Upload wkv.py
Browse files
wkv.py
CHANGED
@@ -6,6 +6,8 @@ import math
|
|
6 |
import torch.nn as nn
|
7 |
from torch.nn import functional as F
|
8 |
from .configuration_rwkv_hybrid import RwkvHybridConfig
|
|
|
|
|
9 |
|
10 |
try:
|
11 |
import triton
|
@@ -13,6 +15,7 @@ try:
|
|
13 |
fused_recurrent_rwkv7,
|
14 |
chunk_rwkv7,
|
15 |
native_recurrent_rwkv7,
|
|
|
16 |
) # pylint: disable=C0411
|
17 |
from rwkvfla.ops.rwkv6 import (
|
18 |
fused_recurrent_rwkv6,
|
@@ -22,11 +25,13 @@ try:
|
|
22 |
except ImportError:
|
23 |
from rwkvfla.ops.rwkv7 import native_recurrent_rwkv7 # pylint: disable=C0411
|
24 |
from rwkvfla.ops.rwkv6 import native_recurrent_rwkv6
|
|
|
25 |
|
26 |
fused_recurrent_rwkv7 = native_recurrent_rwkv7
|
27 |
chunk_rwkv7 = native_recurrent_rwkv7
|
28 |
chunk_rwkv6 = native_recurrent_rwkv6
|
29 |
fused_recurrent_rwkv6 = native_recurrent_rwkv6
|
|
|
30 |
|
31 |
|
32 |
class Rwkv_Tmix_x070(nn.Module):
|
@@ -50,8 +55,7 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
50 |
self.x_k = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
51 |
self.x_v = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
52 |
self.x_a = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
53 |
-
|
54 |
-
|
55 |
D_DECAY_LORA = 64
|
56 |
D_AAA_LORA = 64
|
57 |
D_MV_LORA = 32
|
@@ -70,6 +74,7 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
70 |
self.v0 = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
71 |
|
72 |
if self.args.wkv_has_gate:
|
|
|
73 |
self.g1 = nn.Parameter(torch.Tensor(args.hidden_size, D_GATE_LORA))
|
74 |
self.g2 = nn.Parameter(torch.Tensor(D_GATE_LORA, args.hidden_size))
|
75 |
|
@@ -78,7 +83,8 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
78 |
self.r_k = nn.Parameter(torch.Tensor(H, N))
|
79 |
|
80 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
81 |
-
self.receptance = nn.Linear(
|
|
|
82 |
self.key = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
83 |
self.value = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
84 |
self.output = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
@@ -90,7 +96,8 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
90 |
|
91 |
def post_init(self):
|
92 |
with torch.no_grad():
|
93 |
-
ratio_0_to_1 = self.layer_id /
|
|
|
94 |
ratio_1_to_almost0 = 1.0 - (
|
95 |
self.layer_id / self.args.num_hidden_layers
|
96 |
) # 1 to ~0
|
@@ -99,39 +106,48 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
99 |
for i in range(self.args.hidden_size):
|
100 |
ddd[0, 0, i] = i / self.args.hidden_size
|
101 |
|
102 |
-
nn.init.constant_(
|
103 |
-
|
|
|
|
|
104 |
nn.init.constant_(
|
105 |
self.x_k,
|
106 |
-
1.0 - (torch.pow(ddd, 0.9 * ratio_1_to_almost0) +
|
|
|
107 |
)
|
108 |
nn.init.constant_(
|
109 |
self.x_v,
|
110 |
-
1.0 - (torch.pow(ddd, 0.4 * ratio_1_to_almost0) +
|
|
|
111 |
)
|
112 |
-
nn.init.constant_(
|
113 |
-
|
|
|
114 |
|
115 |
def ortho_init(x, scale):
|
116 |
shape = x.shape
|
117 |
