CosyVoice/cosyvoice/llm/llm.py

# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
#               2025 Alibaba Inc (authors: Xiang Lyu, Yabin Li, Qihua, Shengqiang Li)
#
# 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 queue
import random
import time
import threading
from typing import Dict, Optional, Callable, List, Generator
import torch
from torch import nn
import torch.nn.functional as F
from transformers import Qwen2ForCausalLM
from torch.nn.utils.rnn import pad_sequence, unpad_sequence
from cosyvoice.utils.common import IGNORE_ID
from cosyvoice.transformer.label_smoothing_loss import LabelSmoothingLoss
from cosyvoice.utils.common import th_accuracy
from cosyvoice.utils.file_utils import logging
from cosyvoice.utils.mask import make_pad_mask


class TransformerLM(torch.nn.Module):
    def __init__(
            self,
            text_encoder_input_size: int,
            llm_input_size: int,
            llm_output_size: int,
            text_token_size: int,
            speech_token_size: int,
            text_encoder: torch.nn.Module,
            llm: torch.nn.Module,
            sampling: Callable,
            length_normalized_loss: bool = True,
            lsm_weight: float = 0.0,
            spk_embed_dim: int = 192,
    ):
        super().__init__()
        self.llm_input_size = llm_input_size
        self.speech_token_size = speech_token_size
        # 1. build text token inputs related modules
        self.text_embedding = torch.nn.Embedding(text_token_size, text_encoder_input_size)
        self.text_encoder = text_encoder
        self.text_encoder_affine_layer = nn.Linear(
            self.text_encoder.output_size(),
            llm_input_size
        )

        # 2. build speech token language model related modules
        self.sos = 0
        self.task_id = 1
        self.eos_token = self.speech_token_size
        self.llm_embedding = torch.nn.Embedding(2, llm_input_size)
        self.llm = llm
        self.llm_decoder = nn.Linear(llm_output_size, speech_token_size + 1)
        self.criterion_ce = LabelSmoothingLoss(
            size=speech_token_size + 1,
            padding_idx=IGNORE_ID,
            smoothing=lsm_weight,
            normalize_length=length_normalized_loss,
        )

        # 3. [Optional] build speech token related modules
        self.speech_embedding = torch.nn.Embedding(speech_token_size, llm_input_size)
        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, llm_input_size)

        # 4. sampling method
        self.sampling = sampling

    def encode(
            self,
            text: torch.Tensor,
            text_lengths: torch.Tensor,
    ):
        encoder_out, encoder_mask = self.text_encoder(text, text_lengths, decoding_chunk_size=1, num_decoding_left_chunks=-1)
        encoder_out_lens = encoder_mask.squeeze(1).sum(1)
        encoder_out = self.text_encoder_affine_layer(encoder_out)
        return encoder_out, encoder_out_lens

    def pad_unpad_sequence(self, sos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len):
        text_token = unpad_sequence(text_token, text_token_len.cpu(), batch_first=True)
        speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
        lm_input = [torch.concat([sos_emb.squeeze(dim=0), embedding[i], text_token[i], task_id_emb.squeeze(dim=0), speech_token[i]], dim=0)
                    for i in range(len(text_token))]
        lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32)
        lm_input = pad_sequence(lm_input, batch_first=True, padding_value=IGNORE_ID)
        return lm_input, lm_input_len

    def forward(
            self,
            batch: dict,
            device: torch.device,
    ) -> Dict[str, Optional[torch.Tensor]]:
        """
        Args:
            text: (B, L, D)
            text_lengths: (B,)
            audio: (B, T, N) or (B, T)
            audio_lengths: (B,)
        """
        text_token = batch['text_token'].to(device)
        text_token_len = batch['text_token_len'].to(device)
        speech_token = batch['speech_token'].to(device)
        speech_token_len = batch['speech_token_len'].to(device)
        embedding = batch['embedding'].to(device)

        # 1. prepare llm_target
        lm_target = [torch.tensor([IGNORE_ID] * (2 + text_token_len[i]) + speech_token[i, :speech_token_len[i]].tolist() +
                                  [self.speech_token_size]) for i in range(text_token.size(0))]
        lm_target = pad_sequence(lm_target, batch_first=True, padding_value=IGNORE_ID).to(device)

        # 1. encode text_token
        text_token = self.text_embedding(text_token)
        text_token, text_token_len = self.encode(text_token, text_token_len)

        # 2. embedding projection
        embedding = F.normalize(embedding, dim=1)
        embedding = self.spk_embed_affine_layer(embedding)
        embedding = embedding.unsqueeze(1)

        # 3. sos and task_id
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)

        # 4. encode speech_token
        speech_token = self.speech_embedding(speech_token)

        # 5. unpad and pad
        lm_input, lm_input_len = self.pad_unpad_sequence(sos_emb, embedding, text_token, text_token_len,
                                                         task_id_emb, speech_token, speech_token_len)

        # 6. run lm forward
        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
        logits = self.llm_decoder(lm_output)
        loss = self.criterion_ce(logits, lm_target)
        acc = th_accuracy(logits.view(-1, self.speech_token_size + 1), lm_target, ignore_label=IGNORE_ID)
        return {'loss': loss, 'acc': acc}

    def sampling_ids(
            self,
            weighted_scores: torch.Tensor,
            decoded_tokens: List,
            sampling: int,
            ignore_eos: bool = True,
    ):
        num_trials, max_trials = 0, 100
        while True:
            top_ids = self.sampling(weighted_scores, decoded_tokens, sampling)
            if (not ignore_eos) or (top_ids < self.speech_token_size):
                break
            num_trials += 1
            if num_trials > max_trials:
                raise RuntimeError('sampling reaches max_trials {} and still get eos when ignore_eos is True, check your input!'.format(max_trials))
        return top_ids

    @torch.inference_mode()
    def inference(
            self,
            text: torch.Tensor,
            text_len: torch.Tensor,
            prompt_text: torch.Tensor,
            prompt_text_len: torch.Tensor,
            prompt_speech_token: torch.Tensor,
            prompt_speech_token_len: torch.Tensor,
            embedding: torch.Tensor,
            sampling: int = 25,
            max_token_text_ratio: float = 20,
            min_token_text_ratio: float = 2,
            uuid: str = '',
    ) -> Generator[torch.Tensor, None, None]:
        device = text.device
        text = torch.concat([prompt_text, text], dim=1)
        text_len += prompt_text_len
        text = self.text_embedding(text)

        # 1. encode text
        text, text_len = self.encode(text, text_len)

        # 2. encode embedding
        if embedding.shape[0] != 0:
            embedding = F.normalize(embedding, dim=1)
            embedding = self.spk_embed_affine_layer(embedding)
            embedding = embedding.unsqueeze(dim=1)
        else:
            embedding = torch.zeros(1, 0, self.llm_input_size, dtype=text.dtype).to(device).to(text.dtype)

        # 3. concat llm_input
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
        if prompt_speech_token_len != 0:
            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
        else:
            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size, dtype=text.dtype).to(device)
        lm_input = torch.concat([sos_emb, embedding, text, task_id_emb, prompt_speech_token_emb], dim=1)

        # 4. cal min/max_length
        min_len = int((text_len - prompt_text_len) * min_token_text_ratio)
        max_len = int((text_len - prompt_text_len) * max_token_text_ratio)

