# Copyright (c) 2024 Alibaba Inc # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch.nn as nn import torch from transformers import AutoModelForCausalLM, AutoTokenizer from transformers import Qwen2ForCausalLM, Qwen2Config from torch.nn.utils.rnn import pad_sequence, unpad_sequence from tqdm import tqdm def topk_sampling(weighted_scores, decoded_tokens, top_k=1024): zeros = weighted_scores.new_ones(weighted_scores.shape) * float('-inf') values,indices = torch.topk(weighted_scores,top_k) zeros.scatter_(-1, indices, values) return random_sampling(zeros,decoded_tokens) # Repetition Aware Sampling in VALL-E 2 def ras_sampling(weighted_scores, decoded_tokens, top_p=0.8, top_k=50, win_size=10, tau_r=0.1): top_ids = nucleus_sampling(weighted_scores, top_p=top_p, top_k=top_k) rep_num = (torch.tensor(decoded_tokens[-win_size:]).to(weighted_scores.device) == top_ids).sum().item() if rep_num >= win_size * tau_r: top_ids = random_sampling(weighted_scores, decoded_tokens) return top_ids def caras_sampling(weighted_scores, decoded_tokens, top_p=0.8, top_k=25, win_size=10, tau_r=0.1): weighted_scores, cfg_weighted_scores = weighted_scores top_ids = nucleus_sampling(weighted_scores, top_p=top_p, top_k=top_k) rep_num = (torch.tensor(decoded_tokens[-win_size:]).to(weighted_scores.device) == top_ids).sum().item() if rep_num >= win_size * tau_r: top_ids = random_sampling(cfg_weighted_scores, decoded_tokens) return top_ids def nucleus_sampling(weighted_scores, top_p=0.8, top_k=25): prob, indices = [], [] cum_prob = 0.0 sorted_value, sorted_idx = weighted_scores.softmax(dim=0).sort(descending=True, stable=True) for i in range(len(sorted_idx)): # sampling both top-p and numbers. if cum_prob < top_p and len(prob) < top_k: cum_prob += sorted_value[i] prob.append(sorted_value[i]) indices.append(sorted_idx[i]) else: break prob = torch.tensor(prob).to(weighted_scores) indices = torch.tensor(indices, dtype=torch.long).to(weighted_scores.device) top_ids = indices[prob.multinomial(1, replacement=True)] return top_ids def random_sampling(weighted_scores, decoded_tokens): top_ids = weighted_scores.softmax(dim=0).multinomial(1, replacement=True) return top_ids def th_accuracy(pad_outputs: torch.Tensor, pad_targets: torch.Tensor, ignore_label: int) -> torch.Tensor: """Calculate accuracy. Args: pad_outputs (Tensor): Prediction tensors (B * Lmax, D). pad_targets (LongTensor): Target label tensors (B, Lmax). ignore_label (int): Ignore label id. Returns: torch.Tensor: Accuracy value (0.0 - 1.0). """ pad_pred = pad_outputs.view(pad_targets.size(0), pad_targets.size(1), pad_outputs.size(1)).argmax(2) mask = pad_targets != ignore_label numerator = torch.sum( pad_pred.masked_select(mask) == pad_targets.masked_select(mask)) denominator = torch.sum(mask) return (numerator / denominator).detach() def subsequent_mask( size: int, device: torch.device = torch.device("cpu"), ) -> torch.Tensor: """Create mask for subsequent steps (size, size). This mask is used only in decoder which works in an auto-regressive mode. This means the current step could only do attention with its left steps. In encoder, fully attention is used when streaming is not necessary and the sequence is not long. In this case, no attention mask is needed. When streaming is need, chunk-based attention is used in encoder. See subsequent_chunk_mask for the chunk-based attention mask. Args: size (int): size of mask str device (str): "cpu" or "cuda" or torch.Tensor.device dtype (torch.device): result dtype Returns: torch.Tensor: mask Examples: >>> subsequent_mask(3) [[1, 0, 0], [1, 1, 0], [1, 1, 1]] """ arange = torch.arange(size, device=device) mask = arange.expand(size, size) arange = arange.unsqueeze(-1) mask = mask <= arange return mask def subsequent_chunk_mask( size: int, chunk_size: int, num_left_chunks: int = -1, device: torch.device = torch.device("cpu"), ) -> torch.Tensor: """Create mask for subsequent steps (size, size) with chunk size, this is for streaming encoder Args: size (int): size of mask chunk_size (int): size of chunk num_left_chunks (int): number of left chunks <0: use full chunk >=0: use num_left_chunks device (torch.device): "cpu" or "cuda" or torch.Tensor.device Returns: torch.Tensor: mask Examples: >>> subsequent_chunk_mask(4, 2) [[1, 1, 0, 0], [1, 1, 0, 0], [1, 1, 1, 1], [1, 1, 1, 1]] """ ret = torch.zeros(size, size, device=device, dtype=torch.bool) for i in range(size): if num_left_chunks < 0: start = 0 else: start = max((i // chunk_size - num_left_chunks) * chunk_size, 0) ending = min((i // chunk_size + 1) * chunk_size, size) ret[i, start:ending] = True return ret def add_optional_chunk_mask(xs: torch.Tensor, masks: torch.Tensor, use_dynamic_chunk: bool, use_dynamic_left_chunk: bool, decoding_chunk_size: int, static_chunk_size: int, num_decoding_left_chunks: int, enable_full_context: bool = True): """ Apply optional mask for encoder. Args: xs (torch.Tensor): padded input, (B, L, D), L for max length mask (torch.Tensor): mask for xs, (B, 1, L) use_dynamic_chunk (bool): whether to use dynamic chunk or not use_dynamic_left_chunk (bool): whether to use dynamic left chunk for training. decoding_chunk_size (int): decoding chunk size for dynamic chunk, it's 0: default for training, use random dynamic chunk. <0: for decoding, use full chunk. >0: for decoding, use fixed chunk size as set. static_chunk_size (int): chunk size for static chunk training/decoding if it's greater than 0, if use_dynamic_chunk is true, this parameter will be ignored num_decoding_left_chunks: number of left chunks, this is for decoding, the chunk size is decoding_chunk_size. >=0: use num_decoding_left_chunks <0: use all left chunks enable_full_context (bool): True: chunk size is either [1, 25] or full context(max_len) False: chunk size ~ U[1, 25] Returns: torch.Tensor: chunk mask of the input xs. """ # Whether to use chunk mask or not if use_dynamic_chunk: max_len = xs.size(1) if decoding_chunk_size < 0: chunk_size = max_len num_left_chunks = -1 elif decoding_chunk_size > 0: chunk_size = decoding_chunk_size num_left_chunks = num_decoding_left_chunks else: # chunk size is either [1, 25] or full context(max_len). # Since we use 4 times subsampling and allow up to 1s(100 frames) # delay, the maximum frame is 100 / 4 = 25. chunk_size = torch.randint(1, max_len, (1, )).item() num_left_chunks = -1 if chunk_size > max_len // 2 and enable_full_context: chunk_size = max_len else: chunk_size = chunk_size % 25 + 1 if use_dynamic_left_chunk: max_left_chunks = (max_len - 1) // chunk_size num_left_chunks = torch.randint(0, max_left_chunks, (1, )).item() chunk_masks = subsequent_chunk_mask(xs.size(1), chunk_size, num_left_chunks, xs.device) # (L, L) chunk_masks = chunk_masks.unsqueeze(0) # (1, L, L) chunk_masks = masks & chunk_masks # (B, L, L) elif static_chunk_size > 0: num_left_chunks = num_decoding_left_chunks chunk_masks = subsequent_chunk_mask(xs.size(1), static_chunk_size, num_left_chunks, xs.device) # (L, L) chunk_masks = chunk_masks.unsqueeze(0) # (1, L, L) chunk_masks = masks & chunk_masks # (B, L, L) else: chunk_masks = masks return chunk_masks def make_pad_mask(lengths: torch.Tensor, max_len: int = 0) -> torch.Tensor: """Make mask tensor containing indices of padded part. See description of make_non_pad_mask. Args: lengths (torch.Tensor): Batch of lengths (B,). Returns: torch.Tensor: Mask tensor containing indices of padded part. Examples: >>> lengths = [5, 3, 2] >>> make_pad_mask(lengths) masks = [[0, 0, 0, 0 ,0], [0, 0, 0, 1, 1], [0, 0, 1, 1, 1]] """ batch_size = lengths.size(0) max_len = max_len if max_len > 0 else lengths.max().item() seq_range = torch.arange(0, max_len, dtype=torch.int64, device=lengths.device) seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len) seq_length_expand = lengths.unsqueeze(-1) mask = seq_range_expand >= seq_length_expand return mask qwen_conf = { "architectures": [ "Qwen2ForCausalLM" ], "attention_dropout": 0.0, "bos_token_id": 151643, "eos_token_id": 151643, "hidden_act": "silu", "hidden_size": 896, "initializer_range": 0.02, "intermediate_size": 4864, "max_position_embeddings": 32768, "max_window_layers": 24, "model_type": "qwen2", "num_attention_heads": 14, "num_hidden_layers": 24, "num_key_value_heads": 2, "rms_norm_eps": 1e-06, "rope_theta": 1000000.0, "sliding_window": 32768, "tie_word_embeddings": True, "torch_dtype": "bfloat16", "transformers_version": "4.40.1", "use_cache": True, "use_mrope": False, "use_sliding_window": False, "vocab_size": 1024 } # qwen_conf = { # "architectures": [ # "Qwen2ForCausalLM" # ], # "attention_dropout": 0.0, # "bos_token_id": 151643, # "eos_token_id": 151643, # "hidden_act": "silu", # "hidden_size": 768, # "initializer_range": 0.02, # "intermediate_size": 3072, # "max_position_embeddings": 32768, # "max_window_layers": 24, # "model_type": "qwen2", # "num_attention_heads": 12, # "num_hidden_layers": 24, # "num_key_value_heads": 2, # "rms_norm_eps": 1e-06, # "rope_theta": 1000000.0, # "sliding_window": 32768, # "tie_word_embeddings": True, # "torch_dtype": "bfloat16", # "transformers_version": "4.40.1", # "use_cache": True, # "use_mrope": False, # "use_sliding_window": False, # "vocab_size": 1024 # } DTYPES = { "bf16": torch.bfloat16, "fp16": torch.float16, } IGNORE_ID = -100 MAX_TEXT_TOKEN = 1024 BERT_DIM = 1024 SPK_EMB_DIM = 256 class QwenEmbeddingEncoder(nn.Module): def __init__( self, dtype: str = "bf16" ): super(QwenEmbeddingEncoder, self).__init__() self.dtype = DTYPES.get(dtype, torch.float32) config = Qwen2Config(**qwen_conf) self.model = AutoModelForCausalLM.from_config(config) self._output_size = self.model.config.hidden_size def output_size(self) -> int: return self._output_size def forward( self, input_embeds: torch.Tensor, ilens: torch.Tensor, ): input_masks = (~make_pad_mask(ilens)).to(input_embeds.device).long() outs = self.model( inputs_embeds=input_embeds, attention_mask=input_masks, output_hidden_states=True, return_dict=True, ) return outs.hidden_states[-1], input_masks def forward_one_step(self, xs, masks, cache=None): outs = self.model( inputs_embeds=xs, attention_mask=masks, output_hidden_states=True, return_dict=True, use_cache=True, past_key_values=cache, ) xs = outs.hidden_states[-1] new_cache = outs.past_key_values assert new_cache is not None return xs, masks, new_cache class LLM(torch.nn.Module): def __init__( self, audio_token_size: int, dtype: str = "fp32", ): super().__init__() self.dtype = DTYPES.get(dtype, torch.float32) # 1. build text token inputs related modules self.llm = QwenEmbeddingEncoder(dtype=dtype) self.llm.model.model.embed_tokens = None self.llm_input_size = self.llm.model.config.hidden_size self.llm_output_size = self.llm.model.config.hidden_size self.audio_token_size = audio_token_size self.llm_decoder = nn.Linear(self.llm_output_size, audio_token_size + 1) self.criterion_ce = torch.nn.CrossEntropyLoss(ignore_index=IGNORE_ID) # 3. [Optional] build audio token related modules self.speech_embedding = torch.nn.Embedding(audio_token_size, self.llm_input_size) self.text_embedding = nn.Embedding(MAX_TEXT_TOKEN, self.llm_input_size) self.bert_embedding = nn.Linear(BERT_DIM, self.llm_input_size) self.spk_embedding = nn.Linear(SPK_EMB_DIM, self.llm_input_size) self.audio_bos = nn.Parameter(torch.randn(1, self.llm_input_size)) def forward( self, text_tokens, bert_features, spk_features, prosody_tokens, lengths, ): """ Args: text_tokens: (B, L) bert_features: (B, T, 1024) spk_features: (B, 256) prosody_tokens: (B, L) lengths: (B,) """ device = text_tokens.device bsz = text_tokens.size(0) # get input embeddings spk_embedding = self.spk_embedding(spk_features) # [b, 256] -> [b, h] text_embedding = self.text_embedding(text_tokens) # [b, t] -> [b, t, h] bert_embedding = self.bert_embedding(bert_features) # [b, t, 1024] -> [b, t, h] text_embedding = text_embedding + bert_embedding audio_embedding = self.speech_embedding(prosody_tokens) #[b, t] -> [b, t, h] # get lm input and target lm_input = [] lm_target = [] for b in range(bsz): # cond part spk_emb = spk_embedding[b].unsqueeze(0) # [h] -> [1, h] text_emb = text_embedding[b][:lengths[b]] # [t, h] # ar input audio_emb = audio_embedding[b][ :lengths[b]] # [t, h] audio_emb = torch.cat([self.audio_bos, audio_emb], dim=0) # [t+1, h] cur_inp = torch.cat([spk_emb, text_emb, audio_emb], dim=0) # [2t+2, h] lm_input.append(cur_inp) # ar target cur_target = prosody_tokens[b][:lengths[b]] # [t] eos = torch.tensor([self.audio_token_size], dtype=torch.int64).to(device) cur_target = torch.cat([cur_target, eos], dim=0) # [t+1] cur_target = torch.nn.functional.pad(cur_target, (cur_inp.shape[0] - cur_target.size(0), 0), value=IGNORE_ID) # [2t+2] lm_target.append(cur_target) lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32).to(device) lm_input = pad_sequence(lm_input, batch_first=True, padding_value=IGNORE_ID) lm_target = pad_sequence(lm_target, batch_first=True, padding_value=IGNORE_ID) # 6. run lm forward lm_output, lm_output_mask = self.llm(lm_input.to(self.dtype), lm_input_len.to(device)) logits = self.llm_decoder(lm_output) loss = self.criterion_ce(logits.view(-1, self.audio_token_size + 1), lm_target.view(-1)) # print(loss,lm_input.shape) acc = th_accuracy(logits.view(-1, self.audio_token_size + 1), lm_target, ignore_label=IGNORE_ID) return {'loss': loss, 'acc': acc} @torch.inference_mode() @torch.no_grad() def inference( self, text_tokens, bert_features, spk_features, prompt_audio_token=None, sampling_method=None, sampling_kwargs=None, ): device = text_tokens.device