KaleiNeely
commited on
Commit
•
2027efa
1
Parent(s):
7a56a0e
Upload 2 files
Browse files- configuration_rwkv5.py +4 -6
- modeling_rwkv5.py +140 -189
configuration_rwkv5.py
CHANGED
@@ -53,11 +53,9 @@ class Rwkv5Config(PretrainedConfig):
|
|
53 |
layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
|
54 |
The epsilon to use in the layer normalization layers.
|
55 |
bos_token_id (`int`, *optional*, defaults to 0):
|
56 |
-
The id of the beginning of sentence token in the vocabulary. Defaults to 0
|
57 |
-
as GPTNeoX.
|
58 |
eos_token_id (`int`, *optional*, defaults to 0):
|
59 |
-
The id of the end of sentence token in the vocabulary. Defaults to 0
|
60 |
-
GPTNeoX.
|
61 |
rescale_every (`int`, *optional*, defaults to 6):
|
62 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
63 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
@@ -90,8 +88,8 @@ class Rwkv5Config(PretrainedConfig):
|
|
90 |
hidden_size=768,
|
91 |
num_hidden_layers=24,
|
92 |
attention_hidden_size=None,
|
93 |
-
num_attention_heads=64,
|
94 |
head_size=64,
|
|
|
95 |
intermediate_size=None,
|
96 |
layer_norm_epsilon=1e-5,
|
97 |
bos_token_id=0,
|
@@ -105,8 +103,8 @@ class Rwkv5Config(PretrainedConfig):
|
|
105 |
self.hidden_size = hidden_size
|
106 |
self.num_hidden_layers = num_hidden_layers
|
107 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
108 |
-
self.num_attention_heads = num_attention_heads
|
109 |
self.head_size = head_size
|
|
|
110 |
self.intermediate_size = None
|
111 |
self.layer_norm_epsilon = layer_norm_epsilon
|
112 |
self.rescale_every = rescale_every
|
|
|
53 |
layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
|
54 |
The epsilon to use in the layer normalization layers.
|
55 |
bos_token_id (`int`, *optional*, defaults to 0):
|
56 |
+
The id of the beginning of sentence token in the vocabulary. Defaults to 0.
|
|
|
57 |
eos_token_id (`int`, *optional*, defaults to 0):
|
58 |
+
The id of the end of sentence token in the vocabulary. Defaults to 0.
|
|
|
59 |
rescale_every (`int`, *optional*, defaults to 6):
|
60 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
61 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
|
|
88 |
hidden_size=768,
|
89 |
num_hidden_layers=24,
|
90 |
attention_hidden_size=None,
|
|
|
91 |
head_size=64,
|
92 |
+
head_size_divisor=8,
|
93 |
intermediate_size=None,
|
94 |
layer_norm_epsilon=1e-5,
|
95 |
bos_token_id=0,
|
|
|
103 |
self.hidden_size = hidden_size
|
104 |
self.num_hidden_layers = num_hidden_layers
|
105 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
|
|
106 |
self.head_size = head_size
|
107 |
+
self.head_size_divisor = head_size_divisor
|
108 |
self.intermediate_size = None
|
109 |
self.layer_norm_epsilon = layer_norm_epsilon
|
110 |
self.rescale_every = rescale_every
|
modeling_rwkv5.py
CHANGED
@@ -1,6 +1,5 @@
|
|
1 |
# coding=utf-8
|
2 |
-
# Copyright
|
3 |
-
# Copyright (c) 2018, NVIDIA CORPORATION. All rights reserved.
|
4 |
#
|
5 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
6 |
# you may not use this file except in compliance with the License.
|
@@ -45,12 +44,6 @@ logger = logging.get_logger(__name__)
|
|
45 |
_CHECKPOINT_FOR_DOC = "RWKV/rwkv-5-world-1b5"
|
46 |
_CONFIG_FOR_DOC = "Rwkv5Config"
|
47 |
|
48 |
-
RWKV5_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
49 |
-
"RWKV/rwkv-5-world-1b5",
|
50 |
-
"RWKV/rwkv-5-world-3b",
|
51 |
-
# See all RWKV models at https://huggingface.co/models?filter=rwkv
|
52 |
-
]
|
53 |
-
|
54 |
rwkv5_cuda_kernel = None
|
55 |
|
56 |
|
@@ -60,14 +53,14 @@ def load_wkv5_cuda_kernel(head_size):
|
|
60 |
|
61 |
global rwkv5_cuda_kernel
|
62 |
|
63 |
-
kernel_folder = Path(__file__).resolve()
|
64 |
cuda_kernel_files = [kernel_folder / f for f in ["wkv5_op.cpp", "wkv5_cuda.cu"]]
|
65 |
|
66 |
# Only load the kernel if it's not been loaded yet or if we changed the context length
|
67 |
if rwkv5_cuda_kernel is not None and rwkv5_cuda_kernel.head_size == head_size:
|
68 |
return
|
69 |
|
70 |
-
logger.info(f"Loading CUDA kernel for
|
71 |
|
72 |
flags = [
|
73 |
"-res-usage",
|
@@ -87,39 +80,45 @@ def load_wkv5_cuda_kernel(head_size):
|
|
87 |
rwkv5_cuda_kernel.head_size = head_size
|
88 |
|
89 |
|
90 |
-
class
|
91 |
@staticmethod
|
92 |
def forward(ctx, receptance, key, value, time_decay, time_first, state):