        # 5. step by step decode
        out_tokens = []
        offset = 0
        att_cache, cnn_cache = torch.zeros((0, 0, 0, 0), device=lm_input.device), torch.zeros((0, 0, 0, 0), device=lm_input.device)
        for i in range(max_len):
            y_pred, att_cache, cnn_cache = self.llm.forward_chunk(lm_input, offset=offset, required_cache_size=-1,
                                                                  att_cache=att_cache, cnn_cache=cnn_cache,
                                                                  att_mask=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]),
                                                                                                 device=lm_input.device)).to(torch.bool))
            logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
            top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=True if i < min_len else False).item()
            if top_ids == self.eos_token:
                break
            # in stream mode, yield token one by one
            yield top_ids
            out_tokens.append(top_ids)
            offset += lm_input.size(1)
            lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)


class Qwen2Encoder(torch.nn.Module):
    def __init__(self, pretrain_path):
        super().__init__()
        self.model = Qwen2ForCausalLM.from_pretrained(pretrain_path)

    def forward(self, xs: torch.Tensor, xs_lens: torch.Tensor):
        T = xs.size(1)
        masks = ~make_pad_mask(xs_lens, T)
        outs = self.model(
            inputs_embeds=xs,
            attention_mask=masks,
            output_hidden_states=True,
            return_dict=True,
        )
        return outs.hidden_states[-1], masks.unsqueeze(1)

    def forward_one_step(self, xs, masks, cache=None):
        input_masks = masks[:, -1, :]
        outs = self.model(
            inputs_embeds=xs,
            attention_mask=input_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
        return xs, new_cache


class Qwen2LM(TransformerLM):
    def __init__(
            self,
            llm_input_size: int,
            llm_output_size: int,
            speech_token_size: int,
            llm: torch.nn.Module,
            sampling: Callable,
            length_normalized_loss: bool = True,
            lsm_weight: float = 0.0,
            mix_ratio: List[int] = [5, 15],
    ):
        torch.nn.Module.__init__(self)
        self.llm_input_size = llm_input_size
        self.llm_output_size = llm_output_size
        self.speech_token_size = speech_token_size
        # 2. build speech token language model related modules
        self.sos = 0
        self.task_id = 1
        self.eos_token = speech_token_size
        self.fill_token = speech_token_size + 2

        self.llm_embedding = torch.nn.Embedding(2, llm_input_size)
        self.llm = llm
        self.llm_decoder = nn.Linear(llm_output_size, speech_token_size + 3)
        self.criterion_ce = LabelSmoothingLoss(
            size=speech_token_size + 3,
            padding_idx=IGNORE_ID,
            smoothing=lsm_weight,
            normalize_length=length_normalized_loss,
        )

        # 3. [Optional] build speech token related modules
        self.speech_embedding = torch.nn.Embedding(speech_token_size + 3, llm_input_size)

        # 4. sampling method
        self.sampling = sampling
        self.mix_ratio = mix_ratio

        # 5. vllm related
        self.stop_token_ids = [speech_token_size + i for i in range(3)]
        self.vllm_output_queue = {}

