xinference 1.1.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

xinference/thirdparty/cosyvoice/flow/length_regulator.py

@@ -49,13 +49,14 @@ class InterpolateRegulator(nn.Module):
         olens = ylens
         return out * mask, olens
 
-    def inference(self, x1, x2, mel_len1, mel_len2):
+    def inference(self, x1, x2, mel_len1, mel_len2, input_frame_rate=50):
         # in inference mode, interploate prompt token and token(head/mid/tail) seprately, so we can get a clear separation point of mel
         # x in (B, T, D)
         if x2.shape[1] > 40:
-            x2_head = F.interpolate(x2[:, :20].transpose(1, 2).contiguous(), size=34, mode='linear')
-            x2_mid = F.interpolate(x2[:, 20:-20].transpose(1, 2).contiguous(), size=mel_len2 - 34 * 2, mode='linear')
-            x2_tail = F.interpolate(x2[:, -20:].transpose(1, 2).contiguous(), size=34, mode='linear')
+            x2_head = F.interpolate(x2[:, :20].transpose(1, 2).contiguous(), size=int(20 / input_frame_rate * 22050 / 256), mode='linear')
+            x2_mid = F.interpolate(x2[:, 20:-20].transpose(1, 2).contiguous(), size=mel_len2 - int(20 / input_frame_rate * 22050 / 256) * 2,
+                                   mode='linear')
+            x2_tail = F.interpolate(x2[:, -20:].transpose(1, 2).contiguous(), size=int(20 / input_frame_rate * 22050 / 256), mode='linear')
             x2 = torch.concat([x2_head, x2_mid, x2_tail], dim=2)
         else:
             x2 = F.interpolate(x2.transpose(1, 2).contiguous(), size=mel_len2, mode='linear')

