File size: 9,512 Bytes
b66f552
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

import warnings

import torch
import triton

from fla.ops.common.chunk_h import chunk_bwd_dh, chunk_fwd_h
from fla.ops.common.chunk_o import chunk_bwd_dqkwg, chunk_bwd_dv, chunk_fwd_o
from fla.ops.utils import chunk_local_cumsum
from fla.utils import autocast_custom_bwd, autocast_custom_fwd, input_guard


def chunk_simple_gla_fwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor | None = None,
    g_gamma: torch.Tensor | None = None,
    scale: float | None = None,
    initial_state: torch.Tensor | None = None,
    output_final_state: bool = False,
    cu_seqlens: torch.LongTensor | None = None,
    chunk_size: int = 64,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    h, ht = chunk_fwd_h(
        k=k,
        v=v,
        g=g,
        g_gamma=g_gamma,
        gk=None,
        gv=None,
        h0=initial_state,
        output_final_state=output_final_state,
        states_in_fp32=False,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    o = chunk_fwd_o(
        q=q,
        k=k,
        v=v,
        g=g,
        g_gamma=g_gamma,
        h=h,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    return o, ht


def chunk_simple_gla_bwd(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor,
    g_gamma: torch.Tensor,
    initial_state: torch.Tensor,
    do: torch.Tensor,
    dht: torch.Tensor,
    scale: float,
    cu_seqlens: torch.LongTensor | None = None,
    chunk_size: int = 64,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    # (SY 09/22) states_in_fp32 seems not affecting the error of dg but for safety, set to True
    h, _ = chunk_fwd_h(
        k=k,
        v=v,
        g=g,
        g_gamma=g_gamma,
        gk=None,
        gv=None,
        h0=initial_state,
        output_final_state=False,
        states_in_fp32=True,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    dh, dh0 = chunk_bwd_dh(
        q=q,
        k=k,
        v=v,
        g=g,
        g_gamma=g_gamma,
        gk=None,
        gv=None,
        do=do,
        h0=initial_state,
        dht=dht,
        scale=scale,
        states_in_fp32=True,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    dq, dk, _, dg = chunk_bwd_dqkwg(
        q=q,
        k=k,
        v=v,
        g=g,
        g_gamma=g_gamma,
        h=h,
        do=do,
        dh=dh,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    dv = chunk_bwd_dv(
        q=q,
        k=k,
        g=g,
        g_gamma=g_gamma,
        do=do,
        dh=dh,
        scale=scale,
        cu_seqlens=cu_seqlens,
        chunk_size=chunk_size,
    )
    return dq, dk, dv, dg, dh0


class ChunkSimpleGLAFunction(torch.autograd.Function):

    @staticmethod
    @input_guard
    @autocast_custom_fwd
    def forward(
        ctx,
        q,
        k,
        v,
        g,
        g_gamma,
        scale,
        initial_state,
        output_final_state,
        cu_seqlens,
    ):
        T = q.shape[1]
        chunk_size = min(64, max(16, triton.next_power_of_2(T)))

        g = chunk_local_cumsum(g, chunk_size=chunk_size, cu_seqlens=cu_seqlens) if g is not None else None
        o, ht = chunk_simple_gla_fwd(
            q=q,
            k=k,
            v=v,
            g=g,
            g_gamma=g_gamma,
            scale=scale,
            initial_state=initial_state,
            output_final_state=output_final_state,
            cu_seqlens=cu_seqlens,
            chunk_size=chunk_size,
        )
        ctx.save_for_backward(q, k, v, g, g_gamma, initial_state)
        ctx.chunk_size = chunk_size
        ctx.scale = scale
        ctx.cu_seqlens = cu_seqlens
        return o.to(q.dtype), ht

    @staticmethod
    @input_guard
    @autocast_custom_bwd
    def backward(ctx, do, dht):
        chunk_size, scale, cu_seqlens = ctx.chunk_size, ctx.scale, ctx.cu_seqlens
        q, k, v, g, g_gamma, initial_state = ctx.saved_tensors
        dq, dk, dv, dg, dh0 = chunk_simple_gla_bwd(
            q=q,
            k=k,
            v=v,
            g=g,
            g_gamma=g_gamma,
            initial_state=initial_state,
            do=do,
            dht=dht,
            scale=scale,
            cu_seqlens=cu_seqlens,
            chunk_size=chunk_size,
        )
        if g is not None:
            dg = chunk_local_cumsum(dg, chunk_size=chunk_size, reverse=True, cu_seqlens=cu_seqlens).to(g)
        else:
            dg = None
        return dq.to(q), dk.to(k), dv.to(v), dg, None, None, dh0, None, None