    def prepare_lm_input_target(self, sos_emb, text_token, text_token_emb, text_token_len, task_id_emb, speech_token, speech_token_emb, speech_token_len):
        lm_target, lm_input = [], []
        text_token = unpad_sequence(text_token, text_token_len.cpu(), batch_first=True)
        speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
        text_token_emb = unpad_sequence(text_token_emb, text_token_len.cpu(), batch_first=True)
        speech_token_emb = unpad_sequence(speech_token_emb, speech_token_len.cpu(), batch_first=True)
        for i in range(len(text_token)):
            # bistream sequence
            if random.random() < 0.5 and speech_token_len[i] / text_token_len[i] > self.mix_ratio[1] / self.mix_ratio[0]:
                this_lm_target, this_lm_input = [], []
                this_lm_target.append(IGNORE_ID)
                this_lm_input.append(sos_emb.squeeze(dim=0))
                for j in range(((text_token_len[i] + 1) / self.mix_ratio[0]).ceil().int().item()):
                    this_text_token = text_token[i][j * self.mix_ratio[0]: (j + 1) * self.mix_ratio[0]].tolist()
                    this_speech_token = speech_token[i][j * self.mix_ratio[1]: (j + 1) * self.mix_ratio[1]].tolist()
                    if len(this_text_token) == self.mix_ratio[0]:
                        assert len(this_speech_token) == self.mix_ratio[1]
                        this_lm_target += [IGNORE_ID] * (self.mix_ratio[0] - 1)
                        this_lm_target += this_speech_token
                        this_lm_target.append(self.fill_token)
                        this_lm_input.append(text_token_emb[i][j * self.mix_ratio[0]: (j + 1) * self.mix_ratio[0]])
                        this_lm_input.append(speech_token_emb[i][j * self.mix_ratio[1]: (j + 1) * self.mix_ratio[1]])
                    else:
                        this_lm_target += [-1] * len(this_text_token)
                        this_lm_target += speech_token[i][j * self.mix_ratio[1]:].tolist()
                        this_lm_target.append(self.eos_token)
                        this_lm_input.append(text_token_emb[i][j * self.mix_ratio[0]:])
                        this_lm_input.append(task_id_emb.squeeze(dim=0))
                        this_lm_input.append(speech_token_emb[i][j * self.mix_ratio[1]:])
                this_lm_target, this_lm_input = torch.tensor(this_lm_target), torch.concat(this_lm_input, dim=0)
            # unistream sequence
            else:
                this_lm_target = torch.tensor([IGNORE_ID] * (1 + text_token_len[i]) + speech_token[i].tolist() + [self.eos_token])
                this_lm_input = torch.concat([sos_emb.squeeze(dim=0), text_token_emb[i], task_id_emb.squeeze(dim=0), speech_token_emb[i]], dim=0)
            lm_target.append(this_lm_target)
            lm_input.append(this_lm_input)
        lm_input_len = torch.tensor([i.size(0) for i in lm_input], dtype=torch.int32)
        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)
        return lm_target, lm_input, lm_input_len

    def forward(
            self,
            batch: dict,
            device: torch.device,
    ) -> Dict[str, Optional[torch.Tensor]]:
        """
        Args:
            text: (B, L, D)
            text_lengths: (B,)
            audio: (B, T, N) or (B, T)
            audio_lengths: (B,)
        """
        text_token = batch['text_token'].to(device)
        text_token_len = batch['text_token_len'].to(device)
        speech_token = batch['speech_token'].to(device)
        speech_token_len = batch['speech_token_len'].to(device)

        # 1. encode text_token
        text_token_emb = self.llm.model.model.embed_tokens(text_token)

        # 3. sos and task_id
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)

        # 2. encode speech_token
        speech_token_emb = self.speech_embedding(speech_token)

        # 3. prepare llm_input/target
        lm_target, lm_input, lm_input_len = self.prepare_lm_input_target(sos_emb, text_token, text_token_emb, text_token_len, task_id_emb, speech_token, speech_token_emb, speech_token_len)
        lm_target = lm_target.to(device)

        # 4. run lm forward
        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
        logits = self.llm_decoder(lm_output)
        loss = self.criterion_ce(logits, lm_target.to(device))
        acc = th_accuracy(logits.view(-1, self.speech_token_size + 3), lm_target, ignore_label=IGNORE_ID)
        return {'loss': loss, 'acc': acc}

    def forward_dpo(
            self,
            batch: dict,
            device: torch.device,
    ) -> Dict[str, Optional[torch.Tensor]]:
        text_token = batch['text_token'].to(device)
        text_token_len = batch['text_token_len'].to(device)
        speech_token = batch['speech_token'].to(device)
        speech_token_len = batch['speech_token_len'].to(device)
        reject_speech_token = batch['reject_speech_token'].to(device)
        reject_speech_token_len = batch['reject_speech_token_len'].to(device)

