block_M, block_N], dtype),
        acc_o: T.FragmentBuffer([block_M, dim], accum_dtype),
        k: T.int32,
        hid: T.int32,
        bid: T.int32,
        sid: T.int32,
    ):
        T.copy(
            V[bid, (seqlen_kv // num_split) * sid + k * block_N:(seqlen_kv // num_split) * sid +
              (k + 1) * block_N, hid, :], V_shared)
        T.gemm(acc_s_cast, V_shared, acc_o, policy=T.GemmWarpPolicy.FullRow)

    @T.macro
    def Softmax(
            acc_s: T.FragmentBuffer([block_M, block_N], accum_dtype),
            acc_s_cast: T.FragmentBuffer([block_M, block_N], dtype),
            scores_max: T.FragmentBuffer([block_M], accum_dtype),
            scores_max_prev: T.FragmentBuffer([block_M], accum_dtype),
            scores_scale: T.FragmentBuffer([block_M], accum_dtype),
            scores_sum: T.FragmentBuffer([block_M], accum_dtype),
            logsum: T.FragmentBuffer([block_M], accum_dtype),
    ):
        T.copy(scores_max, scores_max_prev)
        T.fill(scores_max, -T.infinity(accum_dtype))
        T.reduce_max(acc_s, scores_max, dim=1, clear=False)
        # To do causal softmax, we need to set the scores_max to 0 if it is -inf
        # This process is called Check_inf in FlashAttention3 code, and it only need to be done
        # in the first ceil_div(kBlockM, kBlockN) steps.
        # for i in T.Parallel(block_M):
        #     scores_max[i] = T.if_then_else(scores_max[i] == -T.infinity(accum_dtype), 0, scores_max[i])
        for i in T.Parallel(block_M):
            scores_scale[i] = T.exp2(scores_max_prev[i] * scale - scores_max[i] * scale)
        for i, j in T.Parallel(block_M, block_N):
            # Instead of computing exp(x - max), we compute exp2(x * log_2(e) -
            # max * log_2(e)) This allows the compiler to use the ffma
            # instruction instead of fadd and fmul separately.
            acc_s[i, j] = T.exp2(acc_s[i, j] * scale - scores_max[i] * scale)
        T.reduce_sum(acc_s, scores_sum, dim=1)
        for i in T.Parallel(block_M):
            logsum[i] = logsum[i] * scores_scale[i] + scores_sum[i]
        T.copy(acc_s, acc_s_cast)

    @T.macro
    def Rescale(
            acc_o: T.FragmentBuffer([block_M, dim], accum_dtype),
            scores_scale: T.FragmentBuffer([block_M], accum_dtype),
    ):
        for i, j in T.Parallel(block_M, dim):
            acc_o[i, j] *= scores_scale[i]

    @T.macro
    def flash_attn_split(
            Q: T.Tensor(shape_q, dtype),
            K: T.Tensor(shape_kv, dtype),
            V: T.Tensor(shape_kv, dtype),
            glse: T.Tensor([batch, heads, num_split, seqlen_q], dtype),
            Output_partial: T.Tensor(part_shape, dtype),
    ):
        with T.Kernel(
                T.ceildiv(seqlen_q, block_M), heads * batch, num_split,
                threads=128) as (bx, by, bz):
            Q_shared = T.alloc_shared([block_M, dim], dtype)
            K_shared = T.alloc_shared([block_N, dim], dtype)
            V_shared = T.alloc_shared([block_N, dim], dtype)
            O_shared = T.alloc_shared([block_M, dim], dtype)
            acc_s = T.alloc_fragment([block_M, block_N], accum_dtype)
            acc_s_cast = T.alloc_fragment([block_M, block_N], dtype)
            acc_o = T.alloc_fragment([block_M, dim], accum_dtype)
            scores_max = T.alloc_fragment([block_M], accum_dtype)
            scores_max_prev = T.alloc_fragment([block_M], accum_dtype)
            scores_scale = T.alloc_fragment([block_M], accum_dtype)
            scores_sum = T.alloc_fragment([block_M], accum_dtype)
            logsum = T.alloc_fragment([block_M], accum_dtype)

            mid = bx
            hid = by % heads
            bid = by // heads
            sid = bz

            T.copy(Q[bid, mid * block_M:(mid + 1) * block_M, hid, :], Q_shared)
            T.fill(acc_o, 0)
            T.fill(logsum, 0)
            T.fill(scores_max, -T.infinity(accum_dtype))