|
93 |
with torch.no_grad():
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
|
|
|
|
|
|
|
|
102 |
e_time_decay = (-torch.exp(time_decay.float())).contiguous()
|
103 |
ee_time_decay = (torch.exp(e_time_decay)).contiguous()
|
|
|
104 |
ctx.save_for_backward(receptance, key, value, ee_time_decay, e_time_decay, time_first)
|
105 |
out = torch.empty(
|
106 |
-
(
|
107 |
device=receptance.device,
|
108 |
dtype=torch.bfloat16,
|
109 |
memory_format=torch.contiguous_format,
|
110 |
)
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
|
|
|
|
116 |
receptance,
|
117 |
key,
|
118 |
value,
|
119 |
ee_time_decay,
|
120 |
time_first,
|
121 |
out,
|
122 |
-
state,
|
123 |
)
|
124 |
return out, state
|
125 |
|
@@ -127,51 +126,55 @@ class WKV_5(torch.autograd.Function):
|
|
127 |
def backward(ctx, gout):
|
128 |
with torch.no_grad():
|
129 |
assert gout.dtype == torch.bfloat16
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
receptance, key, value, ee_time_decay, e_time_decay, time_first = ctx.saved_tensors
|
|
|
|
|
|
|
|
|
135 |
greceptance = torch.empty(
|
136 |
-
|
137 |
device=gout.device,
|
138 |
requires_grad=False,
|
139 |
dtype=torch.bfloat16,
|
140 |
memory_format=torch.contiguous_format,
|
141 |
)
|
142 |
g_key = torch.empty(
|
143 |
-
|
144 |
device=gout.device,
|
145 |
requires_grad=False,
|
146 |
dtype=torch.bfloat16,
|
147 |
memory_format=torch.contiguous_format,
|
148 |
)
|
149 |
g_value = torch.empty(
|
150 |
-
|
151 |
device=gout.device,
|
152 |
requires_grad=False,
|
153 |
dtype=torch.bfloat16,
|
154 |
memory_format=torch.contiguous_format,
|
155 |
)
|
156 |
g_time_decay = torch.empty(
|
157 |
-
(
|
158 |
device=gout.device,
|
159 |
requires_grad=False,
|
160 |
dtype=torch.bfloat16,
|
161 |
memory_format=torch.contiguous_format,
|
162 |
)
|
163 |
g_time_first = torch.empty(
|
164 |
-
(
|
165 |
device=gout.device,
|
166 |
requires_grad=False,
|
167 |
dtype=torch.bfloat16,
|
168 |
memory_format=torch.contiguous_format,
|
169 |
)
|
170 |
-
rwkv5_cuda_kernel.
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
receptance,
|
176 |
key,
|
177 |
value,
|
@@ -185,133 +188,69 @@ class WKV_5(torch.autograd.Function):
|
|
185 |
g_time_decay,
|
186 |
g_time_first,
|
187 |
)
|
188 |
-
|
189 |
-
|
|
|
190 |
return (None, None, None, None, greceptance, g_key, g_value, g_time_decay, g_time_first)
|
191 |
|
192 |
|
193 |
-
def
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
receptance
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
receptance = receptance.to(torch.float32).view(Batch, SequenceLength, AttentionHeads, HeadSize).transpose(1, 2)
|
213 |
-
time_decay = torch.exp(-torch.exp(time_decay.float())).reshape(-1, 1, 1).reshape(AttentionHeads, -1, 1)
|
214 |
-
time_first = time_first.float().reshape(-1, 1, 1).reshape(AttentionHeads, -1, 1)
|
215 |
-
layer_norm_weight = layer_norm_weight.float()
|
216 |
-
layer_norm_bias = layer_norm_bias.float()
|
217 |
-
out = torch.zeros_like(key).reshape(Batch, SequenceLength, AttentionHeads, HeadSize)
|
218 |
-
for t in range(SequenceLength):
|
219 |
-
rt = receptance[:, :, t : t + 1, :]
|
220 |
-
kt = key[:, :, :, t : t + 1]
|
221 |
-
vt = value[:, :, t : t + 1, :]
|
222 |
-
at = kt @ vt
|
223 |
-
out[:, t] = (rt @ (time_first * at + state)).squeeze(2)
|
224 |
with torch.no_grad():
|
225 |
-
state =
|
226 |
-
|
227 |
-
out = out.reshape(Batch * SequenceLength, AttentionHeads * HeadSize)
|
228 |
-
out = F.group_norm(out, num_groups=AttentionHeads, weight=layer_norm_weight, bias=layer_norm_bias).reshape(
|
229 |
-
Batch, SequenceLength, AttentionHeads * HeadSize
|
230 |
-
)
|
231 |
-
out = out.to(dtype=hidden.dtype) * gate
|
232 |
-
out = out @ output_weight
|
233 |
|
234 |
return out, state
|
235 |
|
236 |
-
|
237 |
-
def
|
238 |
-
|
239 |
-
time_decay,
|
240 |
-
time_first,
|
241 |
-
receptance,
|
242 |
-
key,
|
243 |
-
value,
|
244 |
-
gate,
|
245 |
-
layer_norm_weight,
|
246 |
-
layer_norm_bias,
|
247 |
-
output_weight,
|
248 |
-
state,
|
249 |
-
):
|
250 |
-
Batch = hidden.shape[0]
|
251 |
-
AttentionHeads = time_decay.shape[0]
|
252 |
-
HeadSize = hidden.shape[-1] // AttentionHeads
|
253 |
-
SequenceLength = hidden.shape[1]
|
254 |
-
no_cuda = any(t.device.type != "cuda" for t in [time_decay, time_first, receptance, key, value])
