File size: 8,412 Bytes
2f9282b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
# -*- coding: utf-8 -*-
# Copyright (c) 2024, Songlin Yang, Yu Zhang

from typing import Optional, Tuple

import torch
import triton
import triton.language as tl

from fla.ops.linear_attn.utils import normalize_output
from fla.utils import contiguous


@triton.jit
def fused_recurrent_linear_attn_fwd_kernel(
    q,  # query [B, H, L, K]
    k,  # key [B, H, L, V]
    v,  # value [B, H, L, V]
    o,  # output [B, H, L, V]
    h0,
    ht,  # final hidden state [B, H, K, V]

    s_k_h,  # stride size: L * K
    s_v_h,  # stride size: L * V

    scale,
    B,  # batch size
    H,  # H
    T,  # T
    K: tl.constexpr,  # K
    V: tl.constexpr,  # V
    BK: tl.constexpr,  # BLOCK SIZE along the K dimension
    BV: tl.constexpr,  # BLOCK SIZE along the V dimension
    USE_INITIAL_STATE: tl.constexpr,  # whether to use initial state
    STORE_FINAL_STATE: tl.constexpr,  # whether to store final state
):
    # indices
    i_v, i_k, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)

    p_q = q + i_bh * s_k_h + i_k * BK + tl.arange(0, BK)
    p_k = k + i_bh * s_k_h + i_k * BK + tl.arange(0, BK)
    p_v = v + i_bh * s_v_h + i_v * BV + tl.arange(0, BV)
    p_o = o + (i_bh + i_k * B * H) * s_v_h + i_v * BV + tl.arange(0, BV)

    mask_bk = (i_k * BK + tl.arange(0, BK)) < K
    mask_bv = (i_v * BV + tl.arange(0, BV)) < V
    mask_kv = mask_bk[None, :] & mask_bv[:, None]

    b_h = tl.zeros([BV, BK], dtype=tl.float32)

    if USE_INITIAL_STATE:
        p_h0 = h0 + i_bh * K * V + (i_k * BK + tl.arange(0, BK)[None, :]) * V + (i_v * BV + tl.arange(0, BV)[:, None])
        b_h += tl.load(p_h0, mask=mask_kv, other=0).to(tl.float32)

    for _ in range(0, T):
        b_k = tl.load(p_k, mask=mask_bk, other=0).to(tl.float32)
        b_v = tl.load(p_v, mask=mask_bv, other=0).to(tl.float32)
        b_q = tl.load(p_q, mask=mask_bk, other=0).to(tl.float32) * scale

        b_h += b_k[None, :] * b_v[:, None]
        b_o = b_h * b_q[None, :]
        b_o = tl.sum(b_o, axis=1)
        tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_bv)

        p_q += K
        p_k += K
        p_o += V
        p_v += V

    if STORE_FINAL_STATE:
        p_ht = ht + i_bh * K * V + (i_k * BK + tl.arange(0, BK)[None, :]) * V + (i_v * BV + tl.arange(0, BV)[:, None])
        tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_kv)


# Similar to Algorithm1 of https://arxiv.org/abs/2006.16236
@triton.jit
def fused_recurrent_linear_attn_bwd_kernel(
    q,  # query [B, H, L, K]
    k,  # key [B, H, L, V]
    v,  # value [B, H, L, V]

    do,  # gradient of output [B, H, L, V]
    dq,  # gradient of query [NV, B, H, L, K]
    dk,  # gradient of key [NV, B, H, L, K]
    dv,  # gradient of value [NK, B, H, L, V]
    h0,  # initial hidden state initialization [B, H, K, V]

    s_k_h,  # stride size: L * K
    s_v_h,  # stride size: L * V
    scale,  # K ** -0.5

    B,  # B
    H,  # H
    T,  # T
    K: tl.constexpr,  # K
    V: tl.constexpr,  # V
    BK: tl.constexpr,  # BLOCK SIZE along the K dimension
    BV: tl.constexpr,  # BLOCK SIZE along the V dimension
    USE_INITIAL_STATE: tl.constexpr,  # whether to use initial state
):
    i_v, i_k, i_bh = tl.program_id(0), tl.program_id(1), tl.program_id(2)

    p_q = q + i_bh * s_k_h + i_k * BK + tl.arange(0, BK)
    p_k = k + i_bh * s_k_h + i_k * BK + tl.arange(0, BK)
    p_v = v + i_bh * s_v_h + i_v * BV + tl.arange(0, BV)
    p_do = do + i_bh * s_v_h + i_v * BV + tl.arange(0, BV)

    p_dq = dq + (i_bh + i_v * B * H) * s_k_h + i_k * BK + tl.arange(0, BK)
    mask_bk = i_k * BK + tl.arange(0, BK) < K
    mask_bv = i_v * BV + tl.arange(0, BV) < V

    b_h = tl.zeros([BK, BV], dtype=tl.float32)

    if USE_INITIAL_STATE:
        mask_kv = mask_bk[:, None] & mask_bv[None, :]
        p_h0 = h0 + i_bh * K * V + (i_k * BK + tl.arange(0, BK)[:, None]) * V + (i_v * BV + tl.arange(0, BV)[None, :])
        b_h += tl.load(p_h0, mask=mask_kv, other=0).to(tl.float32)