xinference/thirdparty/cosyvoice/hifigan/discriminator.py

@@ -0,0 +1,140 @@
+import torch
+import torch.nn as nn
+from torch.nn.utils import weight_norm
+from typing import List, Optional, Tuple
+from einops import rearrange
+from torchaudio.transforms import Spectrogram
+
+
+class MultipleDiscriminator(nn.Module):
+    def __init__(
+            self, mpd: nn.Module, mrd: nn.Module
+    ):
+        super().__init__()
+        self.mpd = mpd
+        self.mrd = mrd
+
+    def forward(self, y: torch.Tensor, y_hat: torch.Tensor):
+        y_d_rs, y_d_gs, fmap_rs, fmap_gs = [], [], [], []
+        this_y_d_rs, this_y_d_gs, this_fmap_rs, this_fmap_gs = self.mpd(y.unsqueeze(dim=1), y_hat.unsqueeze(dim=1))
+        y_d_rs += this_y_d_rs
+        y_d_gs += this_y_d_gs
+        fmap_rs += this_fmap_rs
+        fmap_gs += this_fmap_gs
+        this_y_d_rs, this_y_d_gs, this_fmap_rs, this_fmap_gs = self.mrd(y, y_hat)
+        y_d_rs += this_y_d_rs
+        y_d_gs += this_y_d_gs
+        fmap_rs += this_fmap_rs
+        fmap_gs += this_fmap_gs
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiResolutionDiscriminator(nn.Module):
+    def __init__(
+        self,
+        fft_sizes: Tuple[int, ...] = (2048, 1024, 512),
+        num_embeddings: Optional[int] = None,
+    ):
+        """
+        Multi-Resolution Discriminator module adapted from https://github.com/descriptinc/descript-audio-codec.
+        Additionally, it allows incorporating conditional information with a learned embeddings table.
+
+        Args:
+            fft_sizes (tuple[int]): Tuple of window lengths for FFT. Defaults to (2048, 1024, 512).
+            num_embeddings (int, optional): Number of embeddings. None means non-conditional discriminator.
+                Defaults to None.
+        """
+
+        super().__init__()
+        self.discriminators = nn.ModuleList(
+            [DiscriminatorR(window_length=w, num_embeddings=num_embeddings) for w in fft_sizes]
+        )
+
+    def forward(
+        self, y: torch.Tensor, y_hat: torch.Tensor, bandwidth_id: torch.Tensor = None
+    ) -> Tuple[List[torch.Tensor], List[torch.Tensor], List[List[torch.Tensor]], List[List[torch.Tensor]]]:
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+
+        for d in self.discriminators:
+            y_d_r, fmap_r = d(x=y, cond_embedding_id=bandwidth_id)
+            y_d_g, fmap_g = d(x=y_hat, cond_embedding_id=bandwidth_id)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorR(nn.Module):
+    def __init__(
+        self,
+        window_length: int,
+        num_embeddings: Optional[int] = None,
+        channels: int = 32,
+        hop_factor: float = 0.25,
+        bands: Tuple[Tuple[float, float], ...] = ((0.0, 0.1), (0.1, 0.25), (0.25, 0.5), (0.5, 0.75), (0.75, 1.0)),
+    ):
+        super().__init__()
+        self.window_length = window_length
+        self.hop_factor = hop_factor
+        self.spec_fn = Spectrogram(
+            n_fft=window_length, hop_length=int(window_length * hop_factor), win_length=window_length, power=None
+        )
+        n_fft = window_length // 2 + 1
+        bands = [(int(b[0] * n_fft), int(b[1] * n_fft)) for b in bands]
+        self.bands = bands
+        convs = lambda: nn.ModuleList(
+            [
+                weight_norm(nn.Conv2d(2, channels, (3, 9), (1, 1), padding=(1, 4))),
+                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
+                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
+                weight_norm(nn.Conv2d(channels, channels, (3, 9), (1, 2), padding=(1, 4))),
+                weight_norm(nn.Conv2d(channels, channels, (3, 3), (1, 1), padding=(1, 1))),
+            ]
+        )
+        self.band_convs = nn.ModuleList([convs() for _ in range(len(self.bands))])
+
+        if num_embeddings is not None:
+            self.emb = torch.nn.Embedding(num_embeddings=num_embeddings, embedding_dim=channels)
+            torch.nn.init.zeros_(self.emb.weight)
+
+        self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), (1, 1), padding=(1, 1)))
+
+    def spectrogram(self, x):
+        # Remove DC offset
+        x = x - x.mean(dim=-1, keepdims=True)
+        # Peak normalize the volume of input audio
+        x = 0.8 * x / (x.abs().max(dim=-1, keepdim=True)[0] + 1e-9)
+        x = self.spec_fn(x)
+        x = torch.view_as_real(x)
+        x = rearrange(x, "b f t c -> b c t f")
+        # Split into bands
+        x_bands = [x[..., b[0]: b[1]] for b in self.bands]
+        return x_bands
+
+    def forward(self, x: torch.Tensor, cond_embedding_id: torch.Tensor = None):
+        x_bands = self.spectrogram(x)
+        fmap = []
+        x = []
+        for band, stack in zip(x_bands, self.band_convs):
+            for i, layer in enumerate(stack):
+                band = layer(band)
+                band = torch.nn.functional.leaky_relu(band, 0.1)
+                if i > 0:
+                    fmap.append(band)
+            x.append(band)
+        x = torch.cat(x, dim=-1)
+        if cond_embedding_id is not None:
+            emb = self.emb(cond_embedding_id)
+            h = (emb.view(1, -1, 1, 1) * x).sum(dim=1, keepdims=True)
+        else:
+            h = 0
+        x = self.conv_post(x)
+        fmap.append(x)
+        x += h
+
+        return x, fmap

xinference/thirdparty/cosyvoice/hifigan/generator.py

@@ -14,7 +14,7 @@
 
 """HIFI-GAN"""
 
-import typing as tp
+from typing import Dict, Optional, List
 import numpy as np
 from scipy.signal import get_window
 import torch
@@ -38,13 +38,15 @@ This code is modified from https://github.com/jik876/hifi-gan
  https://github.com/NVIDIA/BigVGAN
 
 """
+
+
 class ResBlock(torch.nn.Module):
     """Residual block module in HiFiGAN/BigVGAN."""
     def __init__(
         self,
         channels: int = 512,
         kernel_size: int = 3,
-        dilations: tp.List[int] = [1, 3, 5],
+        dilations: List[int] = [1, 3, 5],
     ):
         super(ResBlock, self).__init__()
         self.convs1 = nn.ModuleList()
@@ -100,6 +102,7 @@ class ResBlock(torch.nn.Module):
             remove_weight_norm(self.convs1[idx])
             remove_weight_norm(self.convs2[idx])
 
