Spaces:
Sleeping
Sleeping
File size: 14,161 Bytes
687d97d 6e01670 687d97d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 |
import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig
from .modeling_llama_kv import LlamaForCausalLM as KVLlamaForCausalLM
from .utils import *
from .kv_cache import initialize_past_key_values
from .choices import mc_sim_7b_63
from transformers import AutoTokenizer
import os
from huggingface_hub import hf_hub_download
from .cnets import Model
from .configs import EConfig
class ResBlock(nn.Module):
"""
A Residual Block module.
This module performs a linear transformation followed by a SiLU activation,
and then adds the result to the original input, creating a residual connection.
Args:
hidden_size (int): The size of the hidden layers in the block.
"""
def __init__(self, hidden_size):
super().__init__()
self.linear = nn.Linear(hidden_size, hidden_size)
# Initialize as an identity mapping
torch.nn.init.zeros_(self.linear.weight)
# Use SiLU activation to keep consistent with the Llama model
self.act = nn.SiLU()
def forward(self, x):
"""
Forward pass of the ResBlock.
Args:
x (torch.Tensor): Input tensor.
Returns:
torch.Tensor: Output after the residual connection and activation.
"""
return x + self.act(self.linear(x))
class EaModel(nn.Module):
def __init__(
self,
base_model,
base_model_name_or_path,
ea_model_path,
):
super().__init__()
self.base_model = base_model
self.config = base_model.config
self.hidden_size = base_model.lm_head.weight.shape[-1]
self.vocab_size = base_model.lm_head.weight.shape[0]
self.base_model_name_or_path = base_model_name_or_path
self.tokenizer = AutoTokenizer.from_pretrained(self.base_model_name_or_path)
config = EConfig.from_pretrained(ea_model_path)
self.ea_layer = Model(config)
device = base_model.model.layers[-1].self_attn.q_proj.weight.device
self.ea_layer.to(torch.float16).to(device)
self.ea_layer.init_tree()
def get_tokenizer(self):
"""Get the tokenizer of the base model.
Returns:
Tokenizer: The tokenizer of the base model.
"""
return self.tokenizer
@classmethod
def from_pretrained(
cls,
base_model_path=None,
ea_model_path=None,
**kwargs,
):
base_model = KVLlamaForCausalLM.from_pretrained(
base_model_path, **kwargs
)
model = cls(
base_model,
base_model_path,
ea_model_path
)
ea_layer_state_dict = torch.load(os.path.join(ea_model_path,"pytorch_model.bin"), map_location=base_model.device)
model.ea_layer.load_state_dict(ea_layer_state_dict, strict=False)
return model
def forward(
self,
input_ids=None,
attention_mask=None,
labels=None,
past_key_values=None,
output_orig=False,
position_ids=None,
init=True,
logits_processor=None
):
with torch.inference_mode():
# Pass input through the base model
outputs = self.base_model.model(
input_ids=input_ids,
attention_mask=attention_mask,
past_key_values=past_key_values,
position_ids=position_ids,
)
if output_orig:
orig = self.base_model.lm_head(outputs[0])
hidden_states = outputs[0].clone()
if init:
if logits_processor is not None:
logits=orig[:, -1]
logits=logits_processor(None,logits)
probabilities = torch.nn.functional.softmax(logits, dim=1)
token=torch.multinomial(probabilities, 1)
else:
token = torch.argmax(orig[:,-1])
token=token[None,None]
input_ids=torch.cat((input_ids,token.to(input_ids.device)),dim=1)
# Clone the output hidden states
ea_logits = self.ea_layer.topK_genrate(hidden_states,input_ids,self.base_model.lm_head,logits_processor)
if output_orig:
return ea_logits, outputs, orig,hidden_states,token
return ea_logits,hidden_states,token
else:
if output_orig:
return outputs,orig,hidden_states
@torch.no_grad()
def eagenerate(
self,
input_ids,
temperature=0.0,
top_p=0.0,
top_k=0.0,
max_new_tokens=512,
max_length=2048,
tree_choices=mc_sim_7b_63,
):
if temperature>1e-5:
logits_processor=prepare_logits_processor(temperature=temperature,top_p=top_p,top_k=top_k)
else:
logits_processor=None
assert input_ids.shape[0] == 1, "Only support batch size 1 for now!!"
# Avoid modifying the input_ids in-place
input_ids = input_ids.clone()
self.ea_layer.reset_kv()
if hasattr(self, "tree_choices") and self.tree_choices == tree_choices:
tree_buffers = self.tree_buffers
else:
tree_buffers = generate_tree_buffers(
tree_choices, device=self.base_model.model.layers[-1].self_attn.q_proj.weight.device
)
self.tree_buffers = tree_buffers
self.tree_choices = tree_choices
# Initialize the past key and value states
if hasattr(self, "past_key_values"):
past_key_values = self.past_key_values
past_key_values_data = self.past_key_values_data
current_length_data = self.current_length_data
# Reset the past key and value states
current_length_data.zero_()
else:
(
past_key_values,
past_key_values_data,
current_length_data,
) = initialize_past_key_values(self.base_model)
self.past_key_values = past_key_values
self.past_key_values_data = past_key_values_data
self.current_length_data = current_length_data
input_len = input_ids.shape[1]
reset_tree_mode(self)
tree_logits, logits, hidden_state, sample_token = initialize_tree(
input_ids, self, tree_buffers["tree_attn_mask"], past_key_values, logits_processor
)
new_token = 0
for idx in range(max_length):
candidates, cart_candidates_prob, tree_candidates = generate_candidates(
tree_logits,
tree_buffers["tree_indices"],
tree_buffers["retrieve_indices"],
sample_token,
logits_processor
)
logits, hidden_state_new, outputs = tree_decoding(
self,
tree_candidates,
past_key_values,
tree_buffers["tree_position_ids"],
input_ids,
tree_buffers["retrieve_indices"],
)
best_candidate, accept_length, sample_p = evaluate_posterior(
logits, candidates, logits_processor, cart_candidates_prob
)
input_ids, tree_logits, new_token, hidden_state, sample_token = update_inference_inputs(
input_ids,
candidates,
best_candidate,
accept_length,
tree_buffers["retrieve_indices"],
logits_processor,
logits,
tree_logits,
new_token,
past_key_values_data,
current_length_data,
self,
hidden_state,
hidden_state_new,
sample_p
)
if self.tokenizer.eos_token_id in input_ids[0, input_len:].tolist():
return input_ids
if new_token > max_new_tokens:
return input_ids
if input_ids.shape[1] > max_length:
return input_ids
@torch.no_grad()
def ea_generate(
self,
input_ids,
temperature=0.0,
top_p=0.0,
top_k=0.0,
max_steps=512,
tree_choices=mc_sim_7b_63,
):
if temperature > 1e-5:
logits_processor = prepare_logits_processor(temperature=temperature, top_p=top_p, top_k=top_k)
else:
logits_processor = None
assert input_ids.shape[0] == 1, "Only support batch size 1 for now!!"
# Avoid modifying the input_ids in-place
input_ids = input_ids.clone()
self.ea_layer.reset_kv()
if hasattr(self, "tree_choices") and self.tree_choices == tree_choices:
tree_buffers = self.tree_buffers
else:
tree_buffers = generate_tree_buffers(
tree_choices, device=self.base_model.model.layers[-1].self_attn.q_proj.weight.device
)
self.tree_buffers = tree_buffers
self.tree_choices = tree_choices
# Initialize the past key and value states
if hasattr(self, "past_key_values"):
past_key_values = self.past_key_values
past_key_values_data = self.past_key_values_data
current_length_data = self.current_length_data
# Reset the past key and value states
current_length_data.zero_()
else:
(
past_key_values,
past_key_values_data,
current_length_data,
) = initialize_past_key_values(self.base_model)
self.past_key_values = past_key_values
self.past_key_values_data = past_key_values_data
self.current_length_data = current_length_data
input_len = input_ids.shape[1]
reset_tree_mode(self)
tree_logits, logits, hidden_state, sample_token = initialize_tree(
input_ids, self, tree_buffers["tree_attn_mask"], past_key_values, logits_processor
)
new_token = 0
for idx in range(max_steps):
candidates, cart_candidates_prob, tree_candidates = generate_candidates(
tree_logits,
tree_buffers["tree_indices"],
tree_buffers["retrieve_indices"],
sample_token,
logits_processor
)
logits, hidden_state_new, outputs = tree_decoding(
self,
tree_candidates,
past_key_values,
tree_buffers["tree_position_ids"],
input_ids,
tree_buffers["retrieve_indices"],
)
best_candidate, accept_length, sample_p = evaluate_posterior(
logits, candidates, logits_processor, cart_candidates_prob
)
input_ids, tree_logits, new_token, hidden_state, sample_token = update_inference_inputs(
input_ids,
candidates,
best_candidate,
accept_length,
tree_buffers["retrieve_indices"],
logits_processor,
logits,
tree_logits,
new_token,
past_key_values_data,
current_length_data,
self,
hidden_state,
hidden_state_new,
sample_p
)
yield input_ids
if self.tokenizer.eos_token_id in input_ids[0, input_len:].tolist():
break
if new_token > 1024:
break
if input_ids.shape[1] > 1960:
break
@torch.no_grad()
def naive_generate(
self,
input_ids,
temperature=0.0,
top_p=0.0,
top_k=0.0,
max_steps=512,
tree_choices=mc_sim_7b_63,
):
if temperature > 1e-5:
logits_processor = prepare_logits_processor(temperature=temperature, top_p=top_p, top_k=top_k)
else:
logits_processor = None
assert input_ids.shape[0] == 1, "Only support batch size 1 for now!!"
# Avoid modifying the input_ids in-place
input_ids = input_ids.clone()
self.ea_layer.reset_kv()
if hasattr(self, "tree_choices") and self.tree_choices == tree_choices:
tree_buffers = self.tree_buffers
else:
tree_buffers = generate_tree_buffers(
tree_choices, device=self.base_model.model.layers[-1].self_attn.q_proj.weight.device
)
self.tree_buffers = tree_buffers
self.tree_choices = tree_choices
# Initialize the past key and value states
if hasattr(self, "past_key_values"):
past_key_values = self.past_key_values
past_key_values_data = self.past_key_values_data
current_length_data = self.current_length_data
# Reset the past key and value states
current_length_data.zero_()
else:
(
past_key_values,
past_key_values_data,
current_length_data,
) = initialize_past_key_values(self.base_model)
self.past_key_values = past_key_values
self.past_key_values_data = past_key_values_data
self.current_length_data = current_length_data
input_len = input_ids.shape[1]
reset_tree_mode(self)
outputs = self.base_model(input_ids, past_key_values=past_key_values, use_cache=True)
new_token = 0
for idx in range(max_steps):
input_id = outputs.logits[:, -1:].argmax(dim=-1)
outputs = self.base_model(input_id, use_cache=True, past_key_values=past_key_values)
input_ids = torch.cat([input_ids, input_id], dim=-1)
yield input_ids
if self.tokenizer.eos_token_id in input_ids[0, input_len:].tolist():
break
if new_token > 1024:
break
if input_ids.shape[1] > 1960:
break |