    for _ in range(0, T):
        b_k = tl.load(p_k, mask=mask_bk, other=0).to(tl.float32)
        b_v = tl.load(p_v, mask=mask_bv, other=0).to(tl.float32)
        b_do = tl.load(p_do, mask=mask_bv, other=0).to(tl.float32)

        b_h += b_k[:, None] * b_v[None, :]
        _d_q = b_h * b_do[None, :]
        d_q = tl.sum(_d_q, axis=1) * scale
        tl.store(p_dq, d_q.to(p_dq.dtype.element_ty), mask=mask_bk)

        p_k += K
        p_do += V
        p_v += V
        p_dq += K

    # sync threads
    tl.debug_barrier()

    p_q = q + i_bh * s_k_h + i_k * BK + tl.arange(0, BK) + (T - 1) * K
    p_k = k + i_bh * s_k_h + i_k * BK + tl.arange(0, BK) + (T - 1) * K
    p_do = do + i_bh * s_v_h + i_v * BV + tl.arange(0, BV) + (T - 1) * V
    p_v = v + i_bh * s_v_h + i_v * BV + tl.arange(0, BV) + (T - 1) * V
    p_dk = dk + (i_bh + i_v * B * H) * s_k_h + i_k * BK + tl.arange(0, BK) + (T - 1) * K
    p_dv = dv + (i_bh + i_k * B * H) * s_v_h + i_v * BV + tl.arange(0, BV) + (T - 1) * V
    d_h = tl.zeros([BK, BV], dtype=tl.float32)

    for _ in range(T):
        b_do = tl.load(p_do, mask=mask_bv, other=0).to(tl.float32)
        b_q = tl.load(p_q, mask=mask_bk, other=0).to(tl.float32) * scale
        b_k = tl.load(p_k, mask=mask_bk, other=0).to(tl.float32)
        b_v = tl.load(p_v, mask=mask_bv, other=0).to(tl.float32)
        d_h += b_q[:, None] * b_do[None, :]
        d_k = tl.sum(d_h * b_v[None, :], axis=1)
        d_v = tl.sum(d_h * b_k[:, None], axis=0)

        tl.store(p_dk, d_k.to(p_dk.dtype.element_ty), mask=mask_bk)
        tl.store(p_dv, d_v.to(p_dv.dtype.element_ty), mask=mask_bv)

        p_do -= V
        p_q -= K
        p_k -= K
        p_v -= V
        p_dk -= K
        p_dv -= V


class FusedRecurrentLinearAttentionFunction(torch.autograd.Function):

    @staticmethod
    @contiguous
    def forward(ctx, q, k, v, scale, initial_state=None, output_final_state=False):
        B, H, T, K = q.shape
        V = v.shape[-1]

        BK, BV = min(K, 32), min(V, 32)
        NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
        num_warps = 1
        num_stages = 1

        o = q.new_empty(NK, B, H, T, V)
        final_state = q.new_empty(B, H, K, V) if output_final_state else None

        grid = (NV, NK, B * H)
        fused_recurrent_linear_attn_fwd_kernel[grid](
            q, k, v, o, initial_state, final_state,
            q.stride(1),
            v.stride(1), scale,
            B=B, H=H, T=T, K=K, V=V, BK=BK, BV=BV,
            USE_INITIAL_STATE=initial_state is not None,
            STORE_FINAL_STATE=final_state is not None,
            num_warps=num_warps,
            num_stages=num_stages
        )

        o = o.sum(0)
        ctx.save_for_backward(q, k, v, initial_state)
        ctx.scale = scale
        return o, final_state

    @staticmethod
    @contiguous
    def backward(ctx, do, dht=None):
        q, k, v, initial_state = ctx.saved_tensors
        B, H, T, K = q.shape
        V = v.shape[-1]
        scale = ctx.scale

        BK, BV = min(K, 32), min(V, 32)
        NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
        num_warps = 1
        num_stages = 1

        dq = q.new_empty(NV, B, H, T, K)
        dk = q.new_empty(NV, B, H, T, K)
        dv = q.new_empty(NK, B, H, T, V)
        grid = (NV, NK, B * H)

        fused_recurrent_linear_attn_bwd_kernel[grid](
            q, k, v, do, dq, dk, dv, initial_state,
            q.stride(1),
            v.stride(1),
            scale,
            B=B, H=H, T=T, K=K, V=V, BK=BK, BV=BV,
            USE_INITIAL_STATE=initial_state is not None,
            num_warps=num_warps,
            num_stages=num_stages
        )
        dq = dq.sum(0)
        dk = dk.sum(0)
        dv = dv.sum(0)
        return dq, dk, dv, None, None, None


def fused_recurrent_linear_attn(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    scale: Optional[float] = None,
    initial_state: torch.Tensor = None,
    output_final_state: bool = False,
    normalize: bool = False,
    head_first: bool = True
) -> Tuple[torch.Tensor, torch.Tensor]:
    if scale is None:
        scale = q.shape[-1] ** -0.5
    if not head_first:
        q, k, v = map(lambda x: x.transpose(1, 2), (q, k, v))
    o, final_state = FusedRecurrentLinearAttentionFunction.apply(q, k, v, scale, initial_state, output_final_state)
    if normalize:
        o = normalize_output(q * scale, k, o)
    if not head_first:
        o = o.transpose(1, 2)
    return o, final_state