original_dtype = x.dtype
|
118 |
x_fp32 = x.float()
|
119 |
if len(shape) == 2:
|
120 |
-
gain = math.sqrt(shape[0] / shape[1]
|
|
|
121 |
nn.init.orthogonal_(x_fp32, gain=gain * scale)
|
122 |
elif len(shape) == 3:
|
123 |
-
gain = math.sqrt(shape[1] / shape[2]
|
|
|
124 |
for i in range(shape[0]):
|
125 |
nn.init.orthogonal_(x_fp32[i], gain=gain * scale)
|
126 |
else:
|
127 |
-
raise ValueError(
|
|
|
128 |
x.data.copy_(x_fp32.to(original_dtype))
|
129 |
return x
|
130 |
|
131 |
D_DECAY_LORA = 64
|
132 |
nn.init.zeros_(self.w1)
|
133 |
self.w2 = nn.Parameter(
|
134 |
-
ortho_init(torch.zeros(
|
|
|
135 |
)
|
136 |
|
137 |
decay_speed = torch.ones(self.args.hidden_size)
|
@@ -161,8 +177,11 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
161 |
if self.args.wkv_has_gate:
|
162 |
nn.init.zeros_(self.g1)
|
163 |
self.g2 = nn.Parameter(
|
164 |
-
ortho_init(torch.zeros(
|
|
|
165 |
)
|
|
|
|
|
166 |
|
167 |
nn.init.constant_(self.k_k, 0.85)
|
168 |
nn.init.constant_(self.k_a, 1.0)
|
@@ -177,77 +196,68 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
177 |
nn.init.ones_(self.ln_x.weight)
|
178 |
nn.init.zeros_(self.ln_x.bias)
|
179 |
|
180 |
-
def apply_wkv7_state(self, r, k, v, w, a, b, s
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
a = rearrange(a, "b l (h d) -> b h l d", h=self.n_head)
|
186 |
-
b = rearrange(b, "b l (h d) -> b h l d", h=self.n_head)
|
187 |
|
188 |
if r.device.type == "cpu":
|
|
|
189 |
o, state = native_recurrent_rwkv7(
|
190 |
-
r,
|
191 |
-
|
192 |
-
v,
|
193 |
-
w,
|
194 |
-
a,
|
195 |
-
b,
|
196 |
scale=1.0,
|
197 |
initial_state=s.transpose(-1, -2),
|
198 |
output_final_state=True,
|
199 |
-
use_log_w=False,
|
200 |
head_first=True,
|
201 |
)
|
202 |
state = state.transpose(-1, -2)
|
203 |
-
|
204 |
-
o, state = chunk_rwkv7(
|
205 |
-
r,
|
206 |
-
k,
|
207 |
-
v,
|
208 |
-
w,
|
209 |
-
a,
|
210 |
-
b,
|
211 |
-
scale=1.0,
|
212 |
-
initial_state=s,
|
213 |
-
output_final_state=True,
|
214 |
-
use_log_w=False,
|
215 |
-
head_first=True,
|
216 |
-
)
|
217 |
else:
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
v,
|
222 |
-
|
223 |
-
a,
|
224 |
-
b,
|
225 |
scale=1.0,
|
226 |
initial_state=s,
|
227 |
-
output_final_state=
|
228 |
-
|
229 |
-
head_first=
|
230 |
)
|
231 |
-
|
232 |
-
x = rearrange(o, "b h l d -> b l (h d)")
|
233 |
return x, state
|
234 |
|
235 |
-
def forward(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
236 |
shift_state = last_state.shift_state
|
237 |
-
B, T, C =
|
238 |
-
|
239 |
if shift_state is not None:
|
240 |
-
xx = torch.concat((shift_state.unsqueeze(
|
|
|
241 |
else:
|
242 |
-
xx = self.time_shift(
|
243 |
-
lx = x[:, -1]
|
244 |
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
|
|
251 |
|
252 |
r = self.receptance(xr)
|
253 |
w = (
|
@@ -269,11 +279,11 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
269 |
if self.args.wkv_has_gate:
|
270 |
g = torch.sigmoid(xg @ self.g1) @ self.g2
|
271 |
kk = k * self.k_k
|
272 |
-
kk = F.normalize(kk.view(B, T,
|
273 |
k = k * (1 + (a - 1) * self.k_a)
|
274 |
|
275 |
wkv_state = last_state.wkv_state
|
276 |
-
|
277 |
r,
|
278 |
k,
|
279 |
v,
|
@@ -281,17 +291,22 @@ class Rwkv_Tmix_x070(nn.Module):
|
|
281 |
-kk,
|
282 |
(kk * a),
|
283 |
s=wkv_state,
|
|
|
|
|
|
|
284 |
)
|
285 |
if self.args.wkv_has_group_norm:
|
286 |
-
|
287 |
-
|
288 |
-
|
|
|
289 |
dim=-1, keepdim=True
|
290 |
)
|
291 |
-
* v.view(B, T,
|
292 |
).view(B, T, C)
|
293 |
-
|
294 |
-
|
|
|
295 |
|
296 |
|
297 |
class Rwkv7Attention(nn.Module):
|
@@ -299,24 +314,43 @@ class Rwkv7Attention(nn.Module):
|
|
299 |
super().__init__()
|
300 |
self.args = args
|
301 |
self.layer_idx = layer_id
|
302 |
-
self.time_mixer = Rwkv_Tmix_x070(
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
307 |
|
308 |
if past_key_value is not None and len(past_key_value) > self.layer_idx:
|
309 |
last_state = past_key_value[self.layer_idx][0]
|
310 |
else:
|
311 |
last_state = self.init_state(
|
312 |
-
batch_size,
|
313 |
)
|
314 |
|
315 |
-
attn_output, states = self.time_mixer(
|
|
|
|
|
|
|
|
|
316 |
last_state.time_mix_state = states
|
317 |
|
318 |
if past_key_value is not None:
|
319 |
past_key_value.update(token_length, last_state, self.layer_idx)
|
|
|
320 |
return attn_output, None
|
321 |
|
322 |
def init_state(self, batch_size, device, dtype) -> BlockState:
|
@@ -357,9 +391,12 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
357 |
ddd[0, 0, i] = i / args.hidden_size
|
358 |
|
359 |
# fancy time_mix
|
360 |
-
self.time_maa_x = nn.Parameter(
|
361 |
-
|
362 |
-
self.
|
|
|
|
|
|
|
363 |
self.time_maa_v = nn.Parameter(
|
364 |
1.0 - (torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
|
365 |
)
|
@@ -377,7 +414,8 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
377 |
torch.zeros(args.hidden_size, D_MIX_LORA * 5)
|
378 |
)
|
379 |
self.time_maa_w2 = nn.Parameter(
|
380 |
-
torch.zeros(5, D_MIX_LORA,
|
|
|
381 |
)
|
382 |
|
383 |
# fancy time_decay
|
@@ -386,7 +424,8 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
386 |
decay_speed[n] = -6 + 5 * (n / (args.head_size - 1)) ** (
|
387 |
0.7 + 1.3 * ratio_0_to_1
|
388 |
)
|
389 |
-
self.time_decay = nn.Parameter(
|
|
|
390 |
|
391 |
D_DECAY_LORA = 64
|
392 |
if args.hidden_size == 4096:
|
@@ -401,13 +440,16 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
401 |
tmp = torch.zeros(args.head_size)
|
402 |
for n in range(args.head_size):
|
403 |
zigzag = ((n + 1) % 3 - 1) * 0.1
|
404 |
-
tmp[n] = ratio_0_to_1 *
|
|
|
405 |
|
406 |
-
self.time_faaaa = nn.Parameter(
|
|
|
407 |
# self.time_state = nn.Parameter(torch.zeros(self.n_head, self.head_size, self.head_size))
|
408 |
|
409 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
410 |
-
self.receptance = nn.Linear(
|
|
|
411 |
self.key = nn.Linear(args.hidden_size, args.head_size, bias=False)
|
412 |
|
413 |
self.value = nn.Linear(args.hidden_size, args.head_size, bias=False)
|
@@ -416,7 +458,8 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
416 |
|
417 |
if self.args.wkv_has_group_norm:
|
418 |
self.ln_x = nn.GroupNorm(
|
419 |
-
self.n_head, args.head_size, eps=(
|
|
|
420 |
)
|
421 |
|
422 |
def post_init(self):
|
@@ -433,7 +476,8 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
433 |
lx = x[:, -1]
|
434 |
|
435 |
xxx = x + xx * self.time_maa_x
|
436 |
-
xxx = torch.tanh(xxx @ self.time_maa_w1).view(B *
|
|
|
437 |
xxx = torch.bmm(xxx, self.time_maa_w2).view(5, B, T, -1)
|
438 |
mw, mk, mv, mr, mg = xxx.unbind(dim=0)
|
439 |
|
@@ -461,10 +505,7 @@ class Rwkv_Tmix_x060(nn.Module):
|
|
461 |
return x, TimeMixState(lx, wkv_state)
|
462 |
|
463 |
def apply_wkv6_state(self, B, T, C, H, r, k, v, w, u, s):
|
464 |
-
r = rearrange(
|
465 |
-
k = rearrange(k, "b l (h d) -> b h l d", h=H)
|
466 |
-
v = rearrange(v, "b l (h d) -> b h l d", h=H)
|
467 |
-
w = rearrange(w, "b l (h d) -> b h l d", h=H)
|
468 |
|
469 |
if r.device.type == "cpu":
|
470 |
wkv6_func = native_recurrent_rwkv6
|
@@ -504,7 +545,8 @@ class Rwkv6Attention(nn.Module):
|
|
504 |
last_state = past_key_value[self.layer_idx][0]
|
505 |
if last_state is None:
|
506 |
wkv_states = torch.zeros(
|
507 |
-
(B, self.args.num_wkv_heads,
|
|
|
508 |
device=attn_output.device,
|
509 |
dtype=torch.float32,
|
510 |
)
|
@@ -514,7 +556,8 @@ class Rwkv6Attention(nn.Module):
|
|
514 |
time_state = TimeMixState(token_shift, wkv_states)
|
515 |
channel_state = None
|
516 |
last_state = BlockState(time_state, channel_state)
|
517 |
-
attn_output, states = self.time_mixer(
|
|
|
518 |
last_state.time_mix_state = states
|
519 |
|
520 |
if past_key_value is not None:
|
|
|
6 |
import torch.nn as nn
|
7 |
from torch.nn import functional as F
|
8 |
from .configuration_rwkv_hybrid import RwkvHybridConfig
|
9 |
+
from typing import TYPE_CHECKING, Optional
|
10 |
+
from transformers.cache_utils import Cache
|
11 |
|
12 |
try:
|
13 |
import triton
|
|
|
15 |
fused_recurrent_rwkv7,
|
16 |
chunk_rwkv7,
|
17 |
native_recurrent_rwkv7,
|
18 |
+
fused_addcmul_rwkv7,
|
19 |
) # pylint: disable=C0411
|
20 |
from rwkvfla.ops.rwkv6 import (
|
21 |
fused_recurrent_rwkv6,
|
|
|
25 |
except ImportError:
|
26 |
from rwkvfla.ops.rwkv7 import native_recurrent_rwkv7 # pylint: disable=C0411
|
27 |
from rwkvfla.ops.rwkv6 import native_recurrent_rwkv6
|
28 |
+
from rwkvfla.ops.rwkv7 import torch_addcmul_rwkv7
|
29 |
|
30 |
fused_recurrent_rwkv7 = native_recurrent_rwkv7
|
31 |
chunk_rwkv7 = native_recurrent_rwkv7
|
32 |
chunk_rwkv6 = native_recurrent_rwkv6
|
33 |
fused_recurrent_rwkv6 = native_recurrent_rwkv6
|
34 |
+
fused_addcmul_rwkv7 = torch_addcmul_rwkv7
|
35 |
|
36 |
|
37 |
class Rwkv_Tmix_x070(nn.Module):
|
|
|
55 |
self.x_k = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
56 |
self.x_v = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
57 |
self.x_a = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
58 |
+
|
|
|
59 |
D_DECAY_LORA = 64
|
60 |
D_AAA_LORA = 64
|
61 |
D_MV_LORA = 32
|
|
|
74 |
self.v0 = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
75 |
|
76 |
if self.args.wkv_has_gate:
|
77 |
+
self.x_g = nn.Parameter(torch.Tensor(1, 1, args.hidden_size))
|
78 |
self.g1 = nn.Parameter(torch.Tensor(args.hidden_size, D_GATE_LORA))
|
79 |
self.g2 = nn.Parameter(torch.Tensor(D_GATE_LORA, args.hidden_size))
|
80 |
|
|
|
83 |
self.r_k = nn.Parameter(torch.Tensor(H, N))
|
84 |
|
85 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
86 |
+
self.receptance = nn.Linear(
|
87 |
+
args.hidden_size, args.hidden_size, bias=False)
|
88 |
self.key = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
89 |
self.value = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
90 |
self.output = nn.Linear(args.hidden_size, args.hidden_size, bias=False)
|
|
|
96 |
|
97 |
def post_init(self):
|
98 |
with torch.no_grad():
|
99 |
+
ratio_0_to_1 = self.layer_id / \
|
100 |
+
(self.args.num_hidden_layers - 1) # 0 to 1
|
101 |
ratio_1_to_almost0 = 1.0 - (
|
102 |
self.layer_id / self.args.num_hidden_layers
|
103 |
) # 1 to ~0
|
|
|
106 |
for i in range(self.args.hidden_size):
|
107 |
ddd[0, 0, i] = i / self.args.hidden_size
|
108 |
|
109 |
+
nn.init.constant_(
|
110 |
+
self.x_r, 1.0 - torch.pow(ddd, 0.2 * ratio_1_to_almost0))
|
111 |
+
nn.init.constant_(
|
112 |
+
self.x_w, 1.0 - torch.pow(ddd, 0.9 * ratio_1_to_almost0))
|
113 |
nn.init.constant_(
|
114 |
self.x_k,
|
115 |
+
1.0 - (torch.pow(ddd, 0.9 * ratio_1_to_almost0) +
|
116 |
+
0.4 * ratio_0_to_1),
|
117 |
)
|
118 |
nn.init.constant_(
|
119 |
self.x_v,
|
120 |
+
1.0 - (torch.pow(ddd, 0.4 * ratio_1_to_almost0) +
|
121 |
+
0.6 * ratio_0_to_1),
|
122 |
)
|
123 |
+
nn.init.constant_(
|
124 |
+
self.x_a, 1.0 - torch.pow(ddd, 0.9 * ratio_1_to_almost0))
|
125 |
+
|
126 |
|
127 |
def ortho_init(x, scale):
|
128 |
shape = x.shape
|
129 |
original_dtype = x.dtype
|
130 |
x_fp32 = x.float()
|
131 |
if len(shape) == 2:
|
132 |
+
gain = math.sqrt(shape[0] / shape[1]
|
133 |
+
) if shape[0] > shape[1] else 1
|
134 |
nn.init.orthogonal_(x_fp32, gain=gain * scale)
|
135 |
elif len(shape) == 3:
|
136 |
+
gain = math.sqrt(shape[1] / shape[2]
|
137 |
+
) if shape[1] > shape[2] else 1
|
138 |
for i in range(shape[0]):
|
139 |
nn.init.orthogonal_(x_fp32[i], gain=gain * scale)
|
140 |
else:
|
141 |
+
raise ValueError(
|
142 |
+
"ortho_init only supports 2D or 3D tensors")
|
143 |
x.data.copy_(x_fp32.to(original_dtype))
|
144 |
return x
|
145 |
|
146 |
D_DECAY_LORA = 64
|
147 |
nn.init.zeros_(self.w1)
|
148 |
self.w2 = nn.Parameter(
|
149 |
+
ortho_init(torch.zeros(
|
150 |
+
D_DECAY_LORA, self.args.hidden_size), 0.1)
|
151 |
)
|
152 |
|
153 |
decay_speed = torch.ones(self.args.hidden_size)
|
|
|
177 |
if self.args.wkv_has_gate:
|
178 |
nn.init.zeros_(self.g1)
|
179 |
self.g2 = nn.Parameter(
|
180 |
+
ortho_init(torch.zeros(
|
181 |
+
D_GATE_LORA, self.args.hidden_size), 0.1)
|
182 |
)
|
183 |
+
nn.init.constant_(
|
184 |
+
self.x_g, 1.0 - torch.pow(ddd, 0.2 * ratio_1_to_almost0))
|
185 |
|
186 |
nn.init.constant_(self.k_k, 0.85)
|
187 |
nn.init.constant_(self.k_a, 1.0)
|
|
|
196 |
nn.init.ones_(self.ln_x.weight)
|
197 |
nn.init.zeros_(self.ln_x.bias)
|
198 |
|
199 |
+
def apply_wkv7_state(self, r, k, v, w, a, b, s,
|
200 |
+
output_final_state,
|
201 |
+
cu_seqlens,
|
202 |
+
head_first
|
203 |
+
):
|
|
|
|
|
204 |
|
205 |
if r.device.type == "cpu":
|
206 |
+
r, w, k, v, a, b = map(lambda x: rearrange(x, 'b l (h d) -> b h l d', h=self.n_head), (r, w, k, v, a, b))
|
207 |
o, state = native_recurrent_rwkv7(
|
208 |
+
r=r, k=k, v=v, w=w,
|
209 |
+
a=a, b=b,
|
|
|
|
|
|
|
|
|
210 |
scale=1.0,
|
211 |
initial_state=s.transpose(-1, -2),
|
212 |
output_final_state=True,
|
|
|
213 |
head_first=True,
|
214 |
)
|
215 |
state = state.transpose(-1, -2)
|
216 |
+
x = rearrange(o, "b h l d -> b l (h d)")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
217 |
else:
|
218 |
+
r, w, k, v, a, b = map(lambda x: rearrange(x, 'b l (h d) -> b l h d', h=self.n_head), (r, w, k, v, a, b))
|
219 |
+
wkv7_func = chunk_rwkv7 if self.training else fused_recurrent_rwkv7
|
220 |
+
o, state = wkv7_func(
|
221 |
+
r=r, k=k, v=v, w=w,
|
222 |
+
a=a, b=b,
|
|
|
|
|
223 |
scale=1.0,
|
224 |
initial_state=s,
|
225 |
+
output_final_state=output_final_state,
|
226 |
+
cu_seqlens=cu_seqlens,
|
227 |
+
head_first=head_first,
|
228 |
)
|
229 |
+
x = rearrange(o, "b l h d -> b l (h d)")
|
|
|
230 |
return x, state
|
231 |
|
232 |
+
def forward(
|
233 |
+
self,
|
234 |
+
hidden_states,
|
235 |
+
last_state: TimeMixState,
|
236 |
+
sequence_mask: Optional[torch.Tensor] = None,
|
237 |
+
use_cache: Optional[bool] = False,
|
238 |
+
cu_seqlens: Optional[torch.Tensor] = None,
|
239 |
+
**kwargs
|
240 |
+
):
|
241 |
+
if sequence_mask is not None:
|
242 |
+
hidden_states = hidden_states.mul(
|
243 |
+
sequence_mask[:, -hidden_states.shape[-2]:, None])
|
244 |
+
|
245 |
shift_state = last_state.shift_state
|
246 |
+
B, T, C = hidden_states.size()
|
247 |
+
|
248 |
if shift_state is not None:
|
249 |
+
xx = torch.concat((shift_state.unsqueeze(
|
250 |
+
1), hidden_states[:, :-1]), dim=1) - hidden_states
|
251 |
else:
|
252 |
+
xx = self.time_shift(hidden_states) - hidden_states
|
|
|
253 |
|
254 |
+
lx = hidden_states[:, -1]
|
255 |
+
|
256 |
+
if self.args.wkv_has_gate:
|
257 |
+
xr, xw, xk, xv, xa, xg = fused_addcmul_rwkv7(
|
258 |
+
hidden_states, xx, self.x_r, self.x_w, self.x_k, self.x_v, self.x_a, self.x_g)
|
259 |
+
else:
|
260 |
+
xr, xw, xk, xv, xa, _ = fused_addcmul_rwkv7(hidden_states, xx, self.x_r, self.x_w, self.x_k, self.x_v, self.x_a)
|
261 |
|
262 |
r = self.receptance(xr)
|
263 |
w = (
|
|
|
279 |
if self.args.wkv_has_gate:
|
280 |
g = torch.sigmoid(xg @ self.g1) @ self.g2
|
281 |
kk = k * self.k_k
|
282 |
+
kk = F.normalize(kk.view(B, T, self.n_head, -1), dim=-1, p=2.0).view(B, T, C)