|
255 |
# Launching the CUDA kernel for just one token will actually be slower (there is no for loop in the CPU version
|
256 |
# in this case).
|
257 |
one_token = key.size(1) == 1
|
258 |
-
if rwkv5_cuda_kernel is None or no_cuda or one_token:
|
259 |
-
return
|
260 |
-
|
261 |
-
time_decay,
|
262 |
-
time_first,
|
263 |
-
receptance,
|
264 |
-
key,
|
265 |
-
value,
|
266 |
-
gate,
|
267 |
-
layer_norm_weight,
|
268 |
-
layer_norm_bias,
|
269 |
-
output_weight,
|
270 |
-
state,
|
271 |
)
|
272 |
else:
|
273 |
-
|
274 |
-
Batch,
|
275 |
-
SequenceLength,
|
276 |
-
AttentionHeads * HeadSize,
|
277 |
-
AttentionHeads,
|
278 |
-
receptance,
|
279 |
-
key,
|
280 |
-
value,
|
281 |
-
time_decay,
|
282 |
-
time_first,
|
283 |
-
state,
|
284 |
-
)
|
285 |
-
out = out.reshape(Batch * SequenceLength, AttentionHeads * HeadSize)
|
286 |
-
out = F.group_norm(out, num_groups=AttentionHeads, weight=layer_norm_weight, bias=layer_norm_bias).reshape(
|
287 |
-
Batch, SequenceLength, AttentionHeads * HeadSize
|
288 |
-
)
|
289 |
-
out = out.to(dtype=hidden.dtype) * gate
|
290 |
-
out = out @ output_weight
|
291 |
-
return out, state
|
292 |
|
293 |
|
294 |
-
class
|
295 |
def __init__(self, config, layer_id=0):
|
296 |
super().__init__()
|
297 |
self.config = config
|
298 |
kernel_loaded = rwkv5_cuda_kernel is not None and rwkv5_cuda_kernel.head_size == config.head_size
|
299 |
if is_ninja_available() and is_torch_cuda_available() and not kernel_loaded:
|
300 |
try:
|
301 |
-
load_wkv5_cuda_kernel(config.
|
302 |
except Exception:
|
303 |
logger.info("Could not load the custom CUDA kernel for RWKV5 attention.")
|
304 |
self.layer_id = layer_id
|
305 |
hidden_size = config.hidden_size
|
306 |
-
|
307 |
-
self.num_attention_heads = num_attention_heads
|
308 |
-
attention_hidden_size = (
|
309 |
-
config.attention_hidden_size if config.attention_hidden_size is not None else hidden_size
|
310 |
-
)
|
311 |
self.attention_hidden_size = attention_hidden_size
|
|
|
|
|
312 |
|
313 |
-
self.time_decay = nn.Parameter(torch.empty(
|
314 |
-
self.time_faaaa = nn.Parameter(torch.empty(
|
315 |
self.time_mix_gate = nn.Parameter(torch.empty(1, 1, hidden_size))
|
316 |
|
317 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
@@ -324,7 +263,7 @@ class RwkvSelfAttention(nn.Module):
|
|
324 |
self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
325 |
self.gate = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
326 |
self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)
|
327 |
-
self.ln_x = nn.GroupNorm(
|
328 |
|
329 |
def extract_key_value(self, hidden, state=None):
|
330 |
# Mix hidden with the previous timestep to produce key, value, receptance
|
@@ -336,6 +275,7 @@ class RwkvSelfAttention(nn.Module):
|
|
336 |
shifted[:, 0] = state[0][:, :, self.layer_id]
|
337 |
if len(shifted.size()) == 2:
|
338 |
shifted = shifted.unsqueeze(1)
|
|
|
339 |
key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
|
340 |
value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
|
341 |
receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
|
@@ -353,28 +293,26 @@ class RwkvSelfAttention(nn.Module):
|
|
353 |
|
354 |
def forward(self, hidden, state=None, use_cache=False, seq_mode=True):
|
355 |
receptance, key, value, gate, state = self.extract_key_value(hidden, state=state)
|
|
|
|
|
|
|
|
|
356 |
layer_state = state[1][:, :, :, :, self.layer_id] if state is not None else None
|
357 |
-
|
358 |
-
|
359 |
-
self.time_decay,
|
360 |
-
self.time_faaaa,
|
361 |
-
receptance,
|
362 |
-
key,
|
363 |
-
value,
|
364 |
-
gate,
|
365 |
-
self.ln_x.weight,
|
366 |
-
self.ln_x.bias,
|
367 |
-
self.output.weight.t(),
|
368 |
-
state=layer_state,
|
369 |
)
|
370 |
|
371 |
if layer_state is not None:
|
372 |
state[1][:, :, :, :, self.layer_id] = layer_state
|
373 |
|
374 |
-
|
375 |
-
|
|
|
|
|
|
|
376 |
|
377 |
-
|
|
|
378 |
def __init__(self, config, layer_id=0):
|
379 |
super().__init__()
|
380 |
self.config = config
|
@@ -416,7 +354,7 @@ class RwkvFeedForward(nn.Module):
|
|
416 |
return receptance * value, state
|
417 |
|
418 |
|
419 |
-
#
|
420 |
class Rwkv5Block(nn.Module):
|
421 |
def __init__(self, config, layer_id):
|
422 |
super().__init__()
|
@@ -429,8 +367,8 @@ class Rwkv5Block(nn.Module):
|
|
429 |
self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
430 |
self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
431 |
|
432 |
-
self.attention =
|
433 |
-
self.feed_forward =
|
434 |
|
435 |
def forward(self, hidden, state=None, use_cache=False, output_attentions=False, seq_mode=True):
|
436 |
if self.layer_id == 0:
|
@@ -450,6 +388,7 @@ class Rwkv5Block(nn.Module):
|
|
450 |
return outputs
|
451 |
|
452 |
|
|
|
453 |
class Rwkv5PreTrainedModel(PreTrainedModel):
|
454 |
"""
|
455 |
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
@@ -457,19 +396,20 @@ class Rwkv5PreTrainedModel(PreTrainedModel):
|
|
457 |
"""
|
458 |
|
459 |
config_class = Rwkv5Config
|
460 |
-
base_model_prefix = "
|
461 |
_no_split_modules = ["Rwkv5Block"]
|
462 |
_keep_in_fp32_modules = ["time_decay", "time_first"]
|
463 |
supports_gradient_checkpointing = True
|
464 |
|
465 |
def _init_weights(self, module):
|
466 |
"""Initialize the weights."""
|
467 |
-
if isinstance(module,
|
468 |
layer_id = module.layer_id
|
469 |
num_hidden_layers = module.config.num_hidden_layers
|
470 |
hidden_size = module.config.hidden_size
|
471 |
attention_hidden_size = module.attention_hidden_size
|
472 |
-
|
|
|
473 |
|
474 |
ratio_0_to_1 = layer_id / (num_hidden_layers - 1) # 0 to 1
|
475 |
ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0
|
@@ -496,15 +436,15 @@ class Rwkv5PreTrainedModel(PreTrainedModel):
|
|
496 |
)
|
497 |
|
498 |
with torch.no_grad():
|
499 |
-
module.time_decay.data = decay_speed.reshape(
|
500 |
-
module.time_faaaa.data = tmp.reshape(
|
501 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
502 |
|
503 |
module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
|
504 |
module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
505 |
module.time_mix_gate.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
506 |
|
507 |
-
elif isinstance(module,
|
508 |
layer_id = module.layer_id
|
509 |
num_hidden_layers = module.config.num_hidden_layers
|
510 |
hidden_size = module.config.hidden_size
|
@@ -523,11 +463,11 @@ class Rwkv5PreTrainedModel(PreTrainedModel):
|
|
523 |
module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0)
|
524 |
|
525 |
|
526 |
-
#
|
527 |
@dataclass
|
528 |
class Rwkv5Output(ModelOutput):
|
529 |
"""
|
530 |
-
Class for the
|
531 |
|
532 |
Args:
|
533 |
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
@@ -551,7 +491,7 @@ class Rwkv5Output(ModelOutput):
|
|
551 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
552 |
|
553 |
|
554 |
-
#
|
555 |
@dataclass
|
556 |
class Rwkv5CausalLMOutput(ModelOutput):
|
557 |
"""
|
@@ -582,7 +522,7 @@ class Rwkv5CausalLMOutput(ModelOutput):
|
|
582 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
583 |
|
584 |
|
585 |
-
|
586 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
587 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
588 |
etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module)
|
@@ -595,7 +535,7 @@ RWKV_START_DOCSTRING = r"""
|
|
595 |
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
596 |
"""
|
597 |
|
598 |
-
|
599 |
Args:
|
600 |
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
601 |
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
@@ -625,8 +565,8 @@ RWKV_INPUTS_DOCSTRING = r"""
|
|
625 |
|
626 |
|
627 |
@add_start_docstrings(
|
628 |
-
"The bare
|
629 |
-
|
630 |
)
|
631 |
class Rwkv5Model(Rwkv5PreTrainedModel):
|
632 |
def __init__(self, config):
|
@@ -648,7 +588,7 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
648 |
def set_input_embeddings(self, new_embeddings):
|
649 |
self.embeddings = new_embeddings
|
650 |
|
651 |
-
@add_start_docstrings_to_model_forward(
|
652 |
@add_code_sample_docstrings(
|
653 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
654 |
output_type=Rwkv5Output,
|
@@ -669,6 +609,7 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
669 |
output_hidden_states = (
|
670 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
671 |
)
|
|
|
672 |
# rwkv5 only support inference in huggingface.
|
673 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
674 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
@@ -686,9 +627,10 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
686 |
if inputs_embeds is None:
|
687 |
inputs_embeds = self.embeddings(input_ids)
|
688 |
|
689 |
-
if
|
690 |
state = []
|
691 |
-
|
|
|
692 |
state_attn_x = torch.zeros(
|
693 |
(inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers),
|
694 |
dtype=inputs_embeds.dtype,
|
@@ -698,9 +640,9 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
698 |
state_attn_kv = torch.zeros(
|
699 |
(
|
700 |
inputs_embeds.size(0),
|
701 |
-
|
702 |
-
|
703 |
-
|
704 |
self.config.num_hidden_layers,
|
705 |
),
|
706 |
dtype=torch.float32,
|
@@ -765,8 +707,16 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
765 |
block.attention.output.weight.mul_(2 ** int(block_id // self.config.rescale_every))
|
766 |
block.feed_forward.value.weight.mul_(2 ** int(block_id // self.config.rescale_every))
|
767 |
else:
|
768 |
-
|
769 |
-
block.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
770 |
|
771 |
self.layers_are_rescaled = not self.training
|
772 |
|
@@ -798,8 +748,9 @@ class Rwkv5Model(Rwkv5PreTrainedModel):
|
|
798 |
The RWKV5 Model transformer with a language modeling head on top (linear layer with weights tied to the input
|
799 |
embeddings).
|
800 |
""",
|
801 |
-
|
802 |
)
|
|
|
803 |
class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
804 |
_tied_weights_keys = ["head.weight"]
|
805 |
|
@@ -831,7 +782,7 @@ class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
|
831 |
model_inputs["state"] = state
|
832 |
return model_inputs
|
833 |
|
834 |
-
@add_start_docstrings_to_model_forward(
|
835 |
@add_code_sample_docstrings(
|
836 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
837 |
output_type=Rwkv5CausalLMOutput,
|
@@ -857,7 +808,7 @@ class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
|
857 |
"""
|
858 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
859 |
|
860 |
-
|
861 |
input_ids,
|
862 |
inputs_embeds=inputs_embeds,
|
863 |
state=state,
|
@@ -866,7 +817,7 @@ class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
|
866 |
output_hidden_states=output_hidden_states,
|
867 |
return_dict=return_dict,
|
868 |
)
|
869 |
-
hidden_states =
|
870 |
|
871 |
logits = self.head(hidden_states)
|
872 |
|
@@ -882,13 +833,13 @@ class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
|
882 |
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
883 |
|
884 |
if not return_dict:
|
885 |
-
output = (logits,) +
|
886 |
return ((loss,) + output) if loss is not None else output
|
887 |
|
888 |
return Rwkv5CausalLMOutput(
|
889 |
loss=loss,
|
890 |
logits=logits,
|
891 |
-
state=
|
892 |
-
hidden_states=
|
893 |
-
attentions=
|
894 |
)
|
|
|
1 |
# coding=utf-8
|
2 |
+
# Copyright 2024 The RWKV team and HuggingFace Inc. team.
|
|
|
3 |
#
|
4 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
# you may not use this file except in compliance with the License.
|
|
|
44 |
_CHECKPOINT_FOR_DOC = "RWKV/rwkv-5-world-1b5"
|
45 |
_CONFIG_FOR_DOC = "Rwkv5Config"
|
46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
rwkv5_cuda_kernel = None
|
48 |
|
49 |
|
|
|
53 |
|
54 |
global rwkv5_cuda_kernel
|
55 |
|
56 |
+
kernel_folder = Path(__file__).parent.resolve()
|
57 |
cuda_kernel_files = [kernel_folder / f for f in ["wkv5_op.cpp", "wkv5_cuda.cu"]]
|
58 |
|
59 |
# Only load the kernel if it's not been loaded yet or if we changed the context length
|
60 |
if rwkv5_cuda_kernel is not None and rwkv5_cuda_kernel.head_size == head_size:
|
61 |
return
|
62 |
|
63 |
+
logger.info(f"Loading CUDA kernel for RWKV5 at head size of {head_size}.")
|
64 |
|
65 |
flags = [
|
66 |
"-res-usage",
|
|
|
80 |
rwkv5_cuda_kernel.head_size = head_size
|
81 |
|
82 |
|
83 |
+
class Rwkv5LinearAttention(torch.autograd.Function):
|
84 |
@staticmethod
|
85 |
def forward(ctx, receptance, key, value, time_decay, time_first, state):
|
86 |
with torch.no_grad():
|
87 |
+
assert receptance.dtype == torch.bfloat16
|
88 |
+
assert key.dtype == torch.bfloat16
|
89 |
+
assert value.dtype == torch.bfloat16
|
90 |
+
assert time_decay.dtype == torch.bfloat16
|
91 |
+
assert time_first.dtype == torch.bfloat16
|
92 |
+
assert state.dtype == torch.float32
|
93 |
+
batch, seq_length, hidden_size = key.shape
|
94 |
+
num_heads = time_decay.shape[0]
|
95 |
+
ctx.batch = batch
|
96 |
+
ctx.seq_length = seq_length
|
97 |
+
ctx.hidden_size = hidden_size
|
98 |
+
ctx.num_heads = num_heads
|
99 |
e_time_decay = (-torch.exp(time_decay.float())).contiguous()
|
100 |
ee_time_decay = (torch.exp(e_time_decay)).contiguous()
|
101 |
+
assert ee_time_decay.dtype == torch.float32
|
102 |
ctx.save_for_backward(receptance, key, value, ee_time_decay, e_time_decay, time_first)
|
103 |
out = torch.empty(
|
104 |
+
(batch, seq_length, hidden_size),
|
105 |
device=receptance.device,
|
106 |
dtype=torch.bfloat16,
|
107 |
memory_format=torch.contiguous_format,
|
108 |
)
|
109 |
+
state = state.clone()
|
110 |
+
rwkv5_cuda_kernel.forward_bf16(
|
111 |
+
batch,
|
112 |
+
seq_length,
|
113 |
+
hidden_size,
|
114 |
+
num_heads,
|
115 |
+
state,
|
116 |
receptance,
|
117 |
key,
|
118 |
value,
|
119 |
ee_time_decay,
|
120 |
time_first,
|
121 |
out,
|
|
|
122 |
)
|
123 |
return out, state
|
124 |
|
|
|
126 |
def backward(ctx, gout):
|
127 |
with torch.no_grad():
|
128 |
assert gout.dtype == torch.bfloat16
|
129 |
+
batch = ctx.batch
|
130 |
+
seq_length = ctx.seq_length
|
131 |
+
hidden_size = ctx.hidden_size
|
132 |
+
num_heads = ctx.num_heads
|
133 |
receptance, key, value, ee_time_decay, e_time_decay, time_first = ctx.saved_tensors
|
134 |
+
|
135 |
+
global_shape = (batch, seq_length, hidden_size)
|
136 |
+
|
137 |
+
# TODO dtype should not be forced here IMO
|
138 |
greceptance = torch.empty(
|
139 |
+
global_shape,
|
140 |
device=gout.device,
|
141 |
requires_grad=False,
|
142 |
dtype=torch.bfloat16,
|
143 |
memory_format=torch.contiguous_format,
|
144 |
)
|
145 |
g_key = torch.empty(
|
146 |
+
global_shape,
|
147 |
device=gout.device,
|
148 |
requires_grad=False,
|
149 |
dtype=torch.bfloat16,
|
150 |
memory_format=torch.contiguous_format,
|
151 |
)
|
152 |
g_value = torch.empty(
|
153 |
+
global_shape,
|
154 |
device=gout.device,
|
155 |
requires_grad=False,
|
156 |
dtype=torch.bfloat16,
|
157 |
memory_format=torch.contiguous_format,
|
158 |
)
|
159 |
g_time_decay = torch.empty(
|
160 |
+
(batch, hidden_size),
|
161 |
device=gout.device,
|
162 |
requires_grad=False,
|
163 |
dtype=torch.bfloat16,
|
164 |
memory_format=torch.contiguous_format,
|
165 |
)
|
166 |
g_time_first = torch.empty(
|
167 |
+
(batch, hidden_size),
|
168 |
device=gout.device,
|
169 |
requires_grad=False,
|
170 |
dtype=torch.bfloat16,
|
171 |
memory_format=torch.contiguous_format,
|
172 |
)
|
173 |
+
rwkv5_cuda_kernel.backward_bf16(
|
174 |
+
batch,
|
175 |
+
seq_length,
|
176 |
+
hidden_size,
|
177 |
+
num_heads,
|
178 |
receptance,
|
179 |
key,
|
180 |
value,
|
|
|
188 |
g_time_decay,
|
189 |
g_time_first,
|
190 |
)
|
191 |
+
head_size = hidden_size // num_heads
|
192 |
+
g_time_decay = torch.sum(g_time_decay, 0).view(num_heads, head_size)
|
193 |
+
g_time_first = torch.sum(g_time_first, 0).view(num_heads, head_size)
|
194 |
return (None, None, None, None, greceptance, g_key, g_value, g_time_decay, g_time_first)
|
195 |
|
196 |
|
197 |
+
def rwkv5_linear_attention_cpu(receptance, key, value, time_decay, time_first, state):
|
198 |
+
input_dtype = receptance.dtype
|
199 |
+
# For CPU fallback. Will be slower and probably take more memory than the custom CUDA kernel if not executed
|
200 |
+
# within a torch.no_grad.
|
201 |
+
batch, seq_length, hidden_size = receptance.shape
|
202 |
+
num_heads, head_size = time_first.shape
|
203 |
+
key = key.float().view(batch, seq_length, num_heads, head_size).transpose(1, 2).transpose(-2, -1)
|
204 |
+
value = value.float().view(batch, seq_length, num_heads, head_size).transpose(1, 2)
|
205 |
+
receptance = receptance.float().view(batch, seq_length, num_heads, head_size).transpose(1, 2)
|
206 |
+
time_decay = torch.exp(-torch.exp(time_decay.float())).reshape(-1, 1, 1).reshape(num_heads, -1, 1)
|
207 |
+
time_first = time_first.float().reshape(-1, 1, 1).reshape(num_heads, -1, 1)
|
208 |
+
out = torch.zeros_like(key).reshape(batch, seq_length, num_heads, head_size)
|
209 |
+
|
210 |
+
for current_index in range(seq_length):
|
211 |
+
current_receptance = receptance[:, :, current_index:current_index+1, :]
|
212 |
+
current_key = key[:, :, :, current_index:current_index+1]
|
213 |
+
current_value = value[:, :, current_index:current_index+1, :]
|
214 |
+
attention_output = current_key @ current_value
|
215 |
+
out[:, current_index] = (current_receptance @ (time_first * attention_output + state)).squeeze(2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
216 |
with torch.no_grad():
|
217 |
+
state = attention_output + time_decay * state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
218 |
|
219 |
return out, state
|
220 |
|
221 |
+
# copied from RWKV but with receptance
|
222 |
+
def RWKV5_linear_attention(training, receptance, key, value, time_decay, time_first, state):
|
223 |
+
no_cuda = any(t.device.type != "cuda" for t in [time_decay, time_first, key, value])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
# Launching the CUDA kernel for just one token will actually be slower (there is no for loop in the CPU version
|
225 |
# in this case).
|
226 |
one_token = key.size(1) == 1
|
227 |
+
if not training or rwkv5_cuda_kernel is None or no_cuda or one_token:
|
228 |
+
return rwkv5_linear_attention_cpu(
|
229 |
+
receptance, key, value, time_decay, time_first, state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
230 |
)
|
231 |
else:
|
232 |
+
return Rwkv5LinearAttention.apply(receptance, key, value, time_decay, time_first, state)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
233 |
|
234 |
|
235 |
+
class Rwkv5SelfAttention(nn.Module):
|
236 |
def __init__(self, config, layer_id=0):
|
237 |
super().__init__()
|
238 |
self.config = config
|
239 |
kernel_loaded = rwkv5_cuda_kernel is not None and rwkv5_cuda_kernel.head_size == config.head_size
|
240 |
if is_ninja_available() and is_torch_cuda_available() and not kernel_loaded:
|
241 |
try:
|
242 |
+
load_wkv5_cuda_kernel(config.head_size)
|
243 |
except Exception:
|
244 |
logger.info("Could not load the custom CUDA kernel for RWKV5 attention.")