+
 class SineGen(torch.nn.Module):
     """ Definition of sine generator
     SineGen(samp_rate, harmonic_num = 0,
@@ -231,13 +234,13 @@ class HiFTGenerator(nn.Module):
             nsf_alpha: float = 0.1,
             nsf_sigma: float = 0.003,
             nsf_voiced_threshold: float = 10,
-            upsample_rates: tp.List[int] = [8, 8],
-            upsample_kernel_sizes: tp.List[int] = [16, 16],
-            istft_params: tp.Dict[str, int] = {"n_fft": 16, "hop_len": 4},
-            resblock_kernel_sizes: tp.List[int] = [3, 7, 11],
-            resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
-            source_resblock_kernel_sizes: tp.List[int] = [7, 11],
-            source_resblock_dilation_sizes: tp.List[tp.List[int]] = [[1, 3, 5], [1, 3, 5]],
+            upsample_rates: List[int] = [8, 8],
+            upsample_kernel_sizes: List[int] = [16, 16],
+            istft_params: Dict[str, int] = {"n_fft": 16, "hop_len": 4},
+            resblock_kernel_sizes: List[int] = [3, 7, 11],
+            resblock_dilation_sizes: List[List[int]] = [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+            source_resblock_kernel_sizes: List[int] = [7, 11],
+            source_resblock_dilation_sizes: List[List[int]] = [[1, 3, 5], [1, 3, 5]],
             lrelu_slope: float = 0.1,
             audio_limit: float = 0.99,
             f0_predictor: torch.nn.Module = None,
@@ -286,8 +289,7 @@ class HiFTGenerator(nn.Module):
         self.source_resblocks = nn.ModuleList()
         downsample_rates = [1] + upsample_rates[::-1][:-1]
         downsample_cum_rates = np.cumprod(downsample_rates)
-        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes,
-                                          source_resblock_dilation_sizes)):
+        for i, (u, k, d) in enumerate(zip(downsample_cum_rates[::-1], source_resblock_kernel_sizes, source_resblock_dilation_sizes)):
             if u == 1:
                 self.source_downs.append(
                     Conv1d(istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), 1, 1)
@@ -304,7 +306,7 @@ class HiFTGenerator(nn.Module):
         self.resblocks = nn.ModuleList()
         for i in range(len(self.ups)):
             ch = base_channels // (2**(i + 1))
-            for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
+            for _, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
                 self.resblocks.append(ResBlock(ch, k, d))
 
         self.conv_post = weight_norm(Conv1d(ch, istft_params["n_fft"] + 2, 7, 1, padding=3))
@@ -314,11 +316,19 @@ class HiFTGenerator(nn.Module):
         self.stft_window = torch.from_numpy(get_window("hann", istft_params["n_fft"], fftbins=True).astype(np.float32))
         self.f0_predictor = f0_predictor
 
-    def _f02source(self, f0: torch.Tensor) -> torch.Tensor:
-        f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
-
-        har_source, _, _ = self.m_source(f0)
-        return har_source.transpose(1, 2)
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+        self.m_source.remove_weight_norm()
+        for l in self.source_downs:
+            remove_weight_norm(l)
+        for l in self.source_resblocks:
+            l.remove_weight_norm()
 
     def _stft(self, x):
         spec = torch.stft(
@@ -332,17 +342,11 @@ class HiFTGenerator(nn.Module):
         magnitude = torch.clip(magnitude, max=1e2)
         real = magnitude * torch.cos(phase)
         img = magnitude * torch.sin(phase)
-        inverse_transform = torch.istft(torch.complex(real, img), self.istft_params["n_fft"], self.istft_params["hop_len"], self.istft_params["n_fft"], window=self.stft_window.to(magnitude.device))
+        inverse_transform = torch.istft(torch.complex(real, img), self.istft_params["n_fft"], self.istft_params["hop_len"],
+                                        self.istft_params["n_fft"], window=self.stft_window.to(magnitude.device))
         return inverse_transform
 
-    def forward(self, x: torch.Tensor, cache_source: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
-        f0 = self.f0_predictor(x)
-        s = self._f02source(f0)
-
-        # use cache_source to avoid glitch
-        if cache_source.shape[2] == 0:
-            s[:, :, :cache_source.shape[2]] = cache_source
-
+    def decode(self, x: torch.Tensor, s: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
         s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
         s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
 