        # 1. encode text_token
        text_token_emb = self.llm.model.model.embed_tokens(text_token)

        # 3. sos and task_id
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)

        # 2. encode speech_token
        speech_token = unpad_sequence(speech_token, speech_token_len.cpu(), batch_first=True)
        reject_speech_token = unpad_sequence(reject_speech_token, reject_speech_token_len.cpu(), batch_first=True)
        speech_token_combined = speech_token + reject_speech_token
        speech_token_combined = pad_sequence(speech_token_combined, batch_first=True, padding_value=0)
        speech_token_combined_len = torch.concat([speech_token_len, reject_speech_token_len], dim=0)
        speech_token_combined_emb = self.speech_embedding(speech_token_combined)

        # 3. prepare llm_input/target
        lm_target, lm_input, lm_input_len = self.prepare_lm_input_target(sos_emb, text_token.repeat(2, 1), text_token_emb.repeat(2, 1, 1), text_token_len.repeat(2),
                                                                         task_id_emb, speech_token_combined, speech_token_combined_emb, speech_token_combined_len)
        lm_target = lm_target.to(device)

        # 4. run lm forward
        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
        logits = self.llm_decoder(lm_output)
        chosen_logits = logits[:text_token.shape[0]]
        rejected_logits = logits[text_token.shape[0]:]
        chosen_lm_target = lm_target[:text_token.shape[0]]
        rejected_lm_target = lm_target[text_token.shape[0]:]
        loss = self.criterion_ce(chosen_logits, chosen_lm_target.to(device))
        acc = th_accuracy(chosen_logits.view(-1, self.speech_token_size + 3), chosen_lm_target, ignore_label=IGNORE_ID)

        # 5. calculate dpo logits
        chosen_lm_mask = chosen_lm_target == IGNORE_ID
        rejected_lm_mask = rejected_lm_target == IGNORE_ID
        chosen_logps = torch.gather(chosen_logits.log_softmax(dim=-1), dim=2, index=chosen_lm_target.masked_fill(chosen_lm_mask, 0).unsqueeze(dim=-1)).squeeze(dim=-1)
        rejected_logps = torch.gather(rejected_logits.log_softmax(dim=-1), dim=2, index=rejected_lm_target.masked_fill(rejected_lm_mask, 0).unsqueeze(dim=-1)).squeeze(dim=-1)
        chosen_logps = (chosen_logps * chosen_lm_mask).sum(dim=-1) / chosen_lm_mask.sum(dim=-1)
        rejected_logps = (rejected_logps * rejected_lm_mask).sum(dim=-1) / rejected_lm_mask.sum(dim=-1)
        return {'loss': loss, 'acc': acc, 'chosen_logps': chosen_logps, 'rejected_logps': rejected_logps}

    @torch.inference_mode()
    def inference(
            self,
            text: torch.Tensor,
            text_len: torch.Tensor,
            prompt_text: torch.Tensor,
            prompt_text_len: torch.Tensor,
            prompt_speech_token: torch.Tensor,
            prompt_speech_token_len: torch.Tensor,
            embedding: torch.Tensor,
            sampling: int = 25,
            max_token_text_ratio: float = 20,
            min_token_text_ratio: float = 2,
            uuid: str = '',
    ) -> Generator[torch.Tensor, None, None]:
        device = text.device
        text = torch.concat([prompt_text, text], dim=1)
        text_len += prompt_text_len
        text = self.llm.model.model.embed_tokens(text)

        # 3. concat llm_input
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
        if prompt_speech_token_len != 0:
            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
        else:
            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size, dtype=text.dtype).to(device)
        lm_input = torch.concat([sos_emb, text, task_id_emb, prompt_speech_token_emb], dim=1)

        # 4. cal min/max_length
        min_len = int((text_len - prompt_text_len) * min_token_text_ratio)
        max_len = int((text_len - prompt_text_len) * max_token_text_ratio)

        # 5. step by step decode
        for token in self.inference_wrapper(lm_input, sampling, min_len, max_len, uuid):
            yield token

