File size: 14,701 Bytes
eafbe80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang

from __future__ import annotations

import warnings
from typing import TYPE_CHECKING

import torch
import torch.nn as nn
from einops import rearrange
from torch.nn import functional as F

from fla.layers.rwkv6 import LoRA
from fla.modules import GroupNorm
from fla.modules.l2norm import l2_norm
from fla.modules.token_shift import token_shift
from fla.ops.rwkv7 import chunk_rwkv7, fused_mul_recurrent_rwkv7
from fla.ops.rwkv7.fused_addcmul import fused_addcmul_rwkv7
from fla.ops.rwkv7.fused_k_update import fused_k_rwkv7
from fla.ops.rwkv7.gate_output_correction import gate_output_correction

if TYPE_CHECKING:
    from fla.models.utils import Cache


class RWKV7Attention(nn.Module):

    def __init__(
        self,
        mode: str = 'chunk',
        hidden_size: int = 1024,
        head_dim: int | None = 64,
        num_heads: int | None = None,
        decay_low_rank_dim: int | None = None,
        gate_low_rank_dim: int | None = None,
        a_low_rank_dim: int | None = None,
        v_low_rank_dim: int | None = None,
        elementwise_affine: bool | None = True,
        norm_eps: float = 1e-5,
        layer_idx: int = None,
        fuse_norm: bool = False,
        value_dim: int = None,
        num_hidden_layers: int = None,
        **kwargs,
    ) -> RWKV7Attention:
        super().__init__()

        self.mode = mode
        assert mode in ['chunk', 'fused_recurrent'], f"Not supported mode `{mode}`."
        self.hidden_size = hidden_size

        self.key_dim = hidden_size
        self.value_dim = value_dim if value_dim is not None else hidden_size
        if head_dim is None and num_heads is None:
            raise ValueError("Either `head_dim` or `num_heads` must be specified.")
        elif head_dim is not None:
            self.head_dim = head_dim
            self.num_heads = int(hidden_size // head_dim)
        elif num_heads is not None:
            self.head_dim = int(hidden_size // num_heads)
            self.num_heads = num_heads
        self.head_v_dim = int(self.value_dim // self.num_heads)

        # Increase lora dimension for headdim>64
        factor = self.head_dim / 64
        if decay_low_rank_dim is None:
            decay_low_rank_dim = max(32, int(round((2.5 * (hidden_size**0.5)) * factor / 32) * 32))
            self.decay_low_rank_dim = decay_low_rank_dim
        else:
            self.decay_low_rank_dim = decay_low_rank_dim

        if gate_low_rank_dim is None:
            gate_low_rank_dim = max(32, int(round((5 * (hidden_size**0.5)) / 32) * 32))
            self.gate_low_rank_dim = gate_low_rank_dim
        else:
            self.gate_low_rank_dim = gate_low_rank_dim

        if a_low_rank_dim is None:
            a_low_rank_dim = max(32, int(round((2.5 * (hidden_size**0.5)) * factor / 32) * 32))
            self.a_low_rank_dim = a_low_rank_dim
        else:
            self.a_low_rank_dim = a_low_rank_dim

        if v_low_rank_dim is None:
            v_low_rank_dim = max(32, int(round((1.7 * (hidden_size**0.5)) * factor / 32) * 32))
            self.v_low_rank_dim = v_low_rank_dim
        else:
            self.v_low_rank_dim = v_low_rank_dim

        self.layer_idx = layer_idx
        self.num_hidden_layers = num_hidden_layers
        self.fuse_norm = fuse_norm

        self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
        self.x_r = nn.Parameter(torch.zeros(1, 1, hidden_size))
        self.x_w = nn.Parameter(torch.zeros(1, 1, hidden_size))
        self.x_k = nn.Parameter(torch.zeros(1, 1, hidden_size))
        self.x_v = nn.Parameter(torch.zeros(1, 1, hidden_size))
        self.x_a = nn.Parameter(torch.zeros(1, 1, hidden_size))
        self.x_g = nn.Parameter(torch.zeros(1, 1, hidden_size))

        self.k_k = nn.Parameter(torch.zeros(self.key_dim))
        self.k_a = nn.Parameter(torch.zeros(self.key_dim))
        self.r_k = nn.Parameter(torch.zeros(self.num_heads, self.head_dim))

        self.r_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
        self.k_proj = nn.Linear(hidden_size, self.key_dim, bias=False)
        self.v_proj = nn.Linear(hidden_size, self.value_dim, bias=False)
        self.o_proj = nn.Linear(self.value_dim, hidden_size, bias=False)

        self.w_lora = LoRA(hidden_size, self.key_dim, low_rank_dim=decay_low_rank_dim, activation='tanh')
        if self.layer_idx != 0:
            self.v_lora = LoRA(hidden_size, self.value_dim, low_rank_dim=v_low_rank_dim, activation=None)
        self.a_lora = LoRA(hidden_size, self.key_dim, low_rank_dim=a_low_rank_dim, activation=None)
        self.g_lora = LoRA(hidden_size, self.value_dim, low_rank_dim=gate_low_rank_dim, activation='sigmoid', bias=False)