|
245 |
self.layer_id = layer_id
|
246 |
hidden_size = config.hidden_size
|
247 |
+
attention_hidden_size = config.attention_hidden_size
|
|
|
|
|
|
|
|
|
248 |
self.attention_hidden_size = attention_hidden_size
|
249 |
+
head_size = config.head_size
|
250 |
+
num_heads = attention_hidden_size // head_size
|
251 |
|
252 |
+
self.time_decay = nn.Parameter(torch.empty(num_heads, head_size))
|
253 |
+
self.time_faaaa = nn.Parameter(torch.empty(num_heads, head_size))
|
254 |
self.time_mix_gate = nn.Parameter(torch.empty(1, 1, hidden_size))
|
255 |
|
256 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
|
|
263 |
self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
264 |
self.gate = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
265 |
self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)
|
266 |
+
self.ln_x = nn.GroupNorm(num_heads, hidden_size)
|
267 |
|
268 |
def extract_key_value(self, hidden, state=None):
|
269 |
# Mix hidden with the previous timestep to produce key, value, receptance
|
|
|
275 |
shifted[:, 0] = state[0][:, :, self.layer_id]
|
276 |
if len(shifted.size()) == 2:
|
277 |
shifted = shifted.unsqueeze(1)
|
278 |
+
|
279 |
key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
|
280 |
value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
|
281 |
receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
|
|
|
293 |
|
294 |
def forward(self, hidden, state=None, use_cache=False, seq_mode=True):
|
295 |
receptance, key, value, gate, state = self.extract_key_value(hidden, state=state)
|
296 |
+
|
297 |
+
B,T,C = receptance.shape
|
298 |
+
H, S = self.time_faaaa.shape
|
299 |
+
|
300 |
layer_state = state[1][:, :, :, :, self.layer_id] if state is not None else None
|
301 |
+
out, layer_state = RWKV5_linear_attention(
|
302 |
+
self.training, receptance, key, value, self.time_decay, self.time_faaaa, layer_state
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
303 |
)
|
304 |
|
305 |
if layer_state is not None:
|
306 |
state[1][:, :, :, :, self.layer_id] = layer_state
|
307 |
|
308 |
+
out = out.reshape(B * T, H * S)
|
309 |
+
out = F.group_norm(out / self.config.head_size_divisor, num_groups=H, weight=self.ln_x.weight.to(out.dtype), bias=self.ln_x.bias.to(out.dtype), eps=self.ln_x.eps).reshape(B, T, H * S)
|
310 |
+
out = out.to(dtype=hidden.dtype) * gate
|
311 |
+
out = self.output(out)
|
312 |
+
return out, state
|
313 |
|
314 |
+
# Copied from rwkv exceot for the intermediate size
|
315 |
+
class Rwkv5FeedForward(nn.Module):
|
316 |
def __init__(self, config, layer_id=0):
|
317 |
super().__init__()
|
318 |
self.config = config
|
|
|
354 |
return receptance * value, state
|
355 |
|
356 |
|
357 |
+
# Copied from transformers.models.rwkv.modeling_rwkv.RwkvBlock with Rwkv->Rwkv5
|
358 |
class Rwkv5Block(nn.Module):
|
359 |
def __init__(self, config, layer_id):
|
360 |
super().__init__()
|
|
|
367 |
self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
368 |
self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
|
369 |
|
370 |
+
self.attention = Rwkv5SelfAttention(config, layer_id)
|
371 |
+
self.feed_forward = Rwkv5FeedForward(config, layer_id)
|
372 |
|
373 |
def forward(self, hidden, state=None, use_cache=False, output_attentions=False, seq_mode=True):
|
374 |
if self.layer_id == 0:
|
|
|
388 |
return outputs
|
389 |
|
390 |
|
391 |
+
# Copied from transformers.models.rwkv.modeling_rwkv.RwkvPreTrainedModel with Rwkv->Rwkv5
|
392 |
class Rwkv5PreTrainedModel(PreTrainedModel):
|
393 |
"""
|
394 |
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
|
|
396 |
"""
|
397 |
|
398 |
config_class = Rwkv5Config
|
399 |
+
base_model_prefix = "rwkv5"
|
400 |
_no_split_modules = ["Rwkv5Block"]
|
401 |
_keep_in_fp32_modules = ["time_decay", "time_first"]
|
402 |
supports_gradient_checkpointing = True
|
403 |
|
404 |
def _init_weights(self, module):
|
405 |
"""Initialize the weights."""
|
406 |
+
if isinstance(module, Rwkv5SelfAttention):
|
407 |
layer_id = module.layer_id
|
408 |
num_hidden_layers = module.config.num_hidden_layers
|
409 |
hidden_size = module.config.hidden_size
|
410 |
attention_hidden_size = module.attention_hidden_size
|
411 |
+
head_size = module.config.head_size
|
412 |
+
num_heads = attention_hidden_size // head_size
|
413 |
|
414 |
ratio_0_to_1 = layer_id / (num_hidden_layers - 1) # 0 to 1
|
415 |
ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0
|
|
|
436 |
)
|
437 |
|
438 |
with torch.no_grad():
|
439 |
+
module.time_decay.data = decay_speed.reshape(num_heads, head_size)
|
440 |
+
module.time_faaaa.data = tmp.reshape(num_heads, head_size)
|
441 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
442 |
|
443 |
module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
|
444 |
module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
445 |
module.time_mix_gate.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
446 |
|
447 |
+
elif isinstance(module, Rwkv5FeedForward):
|
448 |
layer_id = module.layer_id
|
449 |
num_hidden_layers = module.config.num_hidden_layers
|
450 |
hidden_size = module.config.hidden_size
|
|
|
463 |
module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0)
|
464 |
|
465 |
|
466 |
+
# Copied from transformers.models.rwkv.modeling_rwkv.RwkvOutput with Rwkv->Rwkv5
|
467 |
@dataclass
|
468 |
class Rwkv5Output(ModelOutput):
|
469 |
"""
|
470 |
+
Class for the RWKV5 model outputs.