            # TODO: Handle causal split case
            loop_range = (
                T.min(T.ceildiv(seqlen_kv, block_N), T.ceildiv(
                    (mid + 1) * block_M, block_N)) if is_causal else T.ceildiv(
                        (seqlen_kv // num_split), block_N))

            for k in T.Pipelined(loop_range, num_stages=2):
                MMA0(K, Q_shared, K_shared, acc_s, k, mid, hid, bid, sid)
                Softmax(acc_s, acc_s_cast, scores_max, scores_max_prev, scores_scale, scores_sum,
                        logsum)
                Rescale(acc_o, scores_scale)
                MMA1(V, V_shared, acc_s_cast, acc_o, k, hid, bid, sid)
            for i, j in T.Parallel(block_M, dim):
                acc_o[i, j] /= logsum[i]
            for i in T.Parallel(block_M):
                logsum[i] = T.log2(logsum[i]) + scores_max[i] * scale
            T.copy(logsum, glse[bid, hid, sid, mid * block_M:(mid + 1) * block_M])
            T.copy(acc_o, O_shared)
            T.copy(O_shared, Output_partial[bid, mid * block_M:(mid + 1) * block_M, hid, sid, :])

    @T.macro
    def combine(
            glse: T.Tensor([batch, heads, num_split, seqlen_q], dtype),
            Output_partial: T.Tensor(part_shape, dtype),
            Output: T.Tensor(shape_q, dtype),
    ):
        with T.Kernel(T.ceildiv(seqlen_q, block_M), heads, batch, threads=128) as (bx, by, bz):
            po_local = T.alloc_fragment([block_M, dim], dtype)
            po_shared = T.alloc_shared([block_M, dim], dtype)
            o_accum_local = T.alloc_fragment([block_M, dim], accum_dtype)
            o_shared = T.alloc_shared([block_M, dim], dtype)
            lse_local = T.alloc_fragment([num_split, block_M], dtype)
            lse_local_split = T.alloc_fragment([block_M], accum_dtype)
            lse_logsum_local = T.alloc_fragment([block_M], accum_dtype)
            lse_max_local = T.alloc_fragment([block_M], accum_dtype)
            scale_local = T.alloc_fragment([block_M], accum_dtype)

            T.annotate_layout({
                o_accum_local: T.Fragment(o_accum_local.shape, forward_thread_fn=lambda i, j: i),
                o_shared: tilelang.layout.make_swizzled_layout(o_shared),
                po_shared: tilelang.layout.make_swizzled_layout(po_shared),
            })

            T.clear(lse_logsum_local)
            T.clear(o_accum_local)
            T.copy(glse[
                bz,
                by,
                :,
                bx * block_M:(bx + 1) * block_M,
            ], lse_local)
            T.reduce_max(lse_local, lse_max_local, dim=0, clear=False)
            for k in T.Pipelined(num_split):
                T.copy(lse_local[k, :], lse_local_split)
                for i in T.Parallel(block_M):
                    lse_logsum_local[i] += T.exp2(lse_local_split[i] - lse_max_local[i])
            for i in T.Parallel(block_M):
                lse_logsum_local[i] = T.log2(lse_logsum_local[i]) + lse_max_local[i]
            for k in T.Pipelined(num_split, num_stages=2):
                T.copy(Output_partial[bz, bx * block_M:(bx + 1) * block_M, by, k, :], po_shared)
                T.copy(po_shared, po_local)
                for i in T.Parallel(block_M):
                    lse_local_split[i] = lse_local[k, i]
                for i in T.Parallel(block_M):
                    scale_local[i] = T.exp2(lse_local_split[i] - lse_logsum_local[i])
                for i, j in T.Parallel(block_M, dim):
                    o_accum_local[i, j] += po_local[i, j] * scale_local[i]
            T.copy(o_accum_local, o_shared)
            T.copy(o_shared, Output[bz, bx * block_M:(bx + 1) * block_M, by, :])

    @T.prim_func
    def flashattn_mha_inference(
            Q: T.Tensor(shape_q, dtype),
            K: T.Tensor(shape_kv, dtype),
            V: T.Tensor(shape_kv, dtype),
            glse: T.Tensor([batch, heads, num_split, seqlen_q], dtype),
            Output_partial: T.Tensor(part_shape, dtype),  # [batch, seqlen_q, heads, num_split, dim]
            Output: T.Tensor(shape_q, dtype),
    ):
        flash_attn_split(Q, K, V, glse, Output_partial)
        combine(glse, Output_partial, Output)

    return flashattn_mha_inference


def ref_program(Q, K, V, glse, Output_partial, causal):
    assert causal is False
    dim = Q.size(-1)
    scores = torch.einsum('bqhd,bkhd->bhqk', Q, K)
    scores = scores / torch.sqrt(torch.tensor(dim, dtype=scores.dtype))
    attention_weights = F.softmax(scores, dim=-1)
    output = torch.einsum('bhqk,bkhd->bqhd', attention_weights, V)
    return output


def reduce_ref(Q, K, V, glse, Output_partial, causal):
    o = torch.empty_like(Output_partial[:, :, :, 0, :]).fill_(0)
    lse_logsum = torch.empty_like(glse[:, :, 0, :]).fill_(0)  # [batch, seqlen_q, heads]
    lse_max = glse.max(dim=2, keepdim=False).values
    for ks in range(num_split):
        lse = glse[:, :, ks, :]
        lse_logsum += torch.exp2(lse - lse_max)
    lse_logsum = torch.log2(lse_logsum) + lse_max
    for ks in range(num_split):
        lse = glse[:, :, ks, :]
        scale = torch.exp2(lse - lse_logsum)  # [batch, heads, seqlen_q]
        o += Output_partial[:, :, :, ks, :] * scale[:, :, :, None].transpose(1, 2)
    return o.to(torch.float16)