@torch.compiler.disable
def chunk_simple_gla(
    q: torch.Tensor,
    k: torch.Tensor,
    v: torch.Tensor,
    g: torch.Tensor | None = None,
    g_gamma: torch.Tensor | None = None,
    scale: float | None = None,
    initial_state: torch.Tensor | None = None,
    output_final_state: bool = False,
    cu_seqlens: torch.LongTensor | None = None,
    head_first: bool = False,
) -> tuple[torch.Tensor, torch.Tensor]:
    r"""
    Args:
        q (torch.Tensor):
            queries of shape `[B, T, H, K]`.
        k (torch.Tensor):
            keys of shape `[B, T, H, K]`.
        v (torch.Tensor):
            values of shape `[B, T, H, V]`.
        g (torch.Tensor):
            Forget gates of shape `[B, T, H]`.
            Compared to GLA, the gating is head-wise instead of elementwise.
        g_gamma (torch.Tensor):
            Log decay of shape `[H]`.
            Head-wise data-independent decay is used if `g_gamma` is provided.
            Only one of `g` or `g_gamma` should be provided.
        scale (Optional[float]):
            Scale factor for the attention scores.
            If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
        initial_state (Optional[torch.Tensor]):
            Initial state of shape `[N, H, K, V]` for `N` input sequences.
            For equal-length input sequences, `N` equals the batch size `B`.
            Default: `None`.
        output_final_state (Optional[bool]):
            Whether to output the final state of shape `[N, H, K, V]`. Default: `False`.
        cu_seqlens (torch.LongTensor):
            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
            consistent with the FlashAttention API.
        head_first (Optional[bool]):
            Whether the inputs are in the head-first format. Default: `False`.
            This argument has been deprecated.

    Returns:
        o (torch.Tensor):
            Outputs of shape `[B, T, H, V]`.
        final_state (torch.Tensor):
            Final state of shape `[N, H, K, V]` if `output_final_state=True` else `None`.

    Examples::
        >>> import torch
        >>> import torch.nn.functional as F
        >>> from einops import rearrange
        >>> from fla.ops.simple_gla import chunk_simple_gla
        # inputs with equal lengths
        >>> B, T, H, K, V = 4, 2048, 4, 512, 512
        >>> q = torch.randn(B, T, H, K, device='cuda')
        >>> k = torch.randn(B, T, H, K, device='cuda')
        >>> v = torch.randn(B, T, H, V, device='cuda')
        >>> g = F.logsigmoid(torch.randn(B, T, H, device='cuda'))
        >>> o, ht = chunk_simple_gla(
            q, k, v, g,
            initial_state=None,
            output_final_state=True
        )
        # for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
        >>> q, k, v, g = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, g))
        # for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
        >>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.long)
        >>> o_var, ht_var = chunk_simple_gla(
            q, k, v, g,
            initial_state=None,
            output_final_state=True,
            cu_seqlens=cu_seqlens
        )
    """
    if head_first:
        raise DeprecationWarning(
            "head_first is deprecated and will be removed in a future version. "
            "Please use head_first=False for now instead.",
        )
    if not head_first and q.shape[1] < q.shape[2]:
        warnings.warn(
            f"Input tensor shape suggests potential format mismatch: seq_len ({q.shape[1]}) < num_heads ({q.shape[2]}). "
            "This may indicate the inputs were passed in head-first format [B, H, T, ...] "
            "when head_first=False was specified. "
            "Please verify your input tensor format matches the expected shape [B, T, H, ...].",
        )
    if cu_seqlens is not None:
        if q.shape[0] != 1:
            raise ValueError(
                f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
                f"Please flatten variable-length inputs before processing.",
            )
        if initial_state is not None and initial_state.shape[0] != len(cu_seqlens) - 1:
            raise ValueError(
                f"The number of initial states is expected to be equal to the number of input sequences, "
                f"i.e., {len(cu_seqlens) - 1} rather than {initial_state.shape[0]}.",
            )
    if scale is None:
        scale = k.shape[-1] ** -0.5
    o, final_state = ChunkSimpleGLAFunction.apply(
        q,
        k,
        v,
        g,
        g_gamma,
        scale,
        initial_state,
        output_final_state,
        cu_seqlens,
    )
    return o, final_state