|
283 |
k = k * (1 + (a - 1) * self.k_a)
|
284 |
|
285 |
wkv_state = last_state.wkv_state
|
286 |
+
hidden_states, wkv_state = self.apply_wkv7_state(
|
287 |
r,
|
288 |
k,
|
289 |
v,
|
|
|
291 |
-kk,
|
292 |
(kk * a),
|
293 |
s=wkv_state,
|
294 |
+
output_final_state=use_cache,
|
295 |
+
cu_seqlens=cu_seqlens,
|
296 |
+
head_first=False
|
297 |
)
|
298 |
if self.args.wkv_has_group_norm:
|
299 |
+
hidden_states = self.ln_x(
|
300 |
+
hidden_states.view(B * T, C)).view(B, T, C)
|
301 |
+
hidden_states = hidden_states + (
|
302 |
+
(r.view(B, T, self.n_head, -1) * k.view(B, T, self.n_head, -1) * self.r_k).sum(
|
303 |
dim=-1, keepdim=True
|
304 |
)
|
305 |
+
* v.view(B, T, self.n_head, -1)
|
306 |
).view(B, T, C)
|
307 |
+
hidden_states = self.output(
|
308 |
+
hidden_states * g) if self.args.wkv_has_gate else self.output(hidden_states)
|
309 |
+
return hidden_states, TimeMixState(lx, wkv_state)
|
310 |
|
311 |
|
312 |
class Rwkv7Attention(nn.Module):
|
|
|
314 |
super().__init__()
|
315 |
self.args = args
|
316 |
self.layer_idx = layer_id
|
317 |
+
self.time_mixer = Rwkv_Tmix_x070(
|
318 |
+
args, layer_id, update_v_first, get_v_first)
|
319 |
+
|
320 |
+
def forward(
|
321 |
+
self,
|
322 |
+
hidden_states: torch.Tensor,
|
323 |
+
sequence_mask: Optional[torch.Tensor] = None,
|
324 |
+
past_key_value: Optional[Cache] = None,
|
325 |
+
use_cache: Optional[bool] = False,
|
326 |
+
output_attentions: Optional[bool] = False,
|
327 |
+
**kwargs
|
328 |
+
):
|
329 |
+
if sequence_mask is not None:
|
330 |
+
assert len(sequence_mask.shape) == 2, (
|
331 |
+
"Expected attention_mask as a 0-1 matrix with shape [batch_size, seq_len] "
|
332 |
+
"for padding purposes (0 indicating padding). "
|
333 |
+
"Arbitrary attention masks of shape [batch_size, seq_len, seq_len] are not allowed."
|
334 |
+
)
|
335 |
+
batch_size, token_length, _ = hidden_states.shape
|
336 |
|
337 |
if past_key_value is not None and len(past_key_value) > self.layer_idx:
|
338 |
last_state = past_key_value[self.layer_idx][0]
|
339 |
else:
|
340 |
last_state = self.init_state(
|
341 |
+
batch_size, hidden_states.device, hidden_states.dtype
|
342 |
)
|
343 |
|
344 |
+
attn_output, states = self.time_mixer(hidden_states=hidden_states,
|
345 |
+
last_state=last_state.time_mix_state,
|
346 |
+
sequence_mask=sequence_mask,
|
347 |
+
use_cache=use_cache,
|
348 |
+
**kwargs)
|
349 |
last_state.time_mix_state = states
|
350 |
|
351 |
if past_key_value is not None:
|
352 |
past_key_value.update(token_length, last_state, self.layer_idx)
|
353 |
+
|
354 |
return attn_output, None
|
355 |
|
356 |
def init_state(self, batch_size, device, dtype) -> BlockState:
|
|
|
391 |
ddd[0, 0, i] = i / args.hidden_size
|
392 |
|
393 |
# fancy time_mix
|
394 |
+
self.time_maa_x = nn.Parameter(
|
395 |
+
1.0 - torch.pow(ddd, ratio_1_to_almost0))
|
396 |
+
self.time_maa_w = nn.Parameter(
|
397 |
+
1.0 - torch.pow(ddd, ratio_1_to_almost0))
|
398 |