    @torch.inference_mode()
    def inference_wrapper(self, lm_input, sampling, min_len, max_len, uuid):
        if hasattr(self, 'vllm'):
            from vllm import SamplingParams, RequestOutput
            sampling_params = SamplingParams(top_k=sampling,
                                             stop_token_ids=self.stop_token_ids,
                                             min_tokens=min_len,
                                             max_tokens=max_len)
            with self.lock:
                self.vllm.add_request(uuid, {"prompt_embeds": lm_input.squeeze(0).to(torch.bfloat16).to(lm_input.device)}, sampling_params)
                self.vllm_output_queue[uuid] = queue.Queue()
            out_tokens = []
            while True:
                with self.lock:
                    if self.vllm_output_queue[uuid].empty() is True:
                        request_outputs: List[RequestOutput] = self.vllm.step()
                        for request_output in request_outputs:
                            top_ids = list(request_output.outputs[0].token_ids)[-1]
                            self.vllm_output_queue[request_output.request_id].put(top_ids)
                if self.vllm_output_queue[uuid].empty() is False:
                    top_ids = self.vllm_output_queue[uuid].get()
                    if top_ids in self.stop_token_ids:
                        break
                    # in stream mode, yield token one by one
                    yield top_ids
                    out_tokens.append(top_ids)
                    if len(out_tokens) == max_len:
                        break
                time.sleep(0.001)
            with self.lock:
                self.vllm_output_queue.pop(uuid)
        else:
            out_tokens = []
            cache = None
            for i in range(max_len):
                y_pred, cache = self.llm.forward_one_step(lm_input,
                                                          masks=torch.tril(torch.ones((1, lm_input.shape[1], lm_input.shape[1]), device=lm_input.device)).to(torch.bool),
                                                          cache=cache)
                logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
                top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=True if i < min_len else False)
                if top_ids in self.stop_token_ids:
                    break
                # in stream mode, yield token one by one
                yield top_ids
                out_tokens.append(top_ids)
                lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)

    @torch.inference_mode()
    def inference_bistream(
            self,
            text: Generator,
            prompt_text: torch.Tensor,
            prompt_text_len: torch.Tensor,
            prompt_speech_token: torch.Tensor,
            prompt_speech_token_len: torch.Tensor,
            embedding: torch.Tensor,
            sampling: int = 25,
            max_token_text_ratio: float = 20,
            min_token_text_ratio: float = 2,
    ) -> Generator[torch.Tensor, None, None]:

        device = prompt_text.device
        # 1. prepare input
        sos_emb = self.llm_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.llm_embedding.weight[self.task_id].reshape(1, 1, -1)
        if prompt_speech_token_len != 0:
            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
        else:
            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size, dtype=prompt_text.dtype).to(device)
        lm_input = torch.concat([sos_emb], dim=1)

        # 2. iterate text
        out_tokens = []
        cache = None
        # NOTE init prompt_text as text_cache as it is basically impossible prompt_speech_token/prompt_text < 15/5
        text_cache = self.llm.model.model.embed_tokens(prompt_text)
        next_fill_index = -1
        for this_text in text:
            text_cache = torch.concat([text_cache, self.llm.model.model.embed_tokens(this_text)], dim=1)
            # prompt_speech_token_emb not empty, try append to lm_input
            while prompt_speech_token_emb.size(1) != 0:
                if text_cache.size(1) >= self.mix_ratio[0]:
                    lm_input_text, lm_input_speech = text_cache[:, :self.mix_ratio[0]], prompt_speech_token_emb[:, :self.mix_ratio[1]]
                    logging.info('append {} text token {} speech token'.format(lm_input_text.size(1), lm_input_speech.size(1)))
                    lm_input = torch.concat([lm_input, lm_input_text, lm_input_speech], dim=1)
                    text_cache, prompt_speech_token_emb = text_cache[:, self.mix_ratio[0]:], prompt_speech_token_emb[:, self.mix_ratio[1]:]
                else:
                    logging.info('not enough text token to decode, wait for more')
                    break
            # no prompt_speech_token_emb remain, can decode some speech token
            if prompt_speech_token_emb.size(1) == 0:
                if (len(out_tokens) != 0 and out_tokens[-1] == self.fill_token) or (len(out_tokens) == 0 and lm_input.size(1) == 1):
                    logging.info('get fill token, need to append more text token')
                    if text_cache.size(1) >= self.mix_ratio[0]:
                        lm_input_text = text_cache[:, :self.mix_ratio[0]]
                        logging.info('append {} text token'.format(lm_input_text.size(1)))
                        if len(out_tokens) != 0 and out_tokens[-1] == self.fill_token:
                            lm_input = lm_input_text
                        else:
                            lm_input = torch.concat([lm_input, lm_input_text], dim=1)
                        text_cache = text_cache[:, self.mix_ratio[0]:]
                    else:
                        logging.info('not enough text token to decode, wait for more')
                        continue
                while True:
                    seq_len = lm_input.shape[1] if cache is None else lm_input.shape[1] + cache[0][0].size(2)
                    y_pred, cache = self.llm.forward_one_step(lm_input,
                                                              masks=torch.tril(torch.ones((1, seq_len, seq_len), device=lm_input.device)).to(torch.bool),
                                                              cache=cache)
                    logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
                    if next_fill_index != -1 and len(out_tokens) == next_fill_index:
                        top_ids = self.fill_token
                        next_fill_index += (self.mix_ratio[1] + 1)
                    else:
                        top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=True).item()
                    if top_ids == self.fill_token:
                        next_fill_index = len(out_tokens) + self.mix_ratio[1] + 1
                        logging.info('fill_token index {} next fill_token index {}'.format(len(out_tokens), next_fill_index))
                    out_tokens.append(top_ids)
                    if top_ids >= self.speech_token_size:
                        if top_ids == self.fill_token:
                            break
                        else:
                            raise ValueError('should not get token {}'.format(top_ids))
                    yield top_ids
                    lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)

        # 3. final decode
        lm_input = torch.concat([lm_input, text_cache, task_id_emb], dim=1)
        logging.info('no more text token, decode until met eos')
        while True:
            seq_len = lm_input.shape[1] if cache is None else lm_input.shape[1] + cache[0][0].size(2)
            y_pred, cache = self.llm.forward_one_step(lm_input,
                                                      masks=torch.tril(torch.ones((1, seq_len, seq_len), device=lm_input.device)).to(torch.bool),
                                                      cache=cache)
            logp = self.llm_decoder(y_pred[:, -1]).log_softmax(dim=-1)
            top_ids = self.sampling_ids(logp.squeeze(dim=0), out_tokens, sampling, ignore_eos=False).item()
            out_tokens.append(top_ids)
            if top_ids >= self.speech_token_size:
                if top_ids == self.eos_token:
                    break
                else:
                    raise ValueError('should not get token {}'.format(top_ids))
            # in stream mode, yield token one by one
            yield top_ids
            lm_input = self.speech_embedding.weight[top_ids].reshape(1, 1, -1)


class CosyVoice3LM(Qwen2LM):
    def __init__(
            self,
            llm_input_size: int,
            llm_output_size: int,
            speech_token_size: int,
            llm: torch.nn.Module,
            sampling: Callable,
            length_normalized_loss: bool = True,
            lsm_weight: float = 0.0,
            mix_ratio: List[int] = [5, 15],
    ):
        torch.nn.Module.__init__(self)
        self.llm_input_size = llm_input_size
        self.llm_output_size = llm_output_size
        self.speech_token_size = speech_token_size
        # 2. build speech token language model related modules
        self.sos = speech_token_size + 0
        self.eos_token = speech_token_size + 1
        self.task_id = speech_token_size + 2
        self.fill_token = speech_token_size + 3

        self.llm = llm
        self.llm_decoder = nn.Linear(llm_output_size, speech_token_size + 200, bias=False)
        self.criterion_ce = LabelSmoothingLoss(
            size=speech_token_size + 200,
            padding_idx=IGNORE_ID,
            smoothing=lsm_weight,
            normalize_length=length_normalized_loss,
        )