@@ -374,22 +378,34 @@ class HiFTGenerator(nn.Module):
 
         x = self._istft(magnitude, phase)
         x = torch.clamp(x, -self.audio_limit, self.audio_limit)
-        return x, s
+        return x
 
-    def remove_weight_norm(self):
-        print('Removing weight norm...')
-        for l in self.ups:
-            remove_weight_norm(l)
-        for l in self.resblocks:
-            l.remove_weight_norm()
-        remove_weight_norm(self.conv_pre)
-        remove_weight_norm(self.conv_post)
-        self.source_module.remove_weight_norm()
-        for l in self.source_downs:
-            remove_weight_norm(l)
-        for l in self.source_resblocks:
-            l.remove_weight_norm()
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        speech_feat = batch['speech_feat'].transpose(1, 2).to(device)
+        # mel->f0
+        f0 = self.f0_predictor(speech_feat)
+        # f0->source
+        s = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+        s, _, _ = self.m_source(s)
+        s = s.transpose(1, 2)
+        # mel+source->speech
+        generated_speech = self.decode(x=speech_feat, s=s)
+        return generated_speech, f0
 
     @torch.inference_mode()
-    def inference(self, mel: torch.Tensor, cache_source: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
-        return self.forward(x=mel, cache_source=cache_source)
+    def inference(self, speech_feat: torch.Tensor, cache_source: torch.Tensor = torch.zeros(1, 1, 0)) -> torch.Tensor:
+        # mel->f0
+        f0 = self.f0_predictor(speech_feat)
+        # f0->source
+        s = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+        s, _, _ = self.m_source(s)
+        s = s.transpose(1, 2)
+        # use cache_source to avoid glitch
+        if cache_source.shape[2] != 0:
+            s[:, :, :cache_source.shape[2]] = cache_source
+        generated_speech = self.decode(x=speech_feat, s=s)
+        return generated_speech, s

xinference/thirdparty/cosyvoice/hifigan/hifigan.py

@@ -0,0 +1,67 @@
+from typing import Dict, Optional
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from matcha.hifigan.models import feature_loss, generator_loss, discriminator_loss
+from cosyvoice.utils.losses import tpr_loss, mel_loss
+
+
+class HiFiGan(nn.Module):
+    def __init__(self, generator, discriminator, mel_spec_transform,
+                 multi_mel_spectral_recon_loss_weight=45, feat_match_loss_weight=2.0,
+                 tpr_loss_weight=1.0, tpr_loss_tau=0.04):
+        super(HiFiGan, self).__init__()
+        self.generator = generator
+        self.discriminator = discriminator
+        self.mel_spec_transform = mel_spec_transform
+        self.multi_mel_spectral_recon_loss_weight = multi_mel_spectral_recon_loss_weight
+        self.feat_match_loss_weight = feat_match_loss_weight
+        self.tpr_loss_weight = tpr_loss_weight
+        self.tpr_loss_tau = tpr_loss_tau
+
+    def forward(
+            self,
+            batch: dict,
+            device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        if batch['turn'] == 'generator':
+            return self.forward_generator(batch, device)
+        else:
+            return self.forward_discriminator(batch, device)
+
+    def forward_generator(self, batch, device):
+        real_speech = batch['speech'].to(device)
+        pitch_feat = batch['pitch_feat'].to(device)
+        # 1. calculate generator outputs
+        generated_speech, generated_f0 = self.generator(batch, device)
+        # 2. calculate discriminator outputs
+        y_d_rs, y_d_gs, fmap_rs, fmap_gs = self.discriminator(real_speech, generated_speech)
+        # 3. calculate generator losses, feature loss, mel loss, tpr losses [Optional]
+        loss_gen, _ = generator_loss(y_d_gs)
+        loss_fm = feature_loss(fmap_rs, fmap_gs)
+        loss_mel = mel_loss(real_speech, generated_speech, self.mel_spec_transform)
+        if self.tpr_loss_weight != 0:
+            loss_tpr = tpr_loss(y_d_rs, y_d_gs, self.tpr_loss_tau)
+        else:
+            loss_tpr = torch.zeros(1).to(device)
+        loss_f0 = F.l1_loss(generated_f0, pitch_feat)
+        loss = loss_gen + self.feat_match_loss_weight * loss_fm + \
+            self.multi_mel_spectral_recon_loss_weight * loss_mel + \
+            self.tpr_loss_weight * loss_tpr + loