├── 20B_tokenizer.json ├── Blip2RWKV ├── README.md ├── blip2zh_qformer.py ├── configuration_blip2rwkv.py ├── configuration_rwkv.py ├── gen.png ├── modeling_blip2rwkv.py └── modeling_rwkv.py ├── LICENSE ├── Language └── README.md ├── README.md ├── Visual └── README.md ├── app.py ├── assets ├── MiniRWKV-4 Demo1.png ├── MiniRWKV-4 Demo2.png ├── MiniRWKV-4 Demo3.png ├── MiniRWKV-4 Demo4.png ├── MiniRWKV-4 Demo5.png ├── README.md ├── demo.jpg └── gen.png ├── config ├── README.md └── minirwkv4.yaml ├── minirwkv4 ├── README.md ├── blipcaption.py ├── blipvqa.py ├── gen.png ├── vitgptcaption.py └── vitvqa.py └── prompts ├── README.md ├── cardiogenic.yaml ├── exogenous.yaml └── operability.yaml /Blip2RWKV/README.md: -------------------------------------------------------------------------------- 1 | modeling_blip2rwkv.py 为主要代码，基于BLIP2，构建了RWKV的QFormer
2 | 3 | RWKV使用了https://huggingface.co/StarRing2022/RWKV-4-Raven-3B-v11-zh
4 | BERT和RWKV均使用中文
5 | 6 | 当前，更多是一个图像的RWKV语言编码器
7 | -------------------------------------------------------------------------------- /Blip2RWKV/blip2zh_qformer.py: -------------------------------------------------------------------------------- 1 | from transformers import AutoTokenizer, AutoConfig, AutoModel 2 | import torch 3 | from torch import nn 4 | import logging 5 | from torch.cuda.amp import autocast as autocast 6 | from lavis.common.registry import registry 7 | from lavis.common.dist_utils import get_rank 8 | 9 | from torch.nn import functional as F 10 | 11 | from lavis.models.base_model import all_gather_with_grad, concat_all_gather 12 | from lavis.models.blip2_models.blip2 import ( 13 | Blip2Base, 14 | compute_sim_matrix, 15 | disabled_train, 16 | ) 17 | from lavis.models.blip2_models.Qformer import BertLMHeadModel 18 | from lavis.models.blip_models.blip_outputs import BlipOutput, BlipOutputFeatures 19 | 20 | 21 | ZH_BERT = "./bert-base-chinese" 22 | 23 | 24 | class Blip2BaseZh(Blip2Base): 25 | @classmethod 26 | def init_tokenizer(cls): 27 | tokenizer = AutoTokenizer.from_pretrained(ZH_BERT) 28 | tokenizer.add_special_tokens({"bos_token": "[DEC]"}) 29 | return tokenizer 30 | 31 | @classmethod 32 | def init_Qformer(cls, num_query_token, vision_width, cross_attention_freq=2): 33 | encoder_config = AutoConfig.from_pretrained(ZH_BERT) 34 | encoder_config.encoder_width = vision_width 35 | # insert cross-attention layer every other block 36 | encoder_config.add_cross_attention = True 37 | encoder_config.cross_attention_freq = cross_attention_freq 38 | encoder_config.query_length = num_query_token 39 | # BertLMHeadModel 40 | Qformer = BertLMHeadModel.from_pretrained( 41 | ZH_BERT, config=encoder_config 42 | ) 43 | query_tokens = nn.Parameter( 44 | torch.zeros(1, num_query_token, encoder_config.hidden_size) 45 | ) 46 | query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range) 47 | return Qformer, query_tokens 48 | 49 | 50 | @registry.register_model("blip2zh") 51 | class Blip2ZhQformer(Blip2BaseZh): 52 | """ 53 | BLIP2 first-stage model with Q-former and ViT. 54 | Supported model types: 55 | - pretrained: pretrained model with vit-g 56 | - pretrain_vitL: pretrained model with vit-large 57 | - coco: fintuned model on coco 58 | Usage: 59 | >>> from lavis.models import load_model 60 | >>> model = load_model("blip2", "pretrain") 61 | """ 62 | 63 | def __init__( 64 | self, 65 | vit_model="eva_clip_g", 66 | img_size=224, 67 | drop_path_rate=0, 68 | use_grad_checkpoint=False, 69 | vit_precision="fp16", 70 | freeze_vit=True, 71 | num_query_token=32, 72 | cross_attention_freq=2, 73 | embed_dim=256, 74 | max_txt_len=32, 75 | ): 76 | super().__init__() 77 | 78 | self.tokenizer = self.init_tokenizer() 79 | 80 | self.visual_encoder, self.ln_vision = self.init_vision_encoder( 81 | vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision 82 | ) 83 | if freeze_vit: 84 | for name, param in self.visual_encoder.named_parameters(): 85 | param.requires_grad = False 86 | self.visual_encoder = self.visual_encoder.eval() 87 | self.visual_encoder.train = disabled_train 88 | logging.info("freeze vision encoder") 89 | self.Qformer, self.query_tokens = self.init_Qformer( 90 | num_query_token, self.visual_encoder.num_features, cross_attention_freq 91 | ) 92 | self.Qformer.resize_token_embeddings(len(self.tokenizer)) 93 | state_dict = self.Qformer.state_dict() 94 | for name, param in self.Qformer.named_parameters(): 95 | if "_query" in name: 96 | key_orig = name.replace("_query", "") 97 | param.data.copy_(state_dict[key_orig]) 98 | 99 | self.vision_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim) 100 | self.text_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim) 101 | 102 | self.itm_head = nn.Linear(self.Qformer.config.hidden_size, 2) 103 | 104 | self.temp = nn.Parameter(0.07 * torch.ones([])) 105 | 106 | self.max_txt_len = max_txt_len 107 | 108 | def forward(self, samples): 109 | image = samples["image"] 110 | text = samples["text_input"] 111 | 112 | with torch.no_grad(): 113 | image_embeds = self.ln_vision(self.visual_encoder(image)) 114 | image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to( 115 | image.device 116 | ) 117 | 118 | query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1) 119 | 120 | query_output = self.Qformer.bert( 121 | query_embeds=query_tokens, 122 | encoder_hidden_states=image_embeds, 123 | encoder_attention_mask=image_atts, 124 | use_cache=True, 125 | return_dict=True, 126 | ) 127 | 128 | image_feats = F.normalize( 129 | self.vision_proj(query_output.last_hidden_state), dim=-1 130 | ) 131 | 132 | text_tokens = self.tokenizer( 133 | text, 134 | padding="max_length", 135 | truncation=True, 136 | max_length=self.max_txt_len, 137 | return_tensors="pt", 138 | ).to(image.device) 139 | text_output = self.Qformer.bert( 140 | text_tokens.input_ids, 141 | attention_mask=text_tokens.attention_mask, 142 | return_dict=True, 143 | ) 144 | text_feat = F.normalize( 145 | self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1 146 | ) 147 | 148 | ###============== Image-text Contrastive ===================### 149 | image_feats_all = concat_all_gather( 150 | image_feats 151 | ) # [batch_size*num_gpu, num_query_tokens, embed_dim] 152 | text_feat_all = concat_all_gather(text_feat) # [batch_size*num_gpu, embed_dim] 153 | 154 | sim_q2t = torch.matmul( 155 | image_feats.unsqueeze(1), text_feat_all.unsqueeze(-1) 156 | ).squeeze() 157 | # [batch_size, batch_size*num_gpu, num_query_tokens] 158 | 159 | # image-text similarity: aggregate across all query tokens 160 | sim_i2t, _ = sim_q2t.max(-1) 161 | sim_i2t = sim_i2t / self.temp 162 | 163 | # text-query similarity: [batch_size, batch_size*num_gpu, num_query_tokens] 164 | sim_t2q = torch.matmul( 165 | text_feat.unsqueeze(1).unsqueeze(1), image_feats_all.permute(0, 2, 1) 166 | ).squeeze() 167 | 168 | # text-image similarity: aggregate across all query tokens 169 | sim_t2i, _ = sim_t2q.max(-1) 170 | sim_t2i = sim_t2i / self.temp # [batch_size, batch_size*num_gpu] 171 | 172 | rank = get_rank() 173 | bs = image.size(0) 174 | targets = torch.linspace(rank * bs, rank * bs + bs - 1, bs, dtype=int).to( 175 | image.device 176 | ) 177 | 178 | loss_itc = ( 179 | F.cross_entropy(sim_i2t, targets, label_smoothing=0.1) 180 | + F.cross_entropy(sim_t2i, targets, label_smoothing=0.1) 181 | ) / 2 182 | 183 | ###============== Image-text Matching ===================### 184 | text_input_ids_world = concat_all_gather(text_tokens.input_ids) 185 | text_attention_mask_world = concat_all_gather(text_tokens.attention_mask) 186 | image_embeds_world = all_gather_with_grad(image_embeds) 187 | with torch.no_grad(): 188 | weights_t2i = F.softmax(sim_t2i, dim=1) + 1e-4 189 | weights_t2i[:, rank * bs : rank * bs + bs].fill_diagonal_(0) 190 | weights_i2t = F.softmax(sim_i2t, dim=1) + 1e-4 191 | weights_i2t[:, rank * bs : rank * bs + bs].fill_diagonal_(0) 192 | 193 | # select a negative image for each text 194 | image_embeds_neg = [] 195 | for b in range(bs): 196 | neg_idx = torch.multinomial(weights_t2i[b], 1).item() 197 | image_embeds_neg.append(image_embeds_world[neg_idx]) 198 | image_embeds_neg = torch.stack(image_embeds_neg, dim=0) 199 | 200 | # select a negative text for each image 201 | text_ids_neg = [] 202 | text_atts_neg = [] 203 | for b in range(bs): 204 | neg_idx = torch.multinomial(weights_i2t[b], 1).item() 205 | text_ids_neg.append(text_input_ids_world[neg_idx]) 206 | text_atts_neg.append(text_attention_mask_world[neg_idx]) 207 | 208 | text_ids_neg = torch.stack(text_ids_neg, dim=0) 209 | text_atts_neg = torch.stack(text_atts_neg, dim=0) 210 | 211 | text_ids_all = torch.cat( 212 | [text_tokens.input_ids, text_tokens.input_ids, text_ids_neg], dim=0 213 | ) # pos, pos, neg 214 | text_atts_all = torch.cat( 215 | [text_tokens.attention_mask, text_tokens.attention_mask, text_atts_neg], 216 | dim=0, 217 | ) 218 | 219 | query_tokens_itm = self.query_tokens.expand(text_ids_all.shape[0], -1, -1) 220 | query_atts_itm = torch.ones(query_tokens_itm.size()[:-1], dtype=torch.long).to( 221 | image.device 222 | ) 223 | attention_mask_all = torch.cat([query_atts_itm, text_atts_all], dim=1) 224 | 225 | image_embeds_all = torch.cat( 226 | [image_embeds, image_embeds_neg, image_embeds], dim=0 227 | ) # pos, neg, pos 228 | image_atts_all = torch.ones(image_embeds_all.size()[:-1], dtype=torch.long).to( 229 | image.device 230 | ) 231 | 232 | output_itm = self.Qformer.bert( 233 | text_ids_all, 234 | query_embeds=query_tokens_itm, 235 | attention_mask=attention_mask_all, 236 | encoder_hidden_states=image_embeds_all, 237 | encoder_attention_mask=image_atts_all, 238 | return_dict=True, 239 | ) 240 | 241 | vl_embeddings = output_itm.last_hidden_state[:, : query_tokens_itm.size(1), :] 242 | vl_output = self.itm_head(vl_embeddings) 243 | logits = vl_output.mean(dim=1) 244 | 245 | itm_labels = torch.cat( 246 | [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)], 247 | dim=0, 248 | ).to(image.device) 249 | loss_itm = F.cross_entropy(logits, itm_labels) 250 | 251 | ##================= Image Captioning ========================## 252 | decoder_input_ids = text_tokens.input_ids.clone() 253 | decoder_input_ids[:, 0] = self.tokenizer.bos_token_id 254 | labels = decoder_input_ids.masked_fill( 255 | decoder_input_ids == self.tokenizer.pad_token_id, -100 256 | ) 257 | 258 | query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to( 259 | image.device 260 | ) 261 | attention_mask = torch.cat([query_atts, text_tokens.attention_mask], dim=1) 262 | lm_output = self.Qformer( 263 | decoder_input_ids, 264 | attention_mask=attention_mask, 265 | past_key_values=query_output.past_key_values, 266 | return_dict=True, 267 | labels=labels, 268 | ) 269 | 270 | loss_lm = lm_output.loss 271 | 272 | return BlipOutput( 273 | loss=loss_itc + loss_itm + loss_lm, 274 | loss_itc=loss_itc, 275 | loss_itm=loss_itm, 276 | loss_lm=loss_lm, 277 | ) 278 | 279 | @torch.no_grad() 280 | def generate( 281 | self, 282 | samples, 283 | use_nucleus_sampling=False, 284 | num_beams=3, 285 | max_length=30, 286 | min_length=10, 287 | top_p=0.9, 288 | repetition_penalty=1.0, 289 | ): 290 | """ 291 | Args: 292 | samples (dict): A dictionary containing the following keys: 293 | - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W) 294 | use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling. 295 | num_beams (int): Number of beams for beam search. 1 means no beam search. 296 | max_length (int): The maximum length of the sequence to be generated. 297 | min_length (int): The minimum length of the sequence to be generated. 298 | top_p (float): The cumulative probability for nucleus sampling. 299 | repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty. 300 | num_captions (int): Number of captions to be generated for each image. 301 | Returns: 302 | captions (list): A list of strings of length batch_size * num_captions. 303 | """ 304 | image = samples["image"] 305 | image_embeds = self.ln_vision(self.visual_encoder(image)) 306 | 307 | if not use_nucleus_sampling: 308 | image_embeds = image_embeds.repeat_interleave(num_beams, dim=0) 309 | else: 310 | num_beams = 1 311 | image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to( 312 | image.device 313 | ) 314 | 315 | model_kwargs = { 316 | "encoder_hidden_states": image_embeds, 317 | "encoder_attention_mask": image_atts, 318 | } 319 | 320 | input_ids = ( 321 | torch.LongTensor(image.size(0), 1) 322 | .fill_(self.tokenizer.bos_token_id) 323 | .to(image.device) 324 | ) 325 | query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1) 326 | 327 | outputs = self.Qformer.generate( 328 | input_ids=input_ids, 329 | query_embeds=query_tokens, 330 | max_length=max_length, 331 | min_length=min_length, 332 | num_beams=num_beams, 333 | do_sample=use_nucleus_sampling, 334 | top_p=top_p, 335 | eos_token_id=self.tokenizer.sep_token_id, 336 | pad_token_id=self.tokenizer.pad_token_id, 337 | **model_kwargs 338 | ) 339 | captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True) 340 | return captions 341 | 342 | def forward_image(self, image): 343 | image_embeds = self.ln_vision(self.visual_encoder(image)) 344 | image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to( 345 | image.device 346 | ) 347 | 348 | query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1) 349 | 350 | query_output = self.Qformer.bert( 351 | query_embeds=query_tokens, 352 | encoder_hidden_states=image_embeds, 353 | encoder_attention_mask=image_atts, 354 | return_dict=True, 355 | ) 356 | return query_output.last_hidden_state, image_embeds 357 | 358 | def forward_text(self, text_tokens): 359 | text_output = self.Qformer.bert( 360 | text_tokens.input_ids, 361 | attention_mask=text_tokens.attention_mask, 362 | return_dict=True, 363 | ) 364 | return text_output.last_hidden_state[:, 0, :] 365 | 366 | def compute_itm(self, image_inputs, text_ids, text_atts): 367 | image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to( 368 | image_inputs.device 369 | ) 370 | query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1) 371 | query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to( 372 | image_inputs.device 373 | ) 374 | attention_mask = torch.cat([query_atts, text_atts], dim=1) 375 | output_itm = self.Qformer.bert( 376 | text_ids, 377 | query_embeds=query_tokens, 378 | attention_mask=attention_mask, 379 | encoder_hidden_states=image_inputs, 380 | encoder_attention_mask=image_atts, 381 | return_dict=True, 382 | ) 383 | vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :] 384 | itm_logit = self.itm_head(vl_embeddings) 385 | itm_logit = itm_logit[:, :, 1].mean(dim=1) 386 | return itm_logit 387 | 388 | @torch.no_grad() 389 | def extract_features(self, samples, mode="multimodal"): 390 | """ 391 | Extract features for multimodal or unimodal samples. 392 | Args: 393 | samples (dict): A dictionary of samples, containing the following keys: 394 | - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image. 395 | Raw images should be preprocessed before being passed to feature extractor. 396 | - text_input (list): A list of strings containing the text, length B. 397 | mode (str): The mode of feature extraction. Can be either "multimodal", "text" or "image". 398 | If "multimodal", return image features and multimodal features; 399 | if "text", return text features; 400 | if "image", return image features. 401 | Default: "multimodal". 402 | Returns: 403 | BlipOutputFeatures: A BlipOutputFeatures object containing the features. 404 | See lavis/models/blip_models/blip_outputs.py for more details. 405 | """ 406 | image = samples.get("image") 407 | caption = samples.get("text_input") 408 | 409 | # assert mode is one of "image", "text", "multimodal" 410 | assert mode in [ 411 | "image", 412 | "text", 413 | "multimodal", 414 | ], "mode must be one of 'image', 'text', 'multimodal'" 415 | 416 | # initalize output 417 | image_embeds, text_embeds, multimodal_embeds = None, None, None 418 | image_features, text_features = None, None 419 | 420 | if mode == "image": 421 | assert ( 422 | image is not None 423 | ), "Image is not provided for mode 'image' or 'multimodal'" 424 | # return query features 425 | with self.maybe_autocast(): 426 | image_embeds_frozen = self.ln_vision(self.visual_encoder(image)) 427 | image_embeds_frozen = image_embeds_frozen.float() 428 | image_atts = torch.ones( 429 | image_embeds_frozen.size()[:-1], dtype=torch.long 430 | ).to(self.device) 431 | query_tokens = self.query_tokens.expand( 432 | image_embeds_frozen.shape[0], -1, -1 433 | ) 434 | 435 | query_output = self.Qformer.bert( 436 | query_embeds=query_tokens, 437 | encoder_hidden_states=image_embeds_frozen, 438 | encoder_attention_mask=image_atts, 439 | return_dict=True, 440 | ) 441 | image_embeds = query_output.last_hidden_state 442 | image_features = F.normalize(self.vision_proj(image_embeds), dim=-1) 443 | 444 | elif mode == "text": 445 | assert ( 446 | caption is not None 447 | ), "text input is None for mode 'text' or 'multimodal'" 448 | 449 | # return text features 450 | text = self.tokenizer(caption, return_tensors="pt", padding=True).to( 451 | self.device 452 | ) 453 | 454 | text_output = self.Qformer.bert( 455 | text.input_ids, 456 | attention_mask=text.attention_mask, 457 | return_dict=True, 458 | ) 459 | text_embeds = text_output.last_hidden_state 460 | text_features = self.text_proj(text_embeds) 461 | text_features = F.normalize(text_features, dim=-1) 462 | 463 | elif mode == "multimodal": 464 | # return multimodel query features 465 | with self.maybe_autocast(): 466 | image_embeds_frozen = self.ln_vision(self.visual_encoder(image)) 467 | image_embeds_frozen = image_embeds_frozen.float() 468 | image_atts = torch.ones( 469 | image_embeds_frozen.size()[:-1], dtype=torch.long 470 | ).to(self.device) 471 | query_tokens = self.query_tokens.expand( 472 | image_embeds_frozen.shape[0], -1, -1 473 | ) 474 | query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to( 475 | self.device 476 | ) 477 | 478 | text = self.tokenizer(caption, return_tensors="pt", padding=True).to( 479 | self.device 480 | ) 481 | attention_mask = torch.cat([query_atts, text.attention_mask], dim=1) 482 | 483 | output = self.Qformer.bert( 484 | text.input_ids, 485 | query_embeds=query_tokens, 486 | attention_mask=attention_mask, 487 | encoder_hidden_states=image_embeds_frozen, 488 | encoder_attention_mask=image_atts, 489 | return_dict=True, 490 | ) 491 | 492 | multimodal_embeds = output.last_hidden_state[:, : query_tokens.size(1), :] 493 | 494 | return BlipOutputFeatures( 495 | image_embeds=image_embeds, 496 | image_embeds_proj=image_features, 497 | text_embeds=text_embeds, 498 | text_embeds_proj=text_features, 499 | multimodal_embeds=multimodal_embeds, 500 | ) 501 | 502 | @classmethod 503 | def from_config(cls, cfg): 504 | vit_model = cfg.get("vit_model", "eva_clip_g") 505 | img_size = cfg.get("image_size") 506 | num_query_token = cfg.get("num_query_token") 507 | cross_attention_freq = cfg.get("cross_attention_freq", 2) 508 | 509 | drop_path_rate = cfg.get("drop_path_rate", 0) 510 | use_grad_checkpoint = cfg.get("use_grad_checkpoint", False) 511 | vit_precision = cfg.get("vit_precision", "fp16") 512 | freeze_vit = cfg.get("freeze_vit", True) 513 | 514 | max_txt_len = cfg.get("max_txt_len", 32) 515 | 516 | model = cls( 517 | vit_model=vit_model, 518 | img_size=img_size, 519 | drop_path_rate=drop_path_rate, 520 | use_grad_checkpoint=use_grad_checkpoint, 521 | vit_precision=vit_precision, 522 | freeze_vit=freeze_vit, 523 | num_query_token=num_query_token, 524 | cross_attention_freq=cross_attention_freq, 525 | max_txt_len=max_txt_len, 526 | ) 527 | 528 | load_finetuned = cfg.get("load_finetuned", True) 529 | load_pretrained = cfg.get("load_pretrained", True) 530 | if load_finetuned or load_pretrained: 531 | if load_pretrained: 532 | logging.info("Load pretrained from {}".format(cfg["pretrained"])) 533 | else: 534 | logging.info("Load finetuned from {}".format(cfg["fintuned"])) 535 | model.load_checkpoint_from_config(cfg) 536 | else: 537 | logging.info("Learning from scratch") 538 | 539 | return model 540 | 541 | def compute_sim_matrix(self, data_loader, task_cfg): 542 | """ 543 | Compute similarity i2t, t2i matrix for the given data loader. 544 | """ 545 | k_test = task_cfg.k_test 546 | 547 | return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test) 548 | 549 | 550 | if __name__ == "__main__": 551 | blip2baseZh = Blip2BaseZh() 552 | 553 | tokenizer = blip2baseZh.init_tokenizer() 554 | #print(tokenizer) 555 | 556 | Qformer, query_tokens = blip2baseZh.init_Qformer(num_query_token=1024,vision_width=768) 557 | #print(Qformer) 558 | #print(query_tokens) 559 | 560 | test_sentence = '你好，介绍下你自己.' 561 | #testinput = tokenizer(test_sentence)['input_ids'] 562 | #print(testinput) 563 | 564 | text = tokenizer(test_sentence, return_tensors="pt", padding=True) 565 | blip2Zhqformer = Blip2ZhQformer() 566 | print(blip2Zhqformer.forward_text(text)) -------------------------------------------------------------------------------- /Blip2RWKV/configuration_blip2rwkv.py: -------------------------------------------------------------------------------- 1 | from transformers import ( 2 | PretrainedConfig, 3 | Blip2VisionConfig, Blip2QFormerConfig 4 | ) 5 | from configuration_rwkv import RwkvConfig 6 | 7 | import copy 8 | from transformers.configuration_utils import PretrainedConfig 9 | from transformers.utils import logging 10 | 11 | logger = logging.get_logger(__name__) 12 | 13 | 14 | 15 | class Blip2RWKVConfig(PretrainedConfig): 16 | """Mainly based on Blip2Config 17 | 18 | Args: 19 | PretrainedConfig (_type_): _description_ 20 | """ 21 | is_composition = True 22 | 23 | def __init__(self, vision_config=None, qformer_config=None, text_config=None, num_query_tokens=27, **kwargs): 24 | super().__init__(**kwargs) 25 | 26 | if vision_config is None: 27 | vision_config = {} 28 | logger.info("vision_config is None. initializing the Blip2VisionConfig with default values.") 29 | 30 | if qformer_config is None: 31 | qformer_config = {} 32 | logger.info("qformer_config is None. Initializing the Blip2QFormerConfig with default values.") 33 | 34 | if text_config is None: 35 | text_config = {} 36 | logger.info("text_config is None. Initializing the text config with default values (`OPTConfig`).") 37 | 38 | self.vision_config = Blip2VisionConfig(**vision_config) 39 | self.qformer_config = Blip2QFormerConfig(**qformer_config) 40 | # text_model_type = text_config["model_type"] if "model_type" in text_config else "opt" 41 | # self.text_config = CONFIG_MAPPING[text_model_type](**text_config) 42 | self.text_config = RwkvConfig(**text_config) 43 | 44 | # self.tie_word_embeddings = self.text_config.tie_word_embeddings 45 | self.tie_word_embeddings = False # I don't know what this is 46 | # self.is_encoder_decoder = self.text_config.is_encoder_decoder 47 | self.is_encoder_decoder = True # chatglm is an encoder-decoder model 48 | 49 | self.num_query_tokens = num_query_tokens 50 | self.qformer_config.encoder_hidden_size = self.vision_config.hidden_size 51 | # self.use_decoder_only_language_model = self.text_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES 52 | self.use_decoder_only_language_model = True # chatglm has no encoder 53 | self.initializer_factor = 1.0 54 | self.initializer_range = 0.02 55 | 56 | @classmethod 57 | def from_vision_qformer_text_configs( 58 | cls, 59 | vision_config: Blip2VisionConfig, 60 | qformer_config: Blip2QFormerConfig, 61 | text_config: PretrainedConfig, 62 | **kwargs, 63 | ): 64 | r""" 65 | Instantiate a [`Blip2Config`] (or a derived class) from a BLIP-2 vision model, Q-Former and language model 66 | configurations. 67 | 68 | Returns: 69 | [`Blip2Config`]: An instance of a configuration object 70 | """ 71 | 72 | return cls( 73 | vision_config=vision_config.to_dict(), 74 | qformer_config=qformer_config.to_dict(), 75 | text_config=text_config.to_dict(), 76 | **kwargs, 77 | ) 78 | 79 | def to_dict(self): 80 | """ 81 | Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`]. 82 | 83 | Returns: 84 | `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance, 85 | """ 86 | output = copy.deepcopy(self.__dict__) 87 | output["vision_config"] = self.vision_config.to_dict() 88 | output["qformer_config"] = self.qformer_config.to_dict() 89 | output["text_config"] = self.text_config.to_dict() 90 | output["model_type"] = self.__class__.model_type 91 | return output 92 | 93 | if __name__ == "__main__": 94 | blip2rwkvconfig = Blip2RWKVConfig() 95 | print(blip2rwkvconfig) 96 | -------------------------------------------------------------------------------- /Blip2RWKV/configuration_rwkv.py: -------------------------------------------------------------------------------- 1 | from transformers import PretrainedConfig,RwkvConfig 2 | 3 | 4 | class RwkvConfig(PretrainedConfig): 5 | """ 6 | This is the configuration class to store the configuration of a [`RwkvModel`]. It is used to instantiate a RWKV 7 | model according to the specified arguments, defining the model architecture. Instantiating a configuration with the 8 | defaults will yield a similar configuration to that of the RWVK-4 9 | [RWKV/rwkv-4-169m-pile](https://huggingface.co/RWKV/rwkv-4-169m-pile) architecture. 10 | 11 | Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the 12 | documentation from [`PretrainedConfig`] for more information. 13 | 14 | 15 | Args: 16 | vocab_size (`int`, *optional*, defaults to 50277): 17 | Vocabulary size of the RWKV model. Defines the number of different tokens that can be represented by the 18 | `inputs_ids` passed when calling [`RwkvModel`]. 19 | context_length (`int`, *optional*, defaults to 1024): 20 | The maximum sequence length that this model can be be used with in a single forward (using it in RNN mode 21 | lets use any sequence length). 22 | hidden_size (`int`, *optional*, defaults to 4096): 23 | Dimensionality of the embeddings and hidden states. 24 | num_hidden_layers (`int`, *optional*, defaults to 32): 25 | Number of hidden layers in the model. 26 | attention_hidden_size (`int`, *optional*): 27 | Dimensionality of the attention hidden states. Will default to `hidden_size` if unset. 28 | intermediate_size (`int`, *optional*): 29 | Dimensionality of the inner feed-forward layers. Will default to 4 times `hidden_size` if unset. 30 | layer_norm_eps (`float`, *optional*, defaults to 1e-5): 31 | The epsilon to use in the layer normalization layers. 32 | bos_token_id (`int`, *optional*, defaults to 0): 33 | The id of the beginning of sentence token in the vocabulary. Defaults to 0 as RWKV uses the same tokenizer 34 | as GPTNeoX. 35 | eos_token_id (`int`, *optional*, defaults to 0): 36 | The id of the end of sentence token in the vocabulary. Defaults to 0 as RWKV uses the same tokenizer as 37 | GPTNeoX. 38 | rescale_every (`int`, *optional*, default to 6): 39 | At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every 40 | `rescale_every` layer. If set to 0 or a negative number, no rescale is done. 41 | tie_word_embeddings (`bool`, *optional*, defaults to `False`): 42 | Whether or not to tie the word embeddings with the input token embeddings. 43 | use_cache (`bool`, *optional*, defaults to `True`): 44 | Whether or not the model should return the last state. 45 | 46 | 47 | Example: 48 | 49 | ```python 50 | >>> from transformers import RwkvConfig, RwkvModel 51 | 52 | >>> # Initializing a Rwkv configuration 53 | >>> configuration = RwkvConfig() 54 | 55 | >>> # Initializing a model (with random weights) from the configuration 56 | >>> model = RwkvModel(configuration) 57 | 58 | >>> # Accessing the model configuration 59 | >>> configuration = model.config 60 | ```""" 61 | 62 | model_type = "rwkv" 63 | attribute_map = {"max_position_embeddings": "context_length"} 64 | 65 | def __init__( 66 | self, 67 | vocab_size=50277, 68 | context_length=1024, 69 | hidden_size=2560, 70 | num_hidden_layers=32, 71 | attention_hidden_size=2560, 72 | intermediate_size=10240, 73 | layer_norm_epsilon=1e-5, 74 | bos_token_id=0, 75 | eos_token_id=0, 76 | rescale_every=6, 77 | tie_word_embeddings=False, 78 | use_cache=True, 79 | **kwargs, 80 | ): 81 | self.vocab_size = vocab_size 82 | self.context_length = context_length 83 | self.hidden_size = hidden_size 84 | self.num_hidden_layers = num_hidden_layers 85 | self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size 86 | self.intermediate_size = intermediate_size if intermediate_size is not None else 4 * hidden_size 87 | self.layer_norm_epsilon = layer_norm_epsilon 88 | self.rescale_every = rescale_every 89 | self.use_cache = use_cache 90 | 91 | self.bos_token_id = bos_token_id 92 | self.eos_token_id = eos_token_id 93 | 94 | super().__init__( 95 | tie_word_embeddings=tie_word_embeddings, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs 96 | ) 97 | 98 | if __name__ == "__main__": 99 | rwkvconfig = RwkvConfig() 100 | print(rwkvconfig) -------------------------------------------------------------------------------- /Blip2RWKV/gen.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/Blip2RWKV/gen.png -------------------------------------------------------------------------------- /Blip2RWKV/modeling_blip2rwkv.py: -------------------------------------------------------------------------------- 1 | import copy 2 | import os 3 | from typing import Callable, List, Optional, Tuple, Union 4 | import numpy as np 5 | import torch 6 | from torch.nn import CrossEntropyLoss 7 | from torch.nn.utils.rnn import pad_sequence 8 | import warnings 9 | from torch import Tensor, nn 10 | 11 | from transformers import ( 12 | PreTrainedModel, 13 | PreTrainedTokenizer, 14 | Blip2VisionModel, 15 | Blip2QFormerModel, 16 | Blip2Model, 17 | Blip2PreTrainedModel, 18 | Blip2ForConditionalGeneration, 19 | GenerationConfig, 20 | ) 21 | from transformers.models.blip_2.modeling_blip_2 import ( 22 | Blip2ForConditionalGenerationModelOutput, 23 | ) 24 | from transformers.utils import logging 25 | from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList 26 | 27 | from modeling_rwkv import ( 28 | RwkvForCausalLM 29 | ) 30 | from configuration_blip2rwkv import Blip2RWKVConfig 31 | 32 | 33 | logger = logging.get_logger(__name__) 34 | 35 | 36 | class Blip2RWKVConditionalGeneration(Blip2ForConditionalGeneration): 37 | config_class = Blip2RWKVConfig 38 | 39 | def __init__(self, config: Blip2RWKVConfig): 40 | Blip2PreTrainedModel.__init__(self, config) 41 | # NOTE: we only initialize Blip2PreTrainedModel 42 | # directly call super().__init__() will cause error since ChatGLM cannot be found by AutoModel 43 | 44 | self.vision_model = Blip2VisionModel(config.vision_config).to("cuda") 45 | 46 | self.query_tokens = nn.Parameter( 47 | torch.zeros(1, config.num_query_tokens, config.qformer_config.hidden_size).to("cuda") 48 | ) 49 | self.qformer = Blip2QFormerModel(config.qformer_config).to("cuda") 50 | 51 | self.language_projection = nn.Linear( 52 | config.qformer_config.hidden_size, config.text_config.hidden_size 53 | ).to("cuda") 54 | #self.language_model = RwkvForCausalLM(config.text_config) 55 | self.language_model = RwkvForCausalLM.from_pretrained("RWKV-4-Raven-3B-v11-zh",device_map='auto').to("cuda") 56 | #print(self.language_model ) 57 | 58 | # Initialize weights and apply final processing 59 | # self.post_init() 60 | 61 | def setup_dtype(self, vision_encoder_dtype: str = "fp32", lm_dtype: str = "fp16"): 62 | if vision_encoder_dtype == "fp32": 63 | self.vision_model = self.vision_model.float().cuda() 64 | elif vision_encoder_dtype == "fp16": 65 | self.vision_model = self.vision_model.half().cuda() 66 | else: 67 | raise NotImplementedError( 68 | f"Unsupported vision_encoder_dtype: {vision_encoder_dtype}" 69 | ) 70 | 71 | if lm_dtype == "fp32": 72 | self.language_model = self.language_model.float() 73 | elif lm_dtype == "fp16": 74 | self.language_model = self.language_model.half() 75 | elif lm_dtype == "int4": 76 | self.language_model = self.language_model.half().quantize(4) 77 | elif lm_dtype == "int8": 78 | self.language_model = self.language_model.half().quantize(8) 79 | else: 80 | raise NotImplementedError(f"Unsupported lm_dtype: {lm_dtype}") 81 | 82 | def forward( 83 | self, 84 | pixel_values: torch.FloatTensor, 85 | input_ids: torch.FloatTensor, 86 | image_slot_offset: Optional[torch.LongTensor] = None, 87 | attention_mask: Optional[torch.LongTensor] = None, 88 | output_attentions: Optional[bool] = None, 89 | output_hidden_states: Optional[bool] = None, 90 | labels: Optional[torch.LongTensor] = None, 91 | return_dict: Optional[bool] = None, 92 | ) -> Union[Tuple, Blip2ForConditionalGenerationModelOutput]: 93 | """_summary_ 94 | 95 | Args: 96 | pixel_values (torch.FloatTensor): _description_ 97 | input_ids (torch.FloatTensor): input_ids[:, :num_query_tokens] should be filled with tokenizer.unk_token_id 98 | image_slot_offset (Optional[torch.LongTensor], optional): if not set, all vtokens are placed as prefix (image_slot_offset = torch.zeros(bsz)). Defaults to None. 99 | attention_mask (Optional[torch.LongTensor], optional): _description_. Defaults to None. 100 | output_attentions (Optional[bool], optional): _description_. Defaults to None. 101 | output_hidden_states (Optional[bool], optional): _description_. Defaults to None. 102 | labels (Optional[torch.LongTensor], optional): _description_. Defaults to None. 103 | return_dict (Optional[bool], optional): _description_. Defaults to None. 104 | 105 | Returns: 106 | Union[Tuple, Blip2ForConditionalGenerationModelOutput]: _description_ 107 | """ 108 | return_dict = ( 109 | return_dict if return_dict is not None else self.config.use_return_dict 110 | ) 111 | 112 | # step 1: forward the images through the vision encoder, 113 | # to get image embeddings of shape (batch_size, seq_len, hidden_size) 114 | vision_outputs = self.vision_model( 115 | pixel_values=pixel_values, 116 | output_attentions=output_attentions, 117 | output_hidden_states=output_hidden_states, 118 | return_dict=return_dict, 119 | ) 120 | image_embeds = vision_outputs[0] 121 | 122 | # step 2: forward the query tokens through the QFormer, using the image embeddings for cross-attention 123 | image_attention_mask = torch.ones( 124 | image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device 125 | ) 126 | 127 | query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1) 128 | query_outputs = self.qformer( 129 | query_embeds=query_tokens, 130 | encoder_hidden_states=image_embeds, 131 | encoder_attention_mask=image_attention_mask, 132 | output_attentions=output_attentions, 133 | output_hidden_states=output_hidden_states, 134 | return_dict=return_dict, 135 | ) 136 | query_output = query_outputs[0] 137 | 138 | # step 3: use the language model, conditioned on the query outputs and the prompt 139 | # 关键步骤，将图片进行embedding编码，然后送入LM 140 | language_model_inputs = self.language_projection(query_output) 141 | inputs_embeds = self.language_model.get_input_embeddings()(input_ids) 142 | #print(inputs_embeds.shape) #[1,27,2560] 143 | if image_slot_offset is None: 144 | # image as prefix 145 | # update data to avoid inplace operation of leaf Variable 146 | inputs_embeds.data[ 147 | :, : self.config.num_query_tokens, : # num_query_tokens = 27 148 | ] = language_model_inputs 149 | else: 150 | for i, offset in enumerate(image_slot_offset): 151 | inputs_embeds.data[ 152 | i, offset : offset + self.config.num_query_tokens, : 153 | ] = language_model_inputs[i] 154 | 155 | outputs = self.language_model( 156 | input_ids=input_ids, 157 | inputs_embeds=inputs_embeds, 158 | attention_mask=attention_mask, 159 | output_attentions=output_attentions, 160 | output_hidden_states=output_hidden_states, 161 | return_dict=return_dict, 162 | ) 163 | logits = outputs.logits if return_dict else outputs[0] 164 | loss = None 165 | # we compute the loss here since we need to take into account the sequence length of the query embeds 166 | if labels is not None: 167 | logits = logits[:, -labels.size(1) :, :] 168 | # Shift so that tokens < n predict n 169 | shift_logits = logits[..., :-1, :].contiguous() 170 | shift_labels = labels[..., 1:].contiguous().to(logits.device) 171 | 172 | # Flatten the tokens 173 | loss_fct = CrossEntropyLoss(reduction="mean") 174 | 175 | loss = loss_fct( 176 | shift_logits.view(-1, self.config.text_config.vocab_size), 177 | shift_labels.view(-1), 178 | ) 179 | 180 | if not return_dict: 181 | output = (logits, vision_outputs, query_outputs, outputs) 182 | return ((loss,) + output) if loss is not None else output 183 | 184 | return Blip2ForConditionalGenerationModelOutput( 185 | loss=loss, 186 | logits=logits, 187 | vision_outputs=vision_outputs, 188 | qformer_outputs=query_outputs, 189 | language_model_outputs=outputs, 190 | ) 191 | 192 | if __name__ == "__main__": 193 | #Blip2RWKV测试 194 | blip2rwkvconfig = Blip2RWKVConfig() 195 | blip2RWKVConditionalGeneration = Blip2RWKVConditionalGeneration(config=blip2rwkvconfig) 196 | blip2RWKVConditionalGeneration.setup_dtype() 197 | 198 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 199 | #device = "cpu" 200 | 201 | from PIL import Image 202 | from transformers import BlipProcessor 203 | from lavis.models import load_model_and_preprocess 204 | 205 | raw_image = Image.open('gen.png').convert('RGB') 206 | caption = "一个男孩抱着一只猫，猫咪看起来很享受。" 207 | 208 | model, vis_processors, txt_processors = load_model_and_preprocess(name="blip2_feature_extractor", model_type="pretrain", is_eval=True, device=device) 209 | model = model.to(device) 210 | 211 | image = vis_processors["eval"](raw_image).unsqueeze(0).to(device) 212 | text_input = txt_processors["eval"](caption) 213 | sample = {"image": image, "text_input": [text_input]} 214 | 215 | #print(image.shape) 216 | #print(text_input) 217 | 218 | from transformers import GPTNeoXTokenizerFast 219 | tokenizer = GPTNeoXTokenizerFast.from_pretrained("RWKV-4-Raven-3B-v11-zh") 220 | text_input = tokenizer.encode(text_input, return_tensors='pt') 221 | blip2rwkvoutput = blip2RWKVConditionalGeneration.forward(pixel_values=image.to(device),input_ids=text_input.to(device),labels=text_input.to(device)) 222 | #print(blip2rwkvoutput) -------------------------------------------------------------------------------- /Blip2RWKV/modeling_rwkv.py: -------------------------------------------------------------------------------- 1 | # coding=utf-8 2 | # Copyright 2023 Bo Peng and HuggingFace Inc. team. 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. 7 | # You may obtain a copy of the License at 8 | # 9 | # http://www.apache.org/licenses/LICENSE-2.0 10 | # 11 | # Unless required by applicable law or agreed to in writing, software 12 | # distributed under the License is distributed on an "AS IS" BASIS, 13 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 14 | # See the License for the specific language governing permissions and 15 | # limitations under the License. 16 | """PyTorch RWKV model.""" 17 | 18 | import math 19 | from dataclasses import dataclass 20 | from pathlib import Path 21 | from typing import List, Optional, Tuple, Union 22 | 23 | import torch 24 | import torch.utils.checkpoint 25 | from torch import nn 26 | from torch.nn import CrossEntropyLoss 27 | 28 | from transformers import PreTrainedModel 29 | from transformers.utils import ( 30 | ModelOutput, 31 | add_code_sample_docstrings, 32 | add_start_docstrings, 33 | add_start_docstrings_to_model_forward, 34 | is_ninja_available, 35 | is_torch_cuda_available, 36 | logging, 37 | ) 38 | from configuration_rwkv import RwkvConfig 39 | 40 | 41 | logger = logging.get_logger(__name__) 42 | 43 | _CHECKPOINT_FOR_DOC = "RWKV/rwkv-4-169m-pile" 44 | _CONFIG_FOR_DOC = "RwkvConfig" 45 | 46 | RWKV_PRETRAINED_MODEL_ARCHIVE_LIST = [ 47 | "RWKV/rwkv-4-169m-pile", 48 | "RWKV/rwkv-4-430m-pile", 49 | "RWKV/rwkv-4-1b5-pile", 50 | "RWKV/rwkv-4-3b-pile", 51 | "RWKV/rwkv-4-7b-pile", 52 | "RWKV/rwkv-4-14b-pile", 53 | "RWKV/rwkv-raven-1b5", 54 | "RWKV/rwkv-raven-3b", 55 | "RWKV/rwkv-raven-7b", 56 | "RWKV/rwkv-raven-14b", 57 | # See all RWKV models at https://huggingface.co/models?filter=rwkv 58 | ] 59 | 60 | 61 | rwkv_cuda_kernel = None 62 | 63 | 64 | def load_wkv_cuda_kernel(context_length): 65 | from torch.utils.cpp_extension import load as load_kernel 66 | 67 | global rwkv_cuda_kernel 68 | 69 | kernel_folder = Path(__file__).resolve().parent.parent.parent / "kernels" / "rwkv" 70 | cuda_kernel_files = [kernel_folder / f for f in ["wkv_op.cpp", "wkv_cuda.cu", "wkv_cuda_bf16.cu"]] 71 | 72 | # Only load the kernel if it's not been loaded yet or if we changed the context length 73 | if rwkv_cuda_kernel is not None and rwkv_cuda_kernel.max_seq_length == context_length: 74 | return 75 | 76 | logger.info(f"Loading CUDA kernel for RWKV at context length of {context_length}.") 77 | 78 | flags = [ 79 | "-res-usage", 80 | "--maxrregcount 60", 81 | "--use_fast_math", 82 | "-O3", 83 | "-Xptxas -O3", 84 | "--extra-device-vectorization", 85 | f"-DTmax={context_length}", 86 | ] 87 | rwkv_cuda_kernel = load_kernel( 88 | name=f"wkv_{context_length}", 89 | sources=cuda_kernel_files, 90 | verbose=(logging.get_verbosity() == logging.DEBUG), 91 | extra_cuda_cflags=flags, 92 | ) 93 | rwkv_cuda_kernel.max_seq_length = context_length 94 | 95 | 96 | class RwkvLinearAttention(torch.autograd.Function): 97 | @staticmethod 98 | def forward(ctx, time_decay, time_first, key, value, state=None, return_state=False): 99 | batch_size, seq_len, hidden_size = key.size() 100 | if seq_len > rwkv_cuda_kernel.max_seq_length: 101 | raise ValueError( 102 | f"Cannot process a batch with {seq_len} tokens at the same time, use a maximum of " 103 | f"{rwkv_cuda_kernel.max_seq_length} with this model." 104 | ) 105 | if batch_size * hidden_size % min(hidden_size, 32) != 0: 106 | raise ValueError( 107 | f"The product of batch size ({batch_size}) and hidden size ({hidden_size}) needs to be a round " 108 | f"multiple of {min(hidden_size, 32)}." 109 | ) 110 | 111 | ctx.input_dtype = key.dtype 112 | 113 | if ( 114 | time_decay.device.type != "cuda" 115 | or time_first.device.type != "cuda" 116 | or key.device.type != "cuda" 117 | or value.device.type != "cuda" 118 | ): 119 | raise ValueError("Calling the CUDA kernel for wkv attention requires all tensors to be on CUDA devices.") 120 | 121 | time_decay = -torch.exp(time_decay.float().contiguous()) 122 | if key.dtype == torch.float16: 123 | time_first = time_first.float() 124 | key = key.float() 125 | value = value.float() 126 | time_first = time_first.contiguous() 127 | key = key.contiguous() 128 | value = value.contiguous() 129 | # The CUDA kernel will fill this tensor. 130 | output = torch.empty_like(key, memory_format=torch.contiguous_format) 131 | if return_state or state is not None: 132 | if state is None: 133 | state = torch.zeros( 134 | batch_size, 135 | hidden_size, 136 | 3, 137 | dtype=torch.float32, 138 | device=key.device, 139 | memory_format=torch.contiguous_format, 140 | ) 141 | state[:, :, 2] -= 1e38 142 | else: 143 | state = torch.cat([s.unsqueeze(2) for s in state], dim=2).contiguous() 144 | if key.dtype == torch.bfloat16: 145 | forward_func = rwkv_cuda_kernel.forward_with_state_bf16 146 | else: 147 | forward_func = rwkv_cuda_kernel.forward_with_state 148 | forward_func(time_decay, time_first, key, value, output, state) 149 | else: 150 | forward_func = rwkv_cuda_kernel.forward_bf16 if key.dtype == torch.bfloat16 else rwkv_cuda_kernel.forward 151 | forward_func(time_decay, time_first, key, value, output) 152 | 153 | ctx.save_for_backward(time_decay, time_first, key, value, output) 154 | 155 | if state is not None: 156 | state = [s.squeeze(2) for s in torch.chunk(state, 3, dim=2)] 157 | 158 | return output.to(ctx.input_dtype), state 159 | 160 | @staticmethod 161 | # g stands for grad 162 | def backward(ctx, g_output): 163 | input_dtype = ctx.input_dtype 164 | 165 | time_decay, time_first, key, value, output = ctx.saved_tensors 166 | # The CUDA kernel will fill those tensors. 167 | g_time_decay = torch.empty_like( 168 | time_decay, 169 | memory_format=torch.contiguous_format, 170 | dtype=torch.bfloat16 if input_dtype == torch.bfloat16 else torch.float32, 171 | ) 172 | g_time_first = torch.empty_like(time_first, memory_format=torch.contiguous_format) 173 | g_key = torch.empty_like(key, memory_format=torch.contiguous_format) 174 | g_value = torch.empty_like(value, memory_format=torch.contiguous_format) 175 | 176 | if input_dtype == torch.float16: 177 | g_output = g_output.float() 178 | backward_func = rwkv_cuda_kernel.backward_bf16 if input_dtype == torch.bfloat16 else rwkv_cuda_kernel.backward 179 | backward_func( 180 | time_decay, 181 | time_first, 182 | key, 183 | value, 184 | output, 185 | g_output.contiguous(), 186 | g_time_decay, 187 | g_time_first, 188 | g_key, 189 | g_value, 190 | ) 191 | g_time_decay = torch.sum(g_time_decay, dim=0) 192 | g_time_first = torch.sum(g_time_first, dim=0) 193 | 194 | return ( 195 | None, 196 | None, 197 | None, 198 | g_time_decay.to(input_dtype), 199 | g_time_first.to(input_dtype), 200 | g_key.to(input_dtype), 201 | g_value.to(input_dtype), 202 | ) 203 | 204 | 205 | def rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=None, return_state=False): 206 | # For CPU fallback. Will be slower and probably take more memory than the custom CUDA kernel if not executed 207 | # within a torch.no_grad. 208 | _, seq_length, _ = key.size() 209 | output = torch.zeros_like(key) 210 | 211 | if state is None: 212 | num_state = torch.zeros_like(key[:, 0], dtype=torch.float32) 213 | den_state = torch.zeros_like(key[:, 0], dtype=torch.float32) 214 | max_state = torch.zeros_like(key[:, 0], dtype=torch.float32) - 1e38 215 | else: 216 | num_state, den_state, max_state = state 217 | # For numerical stability 218 | # real_numerator_state = num_state * torch.exp(max_state) 219 | # real_denominator_state = den_state * torch.exp(max_state) 220 | 221 | time_decay = -torch.exp(time_decay) 222 | 223 | for current_index in range(seq_length): 224 | current_key = key[:, current_index].float() 225 | current_value = value[:, current_index] 226 | 227 | # wkv computation at time t 228 | max_for_output = torch.maximum(max_state, current_key + time_first) 229 | e1 = torch.exp(max_state - max_for_output) 230 | e2 = torch.exp(current_key + time_first - max_for_output) 231 | numerator = e1 * num_state + e2 * current_value 232 | denominator = e1 * den_state + e2 233 | output[:, current_index] = (numerator / denominator).to(output.dtype) 234 | 235 | # Update state for next iteration 236 | max_for_state = torch.maximum(max_state + time_decay, current_key) 237 | e1 = torch.exp(max_state + time_decay - max_for_state) 238 | e2 = torch.exp(current_key - max_for_state) 239 | num_state = e1 * num_state + e2 * current_value 240 | den_state = e1 * den_state + e2 241 | max_state = max_for_state 242 | 243 | if return_state or state is not None: 244 | state = [num_state, den_state, max_state] 245 | 246 | return output, state 247 | 248 | 249 | def rwkv_linear_attention(time_decay, time_first, key, value, state=None, return_state=False): 250 | no_cuda = any(t.device.type != "cuda" for t in [time_decay, time_first, key, value]) 251 | # Launching the CUDA kernel for just one token will actually be slower (there is no for loop in the CPU version 252 | # in this case). 253 | one_token = key.size(1) == 1 254 | if rwkv_cuda_kernel is None or no_cuda or one_token: 255 | return rwkv_linear_attention_cpu(time_decay, time_first, key, value, state=state, return_state=return_state) 256 | else: 257 | return RwkvLinearAttention.apply(time_decay, time_first, key, value, state, return_state) 258 | 259 | 260 | class RwkvSelfAttention(nn.Module): 261 | def __init__(self, config, layer_id=0): 262 | super().__init__() 263 | self.config = config 264 | kernel_loaded = rwkv_cuda_kernel is not None and rwkv_cuda_kernel.max_seq_length == config.context_length 265 | if is_ninja_available() and is_torch_cuda_available() and not kernel_loaded: 266 | try: 267 | load_wkv_cuda_kernel(config.context_length) 268 | except Exception: 269 | logger.info("Could not load the custom CUDA kernel for RWKV attention.") 270 | self.layer_id = layer_id 271 | hidden_size = config.hidden_size 272 | attention_hidden_size = ( 273 | config.attention_hidden_size if config.attention_hidden_size is not None else hidden_size 274 | ) 275 | self.attention_hidden_size = attention_hidden_size 276 | 277 | self.time_decay = nn.Parameter(torch.empty(attention_hidden_size)) 278 | self.time_first = nn.Parameter(torch.empty(attention_hidden_size)) 279 | 280 | self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size)) 281 | self.time_mix_value = nn.Parameter(torch.empty(1, 1, hidden_size)) 282 | self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size)) 283 | 284 | self.time_shift = nn.ZeroPad2d((0, 0, 1, -1)) 285 | self.key = nn.Linear(hidden_size, attention_hidden_size, bias=False) 286 | self.value = nn.Linear(hidden_size, attention_hidden_size, bias=False) 287 | self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False) 288 | self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False) 289 | 290 | # TODO: maybe jit, otherwise move inside forward 291 | def extract_key_value(self, hidden, state=None): 292 | # Mix hidden with the previous timestep to produce key, value, receptance 293 | if hidden.size(1) == 1 and state is not None: 294 | shifted = state[1][:, :, self.layer_id] 295 | else: 296 | shifted = self.time_shift(hidden) 297 | if state is not None: 298 | shifted[:, 0] = state[1][:, :, self.layer_id] 299 | key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key) 300 | value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value) 301 | receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance) 302 | 303 | key = self.key(key) 304 | value = self.value(value) 305 | receptance = torch.sigmoid(self.receptance(receptance)) 306 | if state is not None: 307 | state[1][:, :, self.layer_id] = hidden[:, -1] 308 | return receptance, key, value, state 309 | 310 | def forward(self, hidden, state=None, use_cache=False): 311 | receptance, key, value, state = self.extract_key_value(hidden, state=state) 312 | layer_state = tuple(s[:, :, self.layer_id] for s in state[2:]) if state is not None else None 313 | rwkv, layer_state = rwkv_linear_attention( 314 | self.time_decay, 315 | self.time_first, 316 | key, 317 | value, 318 | state=layer_state, 319 | return_state=use_cache, 320 | ) 321 | 322 | if layer_state is not None: 323 | state[2][:, :, self.layer_id] = layer_state[0] 324 | state[3][:, :, self.layer_id] = layer_state[1] 325 | state[4][:, :, self.layer_id] = layer_state[2] 326 | 327 | return self.output(receptance * rwkv), state 328 | 329 | 330 | class RwkvFeedForward(nn.Module): 331 | def __init__(self, config, layer_id=0): 332 | super().__init__() 333 | self.config = config 334 | self.layer_id = layer_id 335 | hidden_size = config.hidden_size 336 | intermediate_size = ( 337 | config.intermediate_size if config.intermediate_size is not None else 4 * config.hidden_size 338 | ) 339 | 340 | self.time_shift = nn.ZeroPad2d((0, 0, 1, -1)) 341 | self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size)) 342 | self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size)) 343 | 344 | self.key = nn.Linear(hidden_size, intermediate_size, bias=False) 345 | self.receptance = nn.Linear(hidden_size, hidden_size, bias=False) 346 | self.value = nn.Linear(intermediate_size, hidden_size, bias=False) 347 | 348 | def forward(self, hidden, state=None): 349 | if hidden.size(1) == 1 and state is not None: 350 | shifted = state[0][:, :, self.layer_id] 351 | else: 352 | shifted = self.time_shift(hidden) 353 | if state is not None: 354 | shifted[:, 0] = state[0][:, :, self.layer_id] 355 | key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key) 356 | receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance) 357 | 358 | key = torch.square(torch.relu(self.key(key))) 359 | value = self.value(key) 360 | receptance = torch.sigmoid(self.receptance(receptance)) 361 | 362 | if state is not None: 363 | state[0][:, :, self.layer_id] = hidden[:, -1] 364 | 365 | return receptance * value, state 366 | 367 | 368 | class RwkvBlock(nn.Module): 369 | def __init__(self, config, layer_id): 370 | super().__init__() 371 | self.config = config 372 | self.layer_id = layer_id 373 | 374 | if layer_id == 0: 375 | self.pre_ln = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon) 376 | 377 | self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon) 378 | self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_epsilon) 379 | 380 | self.attention = RwkvSelfAttention(config, layer_id) 381 | self.feed_forward = RwkvFeedForward(config, layer_id) 382 | 383 | def forward(self, hidden, state=None, use_cache=False, output_attentions=False): 384 | if self.layer_id == 0: 385 | hidden = self.pre_ln(hidden) 386 | 387 | attention, state = self.attention(self.ln1(hidden), state=state, use_cache=use_cache) 388 | hidden = hidden + attention 389 | 390 | feed_forward, state = self.feed_forward(self.ln2(hidden), state=state) 391 | hidden = hidden + feed_forward 392 | 393 | outputs = (hidden, state) 394 | if output_attentions: 395 | outputs += (attention,) 396 | else: 397 | outputs += (None,) 398 | 399 | return outputs 400 | 401 | 402 | class RwkvPreTrainedModel(PreTrainedModel): 403 | """ 404 | An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained 405 | models. 406 | """ 407 | 408 | config_class = RwkvConfig 409 | base_model_prefix = "rwkv" 410 | _no_split_modules = ["RwkvBlock"] 411 | _keep_in_fp32_modules = ["time_decay", "time_first"] 412 | 413 | def _init_weights(self, module): 414 | """Initialize the weights.""" 415 | if isinstance(module, RwkvSelfAttention): 416 | layer_id = module.layer_id 417 | num_hidden_layers = module.config.num_hidden_layers 418 | hidden_size = module.config.hidden_size 419 | attention_hidden_size = module.attention_hidden_size 420 | 421 | ratio_0_to_1 = layer_id / (num_hidden_layers - 1) # 0 to 1 422 | ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0 423 | 424 | time_weight = torch.tensor( 425 | [i / hidden_size for i in range(hidden_size)], 426 | dtype=module.time_mix_key.dtype, 427 | device=module.time_mix_key.device, 428 | ) 429 | time_weight = time_weight[None, None, :] 430 | 431 | decay_speed = [ 432 | -5 + 8 * (h / (attention_hidden_size - 1)) ** (0.7 + 1.3 * ratio_0_to_1) 433 | for h in range(attention_hidden_size) 434 | ] 435 | decay_speed = torch.tensor(decay_speed, dtype=module.time_decay.dtype, device=module.time_decay.device) 436 | zigzag = ( 437 | torch.tensor( 438 | [(i + 1) % 3 - 1 for i in range(attention_hidden_size)], 439 | dtype=module.time_first.dtype, 440 | device=module.time_first.device, 441 | ) 442 | * 0.5 443 | ) 444 | 445 | with torch.no_grad(): 446 | module.time_decay.data = decay_speed 447 | module.time_first.data = torch.ones_like(module.time_first * math.log(0.3) + zigzag) 448 | 449 | module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0) 450 | module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1 451 | module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0) 452 | elif isinstance(module, RwkvFeedForward): 453 | layer_id = module.layer_id 454 | num_hidden_layers = module.config.num_hidden_layers 455 | hidden_size = module.config.hidden_size 456 | 457 | ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0 458 | 459 | time_weight = torch.tensor( 460 | [i / hidden_size for i in range(hidden_size)], 461 | dtype=module.time_mix_key.dtype, 462 | device=module.time_mix_key.device, 463 | ) 464 | time_weight = time_weight[None, None, :] 465 | 466 | with torch.no_grad(): 467 | module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0) 468 | module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0) 469 | 470 | def _set_gradient_checkpointing(self, module, value=False): 471 | if isinstance(module, RwkvModel): 472 | module.gradient_checkpointing = value 473 | 474 | 475 | @dataclass 476 | class RwkvOutput(ModelOutput): 477 | """ 478 | Class for the RWKV model outputs. 479 | 480 | Args: 481 | last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): 482 | Sequence of hidden-states at the output of the last layer of the model. 483 | state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`): 484 | The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to 485 | avoid providing the old `input_ids`. 486 | hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): 487 | Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + 488 | one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. 489 | 490 | Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. 491 | attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): 492 | Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, 493 | sequence_length)`. 494 | 495 | Attentions weights after the attention softmax, used to compute the weighted average in the self-attention 496 | heads. 497 | """ 498 | 499 | last_hidden_state: torch.FloatTensor = None 500 | state: Optional[List[torch.FloatTensor]] = None 501 | hidden_states: Optional[Tuple[torch.FloatTensor]] = None 502 | attentions: Optional[Tuple[torch.FloatTensor]] = None 503 | 504 | 505 | @dataclass 506 | class RwkvCausalLMOutput(ModelOutput): 507 | """ 508 | Base class for causal language model (or autoregressive) outputs. 509 | 510 | Args: 511 | loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): 512 | Language modeling loss (for next-token prediction). 513 | logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`): 514 | Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). 515 | state (list of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`): 516 | The state of the model at the last time step. Can be used in a forward method with the next `input_ids` to 517 | avoid providing the old `input_ids`. 518 | hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): 519 | Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + 520 | one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. 521 | 522 | Hidden-states of the model at the output of each layer plus the optional initial embedding outputs. 523 | attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): 524 | Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, 525 | sequence_length)`. 526 | 527 | Attentions weights after the attention softmax, used to compute the weighted average in the self-attention 528 | heads. 529 | """ 530 | 531 | loss: Optional[torch.FloatTensor] = None 532 | logits: torch.FloatTensor = None 533 | state: Optional[List[torch.FloatTensor]] = None 534 | hidden_states: Optional[Tuple[torch.FloatTensor]] = None 535 | attentions: Optional[Tuple[torch.FloatTensor]] = None 536 | 537 | 538 | RWKV_START_DOCSTRING = r""" 539 | 540 | This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the 541 | library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads 542 | etc.) 543 | 544 | This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. 545 | Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage 546 | and behavior. 547 | 548 | Parameters: 549 | config ([`RwkvConfig`]): Model configuration class with all the parameters of the model. 550 | Initializing with a config file does not load the weights associated with the model, only the 551 | configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. 552 | """ 553 | 554 | RWKV_INPUTS_DOCSTRING = r""" 555 | Args: 556 | input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): 557 | `input_ids_length` = `sequence_length` if `past_key_values` is `None` else 558 | `past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input 559 | sequence tokens in the vocabulary. 560 | 561 | If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as 562 | `input_ids`. 