+
self.time_maa_k = nn.Parameter(
|
399 |
+
1.0 - torch.pow(ddd, ratio_1_to_almost0))
|
400 |
self.time_maa_v = nn.Parameter(
|
401 |
1.0 - (torch.pow(ddd, ratio_1_to_almost0) + 0.3 * ratio_0_to_1)
|
402 |
)
|
|
|
414 |
torch.zeros(args.hidden_size, D_MIX_LORA * 5)
|
415 |
)
|
416 |
self.time_maa_w2 = nn.Parameter(
|
417 |
+
torch.zeros(5, D_MIX_LORA,
|
418 |
+
args.hidden_size).uniform_(-0.01, 0.01)
|
419 |
)
|
420 |
|
421 |
# fancy time_decay
|
|
|
424 |
decay_speed[n] = -6 + 5 * (n / (args.head_size - 1)) ** (
|
425 |
0.7 + 1.3 * ratio_0_to_1
|
426 |
)
|
427 |
+
self.time_decay = nn.Parameter(
|
428 |
+
decay_speed.reshape(1, 1, args.head_size))
|
429 |
|
430 |
D_DECAY_LORA = 64
|
431 |
if args.hidden_size == 4096:
|
|
|
440 |
tmp = torch.zeros(args.head_size)
|
441 |
for n in range(args.head_size):
|
442 |
zigzag = ((n + 1) % 3 - 1) * 0.1
|
443 |
+
tmp[n] = ratio_0_to_1 * \
|
444 |
+
(1 - (n / (args.head_size - 1))) + zigzag
|
445 |
|
446 |
+
self.time_faaaa = nn.Parameter(
|
447 |
+
tmp.reshape(self.n_head, self.head_size))
|
448 |
# self.time_state = nn.Parameter(torch.zeros(self.n_head, self.head_size, self.head_size))
|
449 |
|
450 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
451 |
+
self.receptance = nn.Linear(
|
452 |
+
args.hidden_size, args.head_size, bias=False)
|
453 |
self.key = nn.Linear(args.hidden_size, args.head_size, bias=False)
|
454 |
|
455 |
self.value = nn.Linear(args.hidden_size, args.head_size, bias=False)
|
|
|
458 |
|
459 |
if self.args.wkv_has_group_norm:
|
460 |
self.ln_x = nn.GroupNorm(
|
461 |
+
self.n_head, args.head_size, eps=(
|
462 |
+
1e-5) * (args.head_size_divisor**2)
|
463 |
)
|
464 |
|
465 |
def post_init(self):
|
|
|
476 |
lx = x[:, -1]
|
477 |
|
478 |
xxx = x + xx * self.time_maa_x
|
479 |
+
xxx = torch.tanh(xxx @ self.time_maa_w1).view(B *
|
480 |
+
T, 5, -1).transpose(0, 1)
|
481 |
xxx = torch.bmm(xxx, self.time_maa_w2).view(5, B, T, -1)
|
482 |
mw, mk, mv, mr, mg = xxx.unbind(dim=0)
|
483 |
|
|
|
505 |
return x, TimeMixState(lx, wkv_state)
|
506 |
|
507 |
def apply_wkv6_state(self, B, T, C, H, r, k, v, w, u, s):
|
508 |
+
r, w, k, v = map(lambda x: rearrange(x, 'b l (h d) -> b h l d', h=self.n_head), (r, w, k, v))
|
|
|
|
|
|
|
509 |
|
510 |
if r.device.type == "cpu":
|
511 |
wkv6_func = native_recurrent_rwkv6
|
|
|
545 |
last_state = past_key_value[self.layer_idx][0]
|
546 |
if last_state is None:
|
547 |
wkv_states = torch.zeros(
|
548 |
+
(B, self.args.num_wkv_heads,
|
549 |
+
self.args.head_size, self.args.head_size),
|
550 |
device=attn_output.device,
|
551 |
dtype=torch.float32,
|
552 |
)
|
|
|
556 |
time_state = TimeMixState(token_shift, wkv_states)
|
557 |
channel_state = None
|
558 |
last_state = BlockState(time_state, channel_state)
|
559 |
+
attn_output, states = self.time_mixer(
|
560 |
+
attn_output, last_state.time_mix_state)
|
561 |
last_state.time_mix_state = states
|
562 |
|
563 |
if past_key_value is not None:
|