|
471 |
|
472 |
Args:
|
473 |
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
|
|
491 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
492 |
|
493 |
|
494 |
+
# Copied from transformers.models.rwkv.modeling_rwkv.RwkvCausalLMOutput with Rwkv->Rwkv5
|
495 |
@dataclass
|
496 |
class Rwkv5CausalLMOutput(ModelOutput):
|
497 |
"""
|
|
|
522 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
523 |
|
524 |
|
525 |
+
RWKV5_START_DOCSTRING = r"""
|
526 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
527 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
528 |
etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module)
|
|
|
535 |
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
536 |
"""
|
537 |
|
538 |
+
RWKV5_INPUTS_DOCSTRING = r"""
|
539 |
Args:
|
540 |
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
541 |
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
|
|
565 |
|
566 |
|
567 |
@add_start_docstrings(
|
568 |
+
"The bare RWKV5 Model transformer outputting raw hidden-states without any specific head on top.",
|
569 |
+
RWKV5_START_DOCSTRING,
|
570 |
)
|
571 |
class Rwkv5Model(Rwkv5PreTrainedModel):
|
572 |
def __init__(self, config):
|
|
|
588 |
def set_input_embeddings(self, new_embeddings):
|
589 |
self.embeddings = new_embeddings
|
590 |
|
591 |
+
@add_start_docstrings_to_model_forward(RWKV5_INPUTS_DOCSTRING)
|
592 |
@add_code_sample_docstrings(
|
593 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
594 |
output_type=Rwkv5Output,
|
|
|
609 |
output_hidden_states = (
|
610 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
611 |
)
|
612 |
+
# FIXME - training is supportable with the CUDA code
|
613 |
# rwkv5 only support inference in huggingface.
|
614 |
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
615 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
|
|
627 |
if inputs_embeds is None:
|
628 |
inputs_embeds = self.embeddings(input_ids)
|
629 |
|
630 |
+
if state is None:
|
631 |
state = []
|
632 |
+
head_size = self.config.head_size
|
633 |
+
num_heads = self.config.attention_hidden_size // head_size
|
634 |
state_attn_x = torch.zeros(
|
635 |
(inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers),
|
636 |
dtype=inputs_embeds.dtype,
|
|
|
640 |
state_attn_kv = torch.zeros(
|
641 |
(
|
642 |
inputs_embeds.size(0),
|
643 |
+
num_heads,
|
644 |
+
head_size,
|
645 |
+
head_size,
|
646 |
self.config.num_hidden_layers,
|
647 |
),
|
648 |
dtype=torch.float32,
|
|
|
707 |
block.attention.output.weight.mul_(2 ** int(block_id // self.config.rescale_every))
|
708 |
block.feed_forward.value.weight.mul_(2 ** int(block_id // self.config.rescale_every))
|
709 |
else:
|
710 |
+
# Deal with quantization statistics
|
711 |
+
if hasattr(block.attention.output.weight, "SCB"):
|
712 |
+
block.attention.output.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
|
713 |
+
block.feed_forward.value.weight.SCB.div_(2 ** int(block_id // self.config.rescale_every))
|
714 |
+
elif hasattr(block.attention.output.weight, "quant_state"):
|
715 |
+
self._bnb_4bit_dequantize_and_rescale(block.attention.output, block_id)
|
716 |
+
self._bnb_4bit_dequantize_and_rescale(block.feed_forward.value, block_id)
|
717 |
+
else:
|
718 |
+
block.attention.output.weight.div_(2 ** int(block_id // self.config.rescale_every))
|
719 |
+
block.feed_forward.value.weight.div_(2 ** int(block_id // self.config.rescale_every))
|
720 |
|
721 |
self.layers_are_rescaled = not self.training
|
722 |
|
|
|
748 |
The RWKV5 Model transformer with a language modeling head on top (linear layer with weights tied to the input
|
749 |
embeddings).
|
750 |
""",
|
751 |
+
RWKV5_START_DOCSTRING,
|
752 |
)
|
753 |
+
# Copied from transformers.models.rwkv.modeling_rwkv.RwkvForCausalLM with Rwkv->Rwkv5
|
754 |
class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
755 |
_tied_weights_keys = ["head.weight"]
|
756 |
|
|
|
782 |
model_inputs["state"] = state
|
783 |
return model_inputs
|
784 |
|
785 |
+
@add_start_docstrings_to_model_forward(RWKV5_INPUTS_DOCSTRING)
|
786 |
@add_code_sample_docstrings(
|
787 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
788 |
output_type=Rwkv5CausalLMOutput,
|
|
|
808 |
"""
|
809 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
810 |
|
811 |
+
outputs = self.rwkv(
|
812 |
input_ids,
|
813 |
inputs_embeds=inputs_embeds,
|
814 |
state=state,
|
|
|
817 |
output_hidden_states=output_hidden_states,
|
818 |
return_dict=return_dict,
|
819 |
)
|
820 |
+
hidden_states = outputs[0]
|
821 |
|
822 |
logits = self.head(hidden_states)
|
823 |
|
|
|
833 |
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
834 |
|
835 |
if not return_dict:
|
836 |
+
output = (logits,) + outputs[1:]
|
837 |
return ((loss,) + output) if loss is not None else output
|
838 |
|
839 |
return Rwkv5CausalLMOutput(
|
840 |
loss=loss,
|
841 |
logits=logits,
|
842 |
+
state=outputs.state,
|
843 |
+
hidden_states=outputs.hidden_states,
|
844 |
+
attentions=outputs.attentions,
|
845 |
)
|