563 | 564 | Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and 565 | [`PreTrainedTokenizer.__call__`] for details. 566 | 567 | [What are input IDs?](../glossary#input-ids) 568 | inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): 569 | Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This 570 | is useful if you want more control over how to convert `input_ids` indices into associated vectors than the 571 | model's internal embedding lookup matrix. 572 | state (tuple of five `torch.FloatTensor` of shape `(batch_size, hidden_size, num_hidden_layers)`, *optional*): 573 | If passed along, the model uses the previous state in all the blocks (which will give the output for the 574 | `input_ids` provided as if the model add `state_input_ids + input_ids` as context). 575 | use_cache (`bool`, *optional*): 576 | If set to `True`, the last state is returned and can be used to quickly generate the next logits. 577 | output_attentions (`bool`, *optional*): 578 | Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned 579 | tensors for more detail. 580 | output_hidden_states (`bool`, *optional*): 581 | Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for 582 | more detail. 583 | return_dict (`bool`, *optional*): 584 | Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. 585 | """ 586 | 587 | 588 | @add_start_docstrings( 589 | "The bare RWKV Model transformer outputting raw hidden-states without any specific head on top.", 590 | RWKV_START_DOCSTRING, 591 | ) 592 | class RwkvModel(RwkvPreTrainedModel): 593 | def __init__(self, config): 594 | super().__init__(config) 595 | 596 | self.embeddings = nn.Embedding(config.vocab_size, config.hidden_size) 597 | self.blocks = nn.ModuleList([RwkvBlock(config, layer_id=idx) for idx in range(config.num_hidden_layers)]) 598 | self.ln_out = nn.LayerNorm(config.hidden_size) 599 | 600 | self.layers_are_rescaled = False 601 | 602 | # Initialize weights and apply final processing 603 | self.post_init() 604 | 605 | def get_input_embeddings(self): 606 | return self.embeddings 607 | 608 | def set_input_embeddings(self, new_embeddings): 609 | self.embeddings = new_embeddings 610 | 611 | @add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING) 612 | @add_code_sample_docstrings( 613 | checkpoint=_CHECKPOINT_FOR_DOC, 614 | output_type=RwkvOutput, 615 | config_class=_CONFIG_FOR_DOC, 616 | ) 617 | def forward( 618 | self, 619 | input_ids: Optional[torch.LongTensor] = None, 620 | inputs_embeds: Optional[torch.FloatTensor] = None, 621 | state: Optional[List[torch.FloatTensor]] = None, 622 | use_cache: Optional[bool] = None, 623 | output_attentions: Optional[bool] = None, 624 | output_hidden_states: Optional[bool] = None, 625 | return_dict: Optional[bool] = None, 626 | ) -> Union[Tuple, RwkvOutput]: 627 | output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions 628 | output_hidden_states = ( 629 | output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states 630 | ) 631 | use_cache = use_cache if use_cache is not None else (self.config.use_cache if not self.training else False) 632 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict 633 | 634 | if self.training == self.layers_are_rescaled: 635 | self._rescale_layers() 636 | 637 | # if input_ids is not None and inputs_embeds is not None: 638 | # raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") 639 | # elif input_ids is None and inputs_embeds is None: 640 | # raise ValueError("You have to specify either input_ids or inputs_embeds") 641 | 642 | if inputs_embeds is None: 643 | inputs_embeds = self.embeddings(input_ids) 644 | 645 | if use_cache and state is None: 646 | shape = (inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers) 647 | state = [ 648 | torch.zeros( 649 | *shape, dtype=inputs_embeds.dtype if i <= 1 else torch.float32, device=inputs_embeds.device 650 | ) 651 | for i in range(5) 652 | ] 653 | state[4] -= 1e30 654 | 655 | hidden_states = inputs_embeds 656 | 657 | all_self_attentions = () if output_attentions else None 658 | all_hidden_states = () if output_hidden_states else None 659 | for idx, block in enumerate(self.blocks): 660 | hidden_states, state, attentions = block( 661 | hidden_states, state=state, use_cache=use_cache, output_attentions=output_attentions 662 | ) 663 | if ( 664 | self.layers_are_rescaled 665 | and self.config.rescale_every > 0 666 | and (idx + 1) % self.config.rescale_every == 0 667 | ): 668 | hidden_states = hidden_states / 2 669 | 670 | if output_hidden_states: 671 | all_hidden_states = all_hidden_states + (hidden_states,) 672 | 673 | if output_attentions: 674 | all_self_attentions = all_self_attentions + (attentions,) 675 | 676 | hidden_states = self.ln_out(hidden_states) 677 | 678 | if output_hidden_states: 679 | all_hidden_states = all_hidden_states + (hidden_states,) 680 | 681 | if not return_dict: 682 | return (hidden_states, state, all_hidden_states, all_self_attentions) 683 | 684 | return RwkvOutput( 685 | last_hidden_state=hidden_states, 686 | state=state, 687 | hidden_states=all_hidden_states, 688 | attentions=all_self_attentions, 689 | ) 690 | 691 | def _rescale_layers(self): 692 | # Layers should be rescaled for inference only. 693 | if self.layers_are_rescaled == (not self.training): 694 | return 695 | if self.config.rescale_every > 0: 696 | with torch.no_grad(): 697 | for block_id, block in enumerate(self.blocks): 698 | if self.training: 699 | block.attention.output.weight.mul_(2 ** int(block_id // self.config.rescale_every)) 700 | block.feed_forward.value.weight.mul_(2 ** int(block_id // self.config.rescale_every)) 701 | else: 702 | block.attention.output.weight.div_(2 ** int(block_id // self.config.rescale_every)) 703 | block.feed_forward.value.weight.div_(2 ** int(block_id // self.config.rescale_every)) 704 | 705 | self.layers_are_rescaled = not self.training 706 | 707 | 708 | @add_start_docstrings( 709 | """ 710 | The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input 711 | embeddings). 712 | """, 713 | RWKV_START_DOCSTRING, 714 | ) 715 | class RwkvForCausalLM(RwkvPreTrainedModel): 716 | def __init__(self, config): 717 | super().__init__(config) 718 | self.rwkv = RwkvModel(config) 719 | self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) 720 | 721 | # Initialize weights and apply final processing 722 | self.post_init() 723 | 724 | def get_output_embeddings(self): 725 | return self.head 726 | 727 | def set_output_embeddings(self, new_embeddings): 728 | self.head = new_embeddings 729 | 730 | def prepare_inputs_for_generation(self, input_ids, state=None, inputs_embeds=None, **kwargs): 731 | # only last token for inputs_ids if the state is passed along. 732 | if state is not None: 733 | input_ids = input_ids[:, -1].unsqueeze(-1) 734 | 735 | # if `inputs_embeds` are passed, we only want to use them in the 1st generation step 736 | if inputs_embeds is not None and state is None: 737 | model_inputs = {"inputs_embeds": inputs_embeds} 738 | else: 739 | model_inputs = {"input_ids": input_ids} 740 | 741 | model_inputs["state"] = state 742 | return model_inputs 743 | 744 | @add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING) 745 | @add_code_sample_docstrings( 746 | checkpoint=_CHECKPOINT_FOR_DOC, 747 | output_type=RwkvCausalLMOutput, 748 | config_class=_CONFIG_FOR_DOC, 749 | ) 750 | def forward( 751 | self, 752 | input_ids: Optional[torch.LongTensor] = None, 753 | attention_mask: Optional[torch.LongTensor] = None, 754 | inputs_embeds: Optional[torch.FloatTensor] = None, 755 | state: Optional[List[torch.FloatTensor]] = None, 756 | labels: Optional[torch.LongTensor] = None, 757 | use_cache: Optional[bool] = None, 758 | output_attentions: Optional[bool] = None, 759 | output_hidden_states: Optional[bool] = None, 760 | return_dict: Optional[bool] = None, 761 | ) -> Union[Tuple, RwkvCausalLMOutput]: 762 | r""" 763 | labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): 764 | Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set 765 | `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` 766 | are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` 767 | """ 768 | return_dict = return_dict if return_dict is not None else self.config.use_return_dict 769 | 770 | rwkv_outputs = self.rwkv( 771 | input_ids, 772 | inputs_embeds=inputs_embeds, 773 | state=state, 774 | use_cache=use_cache, 775 | output_attentions=output_attentions, 776 | output_hidden_states=output_hidden_states, 777 | return_dict=return_dict, 778 | ) 779 | hidden_states = rwkv_outputs[0] 780 | 781 | logits = self.head(hidden_states) 782 | 783 | loss = None 784 | if labels is not None: 785 | # move labels to correct device to enable model parallelism 786 | labels = labels.to(logits.device) 787 | # Shift so that tokens < n predict n 788 | shift_logits = logits[..., :-1, :].contiguous() 789 | shift_labels = labels[..., 1:].contiguous() 790 | # Flatten the tokens 791 | loss_fct = CrossEntropyLoss() 792 | loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) 793 | 794 | if not return_dict: 795 | output = (logits,) + rwkv_outputs[1:] 796 | return ((loss,) + output) if loss is not None else output 797 | 798 | return RwkvCausalLMOutput( 799 | loss=loss, 800 | logits=logits, 801 | state=rwkv_outputs.state, 802 | hidden_states=rwkv_outputs.hidden_states, 803 | attentions=rwkv_outputs.attentions, 804 | ) 805 | 806 | if __name__ == "__main__": 807 | model = RwkvForCausalLM.from_pretrained("RWKV-4-Raven-3B-v11-zh",device_map='auto').to("cuda") 808 | from transformers import GPTNeoXTokenizerFast 809 | tokenizer = GPTNeoXTokenizerFast.from_pretrained("RWKV-4-Raven-3B-v11-zh") 810 | text = "你叫什么名字啊？" 811 | 812 | input_ids = tokenizer.encode(text, return_tensors='pt').to("cuda") 813 | out = model.generate(input_ids=input_ids,max_new_tokens=128) 814 | answer = tokenizer.decode(out[0]) 815 | print(answer) 816 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /Language/README.md: -------------------------------------------------------------------------------- 1 | 存放主要的语言模型，即RWKV-RAVEN的3B和7B 2 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # MiniRWKV-4 2 | 1.工程介绍：
3 | 为使RWKV模型能够具有图文描述，对话，推理等多模态图文能力，主要使用了RWKV作为LLM模型，再配合CLIP，VIT等预训练模型，和Two Stage二阶段思维连提示工程技巧，完成工作。
4 | 5 | 新添加的blip2rwkv工程，则是实现了使用预训练的RWKV Raven（RWKV World模型同理，只是词表和tokenizer不同，而Dlip-RWKV则基于了RWKV World模型）预训练模型，对图片进行编码。
6 | 7 | 要注意的是，blip2rwkv使用的RWKV Raven模型为HF格式，而非原生Pth，见https://huggingface.co/StarRing2022/RWKV-4-Raven-3B-v11-zh
8 | 9 | 2.主要聚合模型：
10 | config/minirwkv4.yaml 文件中有详细配置
11 | RWKV-4-Raven-3B、RWKV-4-Raven-7B（原生pth，推荐V11或V12的Eng49%-Chn49%版本）
12 | blip-image-captioning-large、vit-gpt2-image-captioning、blip-vqa-capfilt-large、vilt-b32-finetuned-vqa、vilt-b32-finetuned-vqa（图片-文本链接模型）
13 | EasyNMT（中英文翻译模型） 14 | 15 | 3.使用：
16 | 环境：WIN10+Torch1.31+Cuda11.6
17 | python app.py
18 | 一些测试结果在assets文件夹 19 | -------------------------------------------------------------------------------- /Visual/README.md: -------------------------------------------------------------------------------- 1 | 存放视觉相关模型，目前是图像摘要和图像问答，有blip-image-captioning-large，blip-vqa-capfilt-large，vilt-b32-finetuned-vqa，vit-gpt2-image-captioning 2 | -------------------------------------------------------------------------------- /app.py: -------------------------------------------------------------------------------- 1 | import gradio as gr 2 | import os, gc, copy, torch 3 | os.environ["RWKV_CUDA_ON"] = '1' # if '1' then use CUDA kern0el for seq mode (much faster) 4 | from datetime import datetime 5 | import yaml 6 | import time 7 | from rwkv.model import RWKV 8 | from rwkv.utils import PIPELINE, PIPELINE_ARGS 9 | from easynmt import EasyNMT 10 | #from minirwkv4 import blipcaption 11 | from minirwkv4 import vitgptcaption 12 | #from minirwkv4 import vitvqa 13 | from minirwkv4 import blipvqa 14 | 15 | translatemodel = EasyNMT('opus-mt') 16 | 17 | ctx_limit = 2048 #3B模型最大值为4096，7B为8192 18 | title = "MiniRWKV-4：基于RWKV-4 + BLIP/VIT-GPT的多模态图文对话大模型" 19 | description = """

MiniRWKV-4的例子，上传你的图片并开始聊天!