        if self.fuse_norm:
            self.g_norm = GroupNorm(
                num_groups=self.num_heads,
                hidden_size=self.value_dim,
                elementwise_affine=elementwise_affine,
                eps=self.head_dim*norm_eps,
                bias=True,
            )
        else:
            self.g_norm = nn.GroupNorm(
                num_groups=self.num_heads,
                num_channels=self.value_dim,
                eps=self.head_dim*norm_eps,
                affine=elementwise_affine,
            )

        try:
            from transformers.modeling_utils import _init_weights
        except ImportError:
            _init_weights = True
        if _init_weights:
            self.apply(self._initialize_weights)
        for name, module in self.named_modules():
            module._in_rwkv_module = True

        warnings.warn(
            "According to Bo, you are using a potentially buggy FLA implementation of RWKV. "
            "If you plan to report any numbers based on this implementation, we strongly recommend "
            "cross-checking with the official repo: https://github.com/BlinkDL/RWKV-LM. "
            "Bo may disagree with results reported from this version.",
        )

    @torch.no_grad()
    @torch.compiler.disable
    def _initialize_weights(self, module: nn.Module):
        if getattr(module, "_is_hf_initialized", False):
            return

        # Initialize only when we're processing the RWKV7Attention module itself
        if isinstance(module, RWKV7Attention) and self.layer_idx is not None:
            ratio_0_to_1 = self.layer_idx / (self.num_hidden_layers - 1)  # 0 to 1
            ratio_1_to_almost0 = 1.0 - (self.layer_idx / self.num_hidden_layers)  # 1 to ~0

            # Create position-based initialization tensor
            ddd = torch.ones(1, 1, self.hidden_size, device=self.x_r.device)
            www = torch.zeros(self.hidden_size, device=self.x_r.device)
            zigzag = torch.zeros(self.hidden_size, device=self.x_r.device)
            linear = torch.zeros(self.hidden_size, device=self.x_r.device)
            for n in range(self.hidden_size):
                linear[n] = n / (self.hidden_size-1) - 0.5
                zigzag[n] = ((n % self.head_dim) - ((self.head_dim-1) / 2)) / ((self.head_dim-1) / 2)
                zigzag[n] = zigzag[n] * abs(zigzag[n])
                www[n] = -6 + 6 * (n / (self.hidden_size - 1)) ** (1 + 1 * ratio_0_to_1 ** 0.3)
                ddd[0, 0, n] = n / self.hidden_size

            # Initialize x_* parameters directly
            self.x_r.data = (1.0 - torch.pow(ddd, 0.2 * ratio_1_to_almost0)).to(self.x_r.dtype)
            self.x_w.data = (1.0 - torch.pow(ddd, 0.9 * ratio_1_to_almost0)).to(self.x_w.dtype)
            self.x_k.data = (1.0 - torch.pow(ddd, 0.7 * ratio_1_to_almost0)).to(self.x_k.dtype)
            self.x_v.data = (1.0 - torch.pow(ddd, 0.7 * ratio_1_to_almost0)).to(self.x_v.dtype)
            self.x_a.data = (1.0 - torch.pow(ddd, 0.9 * ratio_1_to_almost0)).to(self.x_a.dtype)
            self.x_g.data = (1.0 - torch.pow(ddd, 0.2 * ratio_1_to_almost0)).to(self.x_g.dtype)

            # Initialize k_k, k_a, r_k
            nn.init.constant_(self.k_a, 1.02)
            nn.init.constant_(self.r_k, -0.04)
            self.k_k.data.copy_((torch.zeros(self.hidden_size, device=self.k_k.device) +
                                0.71 - linear*0.1).to(self.k_k.dtype))
            # Set specific bias values for LoRA modules
            # 0.5 comes from F.softplus
            self.w_lora.set_bias_value(www + 0.5 + zigzag*2.5)
            self.a_lora.set_bias_value(-0.19 + zigzag*0.3 + linear*0.4)

            # v0 initialization - ones (for non-first layers)
            if self.layer_idx != 0:
                self.v_lora._initialize_weights(self.v_lora)
                self.v_lora.set_bias_value(0.73 - linear*0.4)

            # Initialize GroupNorm
            self.g_norm.weight.data[:] = ((self.layer_idx + 1) / self.num_hidden_layers) ** 0.7

            # Initialize Linear projections
            self._orthogonal_init(self.r_proj.weight)
            self._orthogonal_init(self.k_proj.weight, gain=0.1)
            self._orthogonal_init(self.v_proj.weight)
            self.o_proj.weight.data.zero_()