if prompt_audio_token is not None: res_tokens = prompt_audio_token.clone() audio_embedding = self.speech_embedding(prompt_audio_token) # [b, t] -> [b, t, h] audio_embedding = torch.cat([self.audio_bos.unsqueeze(0), audio_embedding], dim=1) # [b, t+1, h] else: res_tokens = torch.zeros((text_tokens.size(0), 0), dtype=torch.int64).to(device) audio_embedding = self.audio_bos.unsqueeze(0) # [1, 1, h] # get input embeddings spk_embedding = self.spk_embedding(spk_features) # [b, 256] -> [b, h] text_embedding = self.text_embedding(text_tokens) # [b, t] -> [b, t, h] bert_embedding = self.bert_embedding(bert_features) # [b, t, 1024] -> [b, t, h] text_embedding = text_embedding + bert_embedding # cond part spk_embedding = spk_embedding.unsqueeze(0) # [b, h] -> [b, 1, h] lm_input = torch.cat([spk_embedding, text_embedding, audio_embedding], dim=1) # if cfg_scale != 1.0: # print("infer with cfg_scale:", cfg_scale) # lm_cf_input = torch.cat([0*clap_embedding, 0*text_embedding, audio_embedding], dim=1) # lm_input = torch.cat([lm_input, lm_cf_input], dim=0) # 4. cal min/max_length max_len = text_tokens.shape[-1] # 5. step by step decode out_tokens = [] offset = 0 state = None for i in tqdm(range(int(max_len+1))): y_pred, _, state = self.llm.forward_one_step(lm_input.to(self.dtype), torch.ones(lm_input.shape[0], lm_input.shape[1], device=lm_input.device).to(torch.bool), cache=state) logits = self.llm_decoder(y_pred[:, -1]) # if cfg_scale != 1.0: # # perform context free guidance # logits_cf = logits[1] # logits = logits[0] # infer_cfg_ratio = cfg_scale # logits = infer_cfg_ratio * logits + (1 - infer_cfg_ratio) * logits_cf logp = logits.log_softmax(dim=-1) logp = logp.squeeze(dim=0) # if i < int(min_len): # logp[self.audio_token_size] = torch.tensor(float('-inf'), dtype=self.dtype) top_ids = sampling_method(logp, out_tokens, **sampling_kwargs).item() if top_ids == self.audio_token_size: print('end!') break # # in stream mode, yield token one by one current_token = torch.tensor([[top_ids]], dtype=torch.int64, device=device) res_tokens = torch.cat([res_tokens, current_token], dim=1) out_tokens.append(top_ids) offset += lm_input.size(1) lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1) # if cfg_scale != 1.0: # lm_input = lm_input.repeat(2, 1, 1) if res_tokens.shape[-1]!=max_len: print('infer error', res_tokens.shape, max_len) res_tokens = res_tokens[:,:max_len] res_tokens = torch.nn.functional.pad(res_tokens, (0, max_len-res_tokens.size(1)), value=0) return res_tokens if __name__ == '__main__': audio_token_size=1024 device = 'cuda' llm = LLM(audio_token_size=audio_token_size).to(device) print('n_params:', sum(p.numel() for p in llm.parameters() if p.requires_grad)/1e6, 'M') text = torch.randint(0, 1024, (2, 11)).to(device) bert = torch.randn(2, 11, 1024).to(device) spk = torch.randn(2, 256).to(device) prosody = torch.randint(0, 1024, (2, 11)).to(device) lengths = torch.tensor([4, 11]).to(device) llm.train() res = llm(text, bert, spk, prosody, lengths) loss = res['loss'] acc = res['acc'] print('loss:', loss, 'acc:', acc) loss.backward() for name, param in llm.named_parameters(): if param.grad is None: print('name:', name, 'has no grad') llm.eval() text = text[:1] bert = bert[:1] spk = spk[:1] prosody = prosody[:1] res = llm.inference(text, bert, spk, None, topk_sampling, {'top_k': 1024}) print('sampling res', res.shape)