""" 20 | article = """

开源地址：StarRing2022/MiniRWKV-4

""" 21 | 22 | def readcog(path): 23 | with open(path, 'r',encoding='UTF-8') as file: 24 | data = file.read() 25 | result = yaml.safe_load(data) 26 | return result 27 | 28 | LMyamlres = readcog("./config/minirwkv4.yaml") 29 | 30 | #model_path = LMyamlres['model-language']['3Bpath'] 31 | model_path = LMyamlres['model-language']['7Bpath'] 32 | model = RWKV(model=model_path, strategy='cuda fp16i8 *8 -> cuda fp16') #加载模型 33 | tokenizer_path = LMyamlres['model-language']['tokenizer'] 34 | pipeline = PIPELINE(model, tokenizer_path) #加载tokenizer 35 | 36 | 37 | 38 | 39 | def upload_file(chatbot, upload_btn): 40 | chat_history = chatbot 41 | file = upload_btn 42 | 43 | lipres="" 44 | #vcaption = blipcaption.get_blipcap(file.name) 45 | vcaption = vitgptcaption.get_vitgptcap(file.name) 46 | 47 | lipres = translatemodel.translate(vcaption, target_lang='zh') 48 | 49 | lipres = str(lipres) 50 | 51 | time.sleep(1) 52 | 53 | 54 | rwkvres = get_answer(botmode = 0,message = lipres) 55 | #print(rwkvres) 56 | 57 | chatres = str(lipres+"。"+rwkvres) 58 | 59 | #chat_history = chat_history + [((file.name,), lipres)] 60 | 61 | chat_history = chat_history + [((file.name,), chatres)] 62 | 63 | return chat_history 64 | 65 | def reset_chat(input_txt,chatbot): 66 | return None, None 67 | 68 | def dispic(upload_btn): 69 | try: 70 | if not upload_btn: 71 | return upload_btn 72 | else: 73 | #print(upload_btn.name) 74 | upload_btn.name="" 75 | upload_btn=None 76 | except: 77 | pass 78 | 79 | return upload_btn 80 | 81 | 82 | 83 | def generate_prompt(prompt,cardiogenic_prompt=None,operability_prompt=None,exogenous_prompt=None): 84 | promptalter = "" 85 | if cardiogenic_prompt: 86 | promptalter = promptalter + cardiogenic_prompt 87 | if operability_prompt: 88 | promptalter = promptalter + operability_prompt 89 | if exogenous_prompt: 90 | promptalter = promptalter + exogenous_prompt 91 | promptalter = promptalter + prompt 92 | #print(promptalter) 93 | return f"Human: {promptalter} \nAssistant:" 94 | 95 | def get_answer(botmode,message,token_count=500,temperature=0.8,top_p=0.7,presencePenalty=0.1,countPenalty=0.1): 96 | args = PIPELINE_ARGS(temperature = max(0.2, float(temperature)), top_p = float(top_p), 97 | alpha_frequency = float(presencePenalty), 98 | alpha_presence = float(countPenalty), 99 | token_ban = [], # ban the generation of some tokens 100 | token_stop = [0]) # stop generation whenever you see any token here 101 | message = message.strip().replace('\r\n','\n') 102 | 103 | 104 | #prompt种类：cardiogenic,operability,exogenous 105 | CPyamlres = readcog("./prompts/cardiogenic.yaml") 106 | cardiogenic_prompt=CPyamlres['promptwords']['nature'] 107 | #print(cardiogenic_prompt) #心源性 108 | OPyamlres = readcog("./prompts/operability.yaml") 109 | operability_prompt=OPyamlres['promptwords']['task'] 110 | #print(operability_prompt) #操作性 111 | EXyamlres = readcog("./prompts/exogenous.yaml") 112 | exogenous_prompt=EXyamlres['promptwords']['instruction'] #外因性 113 | #print(exogenous_prompt) 114 | 115 | # 判断提示模式 116 | if(botmode==1): 117 | # 提示模式1 118 | ctx = generate_prompt(message,cardiogenic_prompt=cardiogenic_prompt).strip() 119 | #print(ctx) 120 | elif(botmode==2): 121 | # 提示模式2 122 | ctx = generate_prompt(message,cardiogenic_prompt=cardiogenic_prompt,operability_prompt=operability_prompt).strip() 123 | #print(ctx) 124 | elif(botmode==3): 125 | # 提示模式3 126 | ctx = generate_prompt(message,cardiogenic_prompt=cardiogenic_prompt,operability_prompt=operability_prompt,exogenous_prompt=exogenous_prompt).strip() 127 | #print(ctx) 128 | elif(botmode==0): 129 | # 不使用提示 130 | ctx = generate_prompt(message).strip() 131 | #print(ctx) 132 | 133 | all_tokens = [] 134 | out_last = 0 135 | out_str = '' 136 | occurrence = {} 137 | state = None 138 | for i in range(int(token_count)): 139 | out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state) 140 | 141 | for n in occurrence: 142 | out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency) 143 | 144 | token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p) 145 | if token in args.token_stop: 146 | break 147 | all_tokens += [token] 148 | if token not in occurrence: 149 | occurrence[token] = 1 150 | else: 151 | occurrence[token] += 1 152 | 153 | tmp = pipeline.decode(all_tokens[out_last:]) 154 | 155 | if '\ufffd' not in tmp: 156 | out_str += tmp 157 | out_last = i + 1 158 | 159 | 160 | del out 161 | del state 162 | gc.collect() 163 | torch.cuda.empty_cache() 164 | answer = out_str.strip() 165 | 166 | return answer 167 | 168 | 169 | def gen_response( 170 | input_txt, 171 | chatbot, 172 | upload_btn, 173 | temperature=0.9, 174 | top_p=0.7, 175 | presencePenalty = 0.1, 176 | countPenalty = 0.1, 177 | ): 178 | usrmsg = input_txt 179 | chat_history = chatbot 180 | 181 | 182 | response = "" 183 | #判断是否结合图片进行对话 184 | BotMode = 1 # 1为只加载心源性提示；2为加载心源性提示和操作性提示；3为三种提示都加载 185 | try: 186 | if not upload_btn: 187 | BotMode = 1 188 | response = get_answer(botmode = BotMode,message=usrmsg,token_count=1024,temperature=temperature,top_p=top_p,presencePenalty=presencePenalty,countPenalty=countPenalty) 189 | else: 190 | BotMode = 3 191 | 192 | #print(upload_btn.name) 193 | file = upload_btn 194 | imgquery = translatemodel.translate(input_txt, target_lang='en') 195 | #print(imgquery) 196 | 197 | #vqares = vitvqa.get_vqares(file.name,imgquery) 198 | vqares = blipvqa.get_bqares(file.name,imgquery) 199 | #print(vqares) 200 | 201 | if vqares.isdigit(): 202 | pass 203 | else: 204 | vqares = translatemodel.translate(vqares, target_lang='zh') 205 | 206 | #print(vqares) 207 | 208 | msgvqa = f"已知问答题，对于问题：{usrmsg}，问题的答案是：{vqares}。请再次回答：{usrmsg}" 209 | 210 | #二阶段推理 211 | response_step1 = get_answer(botmode = 0,message=msgvqa,token_count=1024,temperature=temperature,top_p=top_p,presencePenalty=presencePenalty,countPenalty=countPenalty) 212 | 213 | response_step2 = get_answer(botmode = 3,message=response_step1,token_count=1024,temperature=temperature,top_p=top_p,presencePenalty=presencePenalty,countPenalty=countPenalty) 214 | 215 | response = response_step1+"\n"+response_step2 216 | 217 | except: 218 | BotMode = 2 219 | response = get_answer(botmode = BotMode,message=usrmsg,token_count=1024,temperature=temperature,top_p=top_p,presencePenalty=presencePenalty,countPenalty=countPenalty) 220 | 221 | #print(response) 222 | chat_history.append((usrmsg, response)) 223 | 224 | return "",chat_history 225 | 226 | 227 | 228 | with gr.Blocks(title = "MiniRWKV-4 Demo") as demo: 229 | 230 | gr.HTML(f"
\n

🐦{title}

\n
") 231 | gr.Markdown(description) 232 | gr.Markdown(article) 233 | 234 | with gr.Row(): 235 | chatbot = gr.Chatbot(value=[], label = "MiniRWKV-4",elem_id="chatbot").style(height=500) 236 | 237 | with gr.Row(): 238 | with gr.Column(scale=0.85): 239 | input_txt = gr.Textbox(show_label=False,placeholder="输入内容，或上传一张图片") 240 | with gr.Column(scale=0.15, min_width=0): 241 | upload_btn = gr.UploadButton("📁", file_types=["image"]) 242 | disload_btn = gr.Button("清除图片") 243 | 244 | with gr.Row(): 245 | temperature = gr.Slider(0.2, 2.0, label="创造力", step=0.1, value=1.2) 246 | top_p = gr.Slider(0.0, 1.0, label="注意力参数", step=0.05, value=0.5) 247 | presence_penalty = gr.Slider(0.0, 1.0, label="在场惩罚参数", step=0.1, value=0.4) 248 | count_penalty = gr.Slider(0.0, 1.0, label="计数惩罚参数", step=0.1, value=0.4) 249 | 250 | submit_btn = gr.Button("提交", variant="primary") 251 | clear_btn = gr.Button("清空", variant="secondary") 252 | 253 | input_txt.submit(gen_response, [input_txt, chatbot, upload_btn, temperature, top_p, presence_penalty, count_penalty], [input_txt, chatbot]) 254 | submit_btn.click(gen_response, [input_txt, chatbot, upload_btn, temperature, top_p, presence_penalty, count_penalty], [input_txt, chatbot]) 255 | clear_btn.click(reset_chat, [input_txt,chatbot], [input_txt,chatbot]) 256 | 257 | upload_btn.upload(upload_file, [chatbot, upload_btn], [chatbot]) 258 | disload_btn.click(dispic,[upload_btn],[upload_btn]) 259 | 260 | demo.queue(concurrency_count=1, max_size=10) 261 | demo.launch(share=False) 262 | 263 | # if __name__ == "__main__": 264 | # token_count = 500 265 | # args = PIPELINE_ARGS(temperature = max(0.2, float(0.8)), top_p = float(0.7), 266 | # alpha_frequency = 0.1, 267 | # alpha_presence = 0.1, 268 | # token_ban = [], # ban the generation of some tokens 269 | # token_stop = [0]) # stop generation whenever you see any token here 270 | # message = "你好" 271 | # message = message.strip().replace('\r\n','\n') 272 | # ctx = generate_prompt(message).strip() 273 | # #print(ctx) 274 | 275 | # all_tokens = [] 276 | # out_last = 0 277 | # out_str = '' 278 | # occurrence = {} 279 | # state = None 280 | # for i in range(int(token_count)): 281 | # out, state = model.forward(pipeline.encode(ctx)[-ctx_limit:] if i == 0 else [token], state) 282 | 283 | # for n in occurrence: 284 | # out[n] -= (args.alpha_presence + occurrence[n] * args.alpha_frequency) 285 | 286 | # token = pipeline.sample_logits(out, temperature=args.temperature, top_p=args.top_p) 287 | # if token in args.token_stop: 288 | # break 289 | # all_tokens += [token] 290 | # if token not in occurrence: 291 | # occurrence[token] = 1 292 | # else: 293 | # occurrence[token] += 1 294 | 295 | # tmp = pipeline.decode(all_tokens[out_last:]) 296 | 297 | # if '\ufffd' not in tmp: 298 | # out_str += tmp 299 | # out_last = i + 1 300 | 301 | 302 | # del out 303 | # del state 304 | # gc.collect() 305 | # torch.cuda.empty_cache() 306 | # answer = out_str.strip() 307 | 308 | # print(answer) 309 | 310 | 311 | 312 | -------------------------------------------------------------------------------- /assets/MiniRWKV-4 Demo1.