        # 3. [Optional] build speech token related modules
        self.speech_embedding = torch.nn.Embedding(speech_token_size + 200, llm_input_size)

        # 4. sampling method
        self.sampling = sampling
        self.mix_ratio = mix_ratio

        # 5. vllm related
        self.stop_token_ids = [speech_token_size + i for i in range(200)]
        self.vllm_output_queue = {}

    def forward(
            self,
            batch: dict,
            device: torch.device,
    ) -> Dict[str, Optional[torch.Tensor]]:
        """
        Args:
            text: (B, L, D)
            text_lengths: (B,)
            audio: (B, T, N) or (B, T)
            audio_lengths: (B,)
        """
        text_token = batch['text_token'].to(device)
        text_token_len = batch['text_token_len'].to(device)
        speech_token = batch['speech_token'].to(device)
        speech_token_len = batch['speech_token_len'].to(device)

        # 1. encode text_token
        text_token_emb = self.llm.model.model.embed_tokens(text_token)

        # 3. sos and task_id
        sos_emb = self.speech_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.speech_embedding.weight[self.task_id].reshape(1, 1, -1)

        # 2. encode speech_token
        speech_token_emb = self.speech_embedding(speech_token)

        # 3. prepare llm_input/target
        lm_target, lm_input, lm_input_len = self.prepare_lm_input_target(sos_emb, text_token, text_token_emb, text_token_len, task_id_emb, speech_token, speech_token_emb, speech_token_len)
        lm_target = lm_target.to(device)

        # 4. run lm forward
        lm_output, lm_output_mask = self.llm(lm_input, lm_input_len.to(device))
        logits = self.llm_decoder(lm_output)
        loss = self.criterion_ce(logits, lm_target.to(device))
        acc = th_accuracy(logits.view(-1, self.speech_token_size + 3), lm_target, ignore_label=IGNORE_ID)
        return {'loss': loss, 'acc': acc}

    @torch.inference_mode()
    def inference(
            self,
            text: torch.Tensor,
            text_len: torch.Tensor,
            prompt_text: torch.Tensor,
            prompt_text_len: torch.Tensor,
            prompt_speech_token: torch.Tensor,
            prompt_speech_token_len: torch.Tensor,
            embedding: torch.Tensor,
            sampling: int = 25,
            max_token_text_ratio: float = 20,
            min_token_text_ratio: float = 2,
            uuid: str = '',
    ) -> Generator[torch.Tensor, None, None]:
        device = text.device
        text = torch.concat([prompt_text, text], dim=1)
        text_len += prompt_text_len
        text = self.llm.model.model.embed_tokens(text)

        # 3. concat llm_input
        sos_emb = self.speech_embedding.weight[self.sos].reshape(1, 1, -1)
        task_id_emb = self.speech_embedding.weight[self.task_id].reshape(1, 1, -1)
        if prompt_speech_token_len != 0:
            prompt_speech_token_emb = self.speech_embedding(prompt_speech_token)
        else:
            prompt_speech_token_emb = torch.zeros(1, 0, self.llm_input_size, dtype=text.dtype).to(device)
        lm_input = torch.concat([sos_emb, text, task_id_emb, prompt_speech_token_emb], dim=1)

        # 4. cal min/max_length
        min_len = int((text_len - prompt_text_len) * min_token_text_ratio)
        max_len = int((text_len - prompt_text_len) * max_token_text_ratio)

        # 5. step by step decode
        for token in self.inference_wrapper(lm_input, sampling, min_len, max_len, uuid):
            yield token