def flash_split_ref(Q, K, V, causal):
    # [batch, seqlen_q, heads, dim]
    batch = Q.size(0)
    block_M = Q.size(1)
    nheads = Q.size(2)
    dim = Q.size(3)
    block_N = 128
    seqlen_kv = K.size(1)

    scale = (1.0 / dim)**0.5 * 1.44269504  # log2(e)
    acc_s = torch.empty((batch, nheads, block_M, block_N), device="cuda", dtype=torch.float)
    acc_s_cast = torch.empty((batch, nheads, block_M, block_N), device="cuda", dtype=torch.float16)
    acc_o = torch.empty((batch, block_M, nheads, dim), device="cuda", dtype=torch.float)
    scores_max = torch.empty((batch, nheads, block_M), device="cuda", dtype=torch.float)
    scores_max_prev = torch.empty((batch, nheads, block_M), device="cuda", dtype=torch.float)
    scores_scale = torch.empty((batch, nheads, block_M), device="cuda", dtype=torch.float)
    scores_sum = torch.empty((batch, nheads, block_M), device="cuda", dtype=torch.float)
    logsum = torch.empty((batch, nheads, block_M), device="cuda", dtype=torch.float)
    gacc_o = torch.empty((num_split, batch, block_M, nheads, dim), device="cuda", dtype=torch.float)
    glogsum = torch.empty((num_split, batch, nheads, block_M), device="cuda", dtype=torch.float)

    Q_ = Q * scale

    for ks in range(num_split):
        acc_o.fill_(0)
        logsum.fill_(0)
        scores_max.fill_(float('-inf'))
        scores_max_prev.fill_(float('-inf'))
        for i in range(int((seqlen_kv // num_split) / block_N)):
            acc_s.fill_(0)
            acc_s = torch.einsum('bqhd,bkhd->bhqk', Q_,
                                 K[:, (seqlen_kv // num_split) * ks +
                                   i * block_N:(seqlen_kv // num_split) * ks +
                                   (i + 1) * block_N, :, :])  # [batch, seqlen, nheads, block_N]
            scores_max_prev = scores_max
            scores_max = acc_s.max(dim=-1, keepdim=False).values  # [blockM]
            scores_scale = torch.exp2(scores_max_prev - scores_max)
            acc_o *= scores_scale[:, :, :, None].transpose(1, 2)
            acc_s = torch.exp2(acc_s - scores_max[:, :, :, None])
            acc_s_cast = acc_s.to(torch.float16)
            acc_o += torch.einsum(
                'bhqk,bkhd->bqhd', acc_s_cast,
                V[:, (seqlen_kv // num_split) * ks + i * block_N:(seqlen_kv // num_split) * ks +
                  (i + 1) * block_N, :, :])
            scores_sum = acc_s.sum(dim=-1, keepdim=False)
            logsum = logsum * scores_scale + scores_sum
        acc_o /= logsum[:, :, :, None].transpose(1, 2)
        logsum = torch.log2(logsum) + scores_max
        gacc_o[ks, :, :, :, :] = acc_o
        glogsum[ks, :, :, :] = logsum

    return glogsum.to(torch.float16).permute(1, 2, 0,
                                             3), gacc_o.to(torch.float16).permute(1, 2, 3, 0, 4)


def main():
    BATCH, H, Q_CTX, KV_CTX, D_HEAD = 1, 32, 128, 8192, 128
    causal = False
    flops_per_matmul = 2.0 * BATCH * H * Q_CTX * KV_CTX * D_HEAD
    total_flops = 2 * flops_per_matmul
    if causal:
        total_flops *= 0.5
    BLOCK_M = 128
    BLOCK_N = 64  # if D_HEAD <= 128 else 32
    kernel = flashattn(BATCH, H, Q_CTX, KV_CTX, D_HEAD, causal, BLOCK_M, BLOCK_N)
    ref_fn = partial(ref_program, causal=causal)
    print(kernel.get_kernel_source())
    profiler = kernel.get_profiler(tensor_supply_type=tilelang.TensorSupplyType.Normal)
    profiler.assert_allclose(ref_fn, rtol=0.01, atol=0.01)
    print("All checks passed!")

    latency = profiler.do_bench(ref_fn, warmup=500)
    print("{:.2f} ms".format(latency))
    print("{:.2f} TFlops".format(total_flops / latency * 1e-9))
    latency = profiler.do_bench(n_warmup=10, n_repeat=10)
    print("{:.4f} ms".format(latency))
    print("{:.2f} TFlops".format(total_flops / latency * 1e-9))


if __name__ == "__main__":
    main()