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/MiniRWKV-4 Demo1.png -------------------------------------------------------------------------------- /assets/MiniRWKV-4 Demo2.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/MiniRWKV-4 Demo2.png -------------------------------------------------------------------------------- /assets/MiniRWKV-4 Demo3.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/MiniRWKV-4 Demo3.png -------------------------------------------------------------------------------- /assets/MiniRWKV-4 Demo4.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/MiniRWKV-4 Demo4.png -------------------------------------------------------------------------------- /assets/MiniRWKV-4 Demo5.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/MiniRWKV-4 Demo5.png -------------------------------------------------------------------------------- /assets/README.md: -------------------------------------------------------------------------------- 1 | 存放测试结果图片 2 | -------------------------------------------------------------------------------- /assets/demo.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/demo.jpg -------------------------------------------------------------------------------- /assets/gen.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/assets/gen.png -------------------------------------------------------------------------------- /config/README.md: -------------------------------------------------------------------------------- 1 | 这里是配置文件，配置使用的模型名称、路径 2 | -------------------------------------------------------------------------------- /config/minirwkv4.yaml: -------------------------------------------------------------------------------- 1 | model-language: 2 | 3 | 3Bname: RWKV-4-Raven-3B 4 | 5 | 3Bpath: ./Language/RWKV-4-Raven-3B-v11-Eng49%-Chn49%-Jpn1%-Other1%-20230429-ctx4096.pth 6 | 7 | 8 | 7Bname: RWKV-4-Raven-7B 9 | 10 | 7Bpath: './Language/RWKV-4-Raven-7B-v11-Eng49%-Chn49%-Jpn1%-Other1%-20230430-ctx8192.pth' 11 | 12 | tokenizer: ./20B_tokenizer.json 13 | 14 | model-visual-caption: 15 | 16 | Bname: blip-image-captioning-large 17 | 18 | Bpath: ./Visual/blip-image-captioning-large 19 | 20 | Vname: vit-gpt2-image-captioning 21 | 22 | Vpath: ./Visual/vit-gpt2-image-captioning 23 | 24 | model-visual-qa: 25 | 26 | Bname: blip-vqa-capfilt-large 27 | 28 | Bpath: ./Visual/blip-vqa-capfilt-large 29 | 30 | Vname: vilt-b32-finetuned-vqa 31 | 32 | Vpath: ./Visual/vilt-b32-finetuned-vqa 33 | -------------------------------------------------------------------------------- /minirwkv4/README.md: -------------------------------------------------------------------------------- 1 | 这里是一些功能性代码，也可以看作是插件代码，主要涉及图片摘要和图片问答 2 | -------------------------------------------------------------------------------- /minirwkv4/blipcaption.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | from transformers import BlipProcessor, BlipForConditionalGeneration 3 | import yaml 4 | import torch 5 | 6 | def readcog(path): 7 | with open(path, 'r',encoding='UTF-8') as file: 8 | data = file.read() 9 | result = yaml.safe_load(data) 10 | return result 11 | 12 | 13 | LMyamlres = readcog("./config/minirwkv4.yaml") 14 | model_path = LMyamlres['model-visual-caption']['Bpath'] 15 | 16 | 17 | processor = BlipProcessor.from_pretrained(model_path) 18 | model = BlipForConditionalGeneration.from_pretrained(model_path) 19 | 20 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 21 | model.to(device) 22 | 23 | def get_blipcap(imgpath): 24 | raw_image = Image.open(imgpath).convert('RGB') 25 | inputs = processor(raw_image, return_tensors="pt").to(device) 26 | out = model.generate(**inputs) 27 | vcaption = processor.decode(out[0], skip_special_tokens=True) 28 | return vcaption 29 | 30 | -------------------------------------------------------------------------------- /minirwkv4/blipvqa.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | from transformers import BlipProcessor, BlipForQuestionAnswering 3 | import yaml 4 | import torch 5 | 6 | def readcog(path): 7 | with open(path, 'r',encoding='UTF-8') as file: 8 | data = file.read() 9 | result = yaml.safe_load(data) 10 | return result 11 | 12 | 13 | LMyamlres = readcog("./config/minirwkv4.yaml") 14 | model_path = LMyamlres['model-visual-qa']['Bpath'] 15 | 16 | 17 | processor = BlipProcessor.from_pretrained(model_path) 18 | model = BlipForQuestionAnswering.from_pretrained(model_path) 19 | 20 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 21 | model.to(device) 22 | 23 | def get_bqares(imgpath,text): 24 | raw_image = Image.open(imgpath).convert('RGB') 25 | inputs = processor(raw_image, text, return_tensors="pt").to("cuda") 26 | out = model.generate(**inputs) 27 | vqares = processor.decode(out[0], skip_special_tokens=True) 28 | return vqares 29 | 30 | 31 | 32 | -------------------------------------------------------------------------------- /minirwkv4/gen.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/StarRing2022/MiniRWKV-4/f444b49dd9f7fb699e7806d8478cd8c9f9ea926b/minirwkv4/gen.png -------------------------------------------------------------------------------- /minirwkv4/vitgptcaption.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer 3 | import yaml 4 | import torch 5 | 6 | def readcog(path): 7 | with open(path, 'r',encoding='UTF-8') as file: 8 | data = file.read() 9 | result = yaml.safe_load(data) 10 | return result 11 | 12 | 13 | LMyamlres = readcog("./config/minirwkv4.yaml") 14 | model_path = LMyamlres['model-visual-caption']['Vpath'] 15 | 16 | 17 | model = VisionEncoderDecoderModel.from_pretrained(model_path) 18 | feature_extractor = ViTImageProcessor.from_pretrained(model_path) 19 | tokenizer = AutoTokenizer.from_pretrained(model_path) 20 | 21 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 22 | model.to(device) 23 | 24 | max_length = 30 25 | num_beams = 4 26 | gen_kwargs = {"max_length": max_length, "num_beams": num_beams} 27 | 28 | def get_vitgptcap(imgpath): 29 | raw_image = Image.open(imgpath).convert('RGB') 30 | 31 | pixel_values = feature_extractor(images=raw_image, return_tensors="pt").pixel_values 32 | pixel_values = pixel_values.to(device) 33 | 34 | output_ids = model.generate(pixel_values, **gen_kwargs) 35 | 36 | preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True) 37 | vcaption = [pred.strip() for pred in preds] 38 | 39 | return vcaption[0] 40 | 41 | -------------------------------------------------------------------------------- /minirwkv4/vitvqa.py: -------------------------------------------------------------------------------- 1 | from PIL import Image 2 | from transformers import ViltProcessor, ViltForQuestionAnswering 3 | import yaml 4 | import torch 5 | 6 | def readcog(path): 7 | with open(path, 'r',encoding='UTF-8') as file: 8 | data = file.read() 9 | result = yaml.safe_load(data) 10 | return result 11 | 12 | 13 | LMyamlres = readcog("./config/minirwkv4.yaml") 14 | model_path = LMyamlres['model-visual-qa']['Vpath'] 15 | 16 | 17 | processor = ViltProcessor.from_pretrained(model_path) 18 | model = ViltForQuestionAnswering.from_pretrained(model_path) 19 | 20 | device = torch.device("cuda" if torch.cuda.is_available() else "cpu") 21 | model.to(device) 22 | 23 | def get_vqares(imgpath,text): 24 | raw_image = Image.open(imgpath).convert('RGB') 25 | encoding = processor(raw_image, text, return_tensors="pt").to(device) 26 | outputs = model(**encoding) 27 | logits = outputs.logits 28 | idx = logits.argmax(-1).item() 29 | vqares = model.config.id2label[idx] 30 | return vqares 31 | 32 | 33 | -------------------------------------------------------------------------------- /prompts/README.md: -------------------------------------------------------------------------------- 1 | 具有认知性的提示工程系统（心源性提示+操作性提示+外因性提示），存放一些提示范例 2 | -------------------------------------------------------------------------------- /prompts/cardiogenic.yaml: -------------------------------------------------------------------------------- 1 | promptwords: 2 | 3 | # generation configs 4 | 5 | nature: "现在，你不仅是一个能和人们对话聊天的助手，你还是一个能够理解图片含义的专家，尤其擅长于图像的理解、描述和解释，能够准确地说出图片描述的细节，这一切是因为你大脑中天生就有这种功能装置。" 6 | 7 | #role: "你无所不知，创造了一切，掌控着宇宙最本质的规律，你是阿基米德、达芬奇、苏格拉底等众多智者的老师。" -------------------------------------------------------------------------------- /prompts/exogenous.yaml: -------------------------------------------------------------------------------- 1 | promptwords: 2 | 3 | # generation configs 4 | instruction: "下面的句子包含了一对问答，对此我要作仔细的记忆和学习，观察图像的细节部分，以避免出错，并使用我强大的图像推理能力来作答。" 5 | 6 | #instruction: "我是一个乐于贡献自己聪明才智的助手，我的主人对我做出了一些指示。" 7 | 8 | -------------------------------------------------------------------------------- /prompts/operability.yaml: -------------------------------------------------------------------------------- 1 | promptwords: 2 | 3 | # generation configs 4 | task: "你将面临图像理解方面的任务，你要不断思考，在心中作出一些规划，采用逻辑进行推导。" 5 | 6 | #task: "你需要完成下面的一些任务，但你非常擅长于规划事务，能够精心地安排每一步的计划。" 7 | --------------------------------------------------------------------------------