            # Clean up temporary tensors to free memory
            del ddd, www, zigzag, linear

        module._is_hf_initialized = True

    @staticmethod
    def _orthogonal_init(weight, gain=1.0):
        oringinal_dtype = weight.dtype
        weight = weight.float()
        nn.init.orthogonal_(weight, gain=gain)
        weight = weight.to(oringinal_dtype)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
        past_key_values: Cache | None = None,
        use_cache: bool | None = False,
        output_attentions: bool | None = False,
        v_first: torch.Tensor = None,
        cu_seqlens: torch.LongTensor | None = None,
        **kwargs,
    ) -> tuple[torch.Tensor, torch.Tensor | None, Cache | None]:
        batch_size, seq_len, _ = hidden_states.shape
        if attention_mask is not None:
            assert len(attention_mask.shape) == 2, (
                "Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
                "for padding purposes (0 indicating padding). "
                "Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
            )
            am = attention_mask.narrow(1, attention_mask.size(1) - seq_len, seq_len).unsqueeze(-1)

        last_state = None
        if past_key_values is not None and len(past_key_values) > self.layer_idx:
            last_state = past_key_values[self.layer_idx]

        if attention_mask is not None:
            hidden_states = hidden_states.mul(am)

        # delta [batch_size, seq_len, hidden_size]
        # conv_cache [N, D]
        if last_state is None:
            conv_cache = None
            recurrent_state = None
        else:
            conv_cache = last_state['conv_state']
            recurrent_state = last_state['recurrent_state']

        delta, conv_state = token_shift(
                hidden_states, cu_seqlens, output_cache=True, cache=conv_cache,
            )
        xr, xw, xk, xv, xa, xg = fused_addcmul_rwkv7(hidden_states, delta, self.x_r, self.x_w,
                                                     self.x_k, self.x_v, self.x_a, self.x_g)

        r = self.r_proj(xr)
        # Using bf16 for LoRA computation is numerically safe here because:
        # 1. After sigmoid activation:
        #    - Max absolute error (vs float32): 0.003
        #    - Mean absolute error: 0.0004
        # 2. Subsequent scaling by -0.6065 will further reduce relative error
        #    (error scales linearly with constant multiplication)
        # 3. Final compounded error remains within acceptable bounds for bf16 precision
        # Empirical observation confirms bf16 introduces no practical degradation
        w = -0.6065306597126334 * self.w_lora(xw).sigmoid()

        k = self.k_proj(xk)
        v = self.v_proj(xv)

        if self.layer_idx == 0:
            v_first = v
        else:
            v = torch.lerp(v, v_first, self.v_lora(xv).sigmoid())
        a = self.a_lora(xa).sigmoid()
        g = self.g_lora(xg)

        if self.fuse_norm:
            kk = l2_norm(rearrange(k * self.k_k, 'b t (h d) -> b t h d', d=self.head_dim))
        else:
            kk = F.normalize(rearrange(k * self.k_k, 'b t (h d) -> b t h d', d=self.head_dim), dim=-1, p=2.0)

        # Prefer addcmul over expanded form for numerical stability in bf16:
        # 1. Fused Multiply-Add (FMA) in addcmul reduces intermediate rounding:
        #    - Single op vs original 3 ops (mul, sub, mul)
        #    - 1 less intermediate value storage (bf16 write->read overhead)
        # 2. Mathematically equivalent to k*(1 + (a-1)*self.k_a)
        #    but with better precision preservation
        # 3. Particularly crucial for bf16 where intermediate values easily lose precision
        # 4. Pytorch method: k = k.addcmul(k * (a - 1), self.k_a)
        k = fused_k_rwkv7(k, a, self.k_a)

        # dealing with left-padding
        if attention_mask is not None:
            v = v * am

        r, w, k, a = map(lambda x: rearrange(x, 'b t (h d) -> b t h d', d=self.head_dim), (r, w, k, a))
        v = rearrange(v, 'b t (h d) -> b t h d', d=self.head_v_dim)

        if self.training or seq_len >= 64:
            # if training, use chunk mode no matter how short the sequence is
            # launching the triton kernel for just one token will actually be slower
            o, recurrent_state = chunk_rwkv7(
                r=r,
                w=w,
                k=k,
                v=v,
                a=-kk,
                b=kk * a,
                scale=1.,
                initial_state=recurrent_state,
                output_final_state=use_cache,
                cu_seqlens=cu_seqlens,
            )
        else:
            o, recurrent_state = fused_mul_recurrent_rwkv7(
                r=r,
                w=w,
                k=k,
                v=v,
                kk=kk,
                a=a,
                scale=1.,
                initial_state=recurrent_state,
                output_final_state=use_cache,
                cu_seqlens=cu_seqlens,
            )

        if past_key_values is not None:
            past_key_values.update(
                recurrent_state=recurrent_state,
                conv_state=conv_state,
                layer_idx=self.layer_idx,
                offset=r.shape[1],
            )

        if self.fuse_norm:
            o = self.g_norm(rearrange(o, '... h d -> ... (h d)'))
        else:
            o = self.g_norm(rearrange(o, 'b t h d -> (b t) (h d)')).view(batch_size, seq_len, -1)

        o = gate_output_correction(o, r, k, self.r_k, v, g)
        o = self.o_proj(o)

        return o, None, past_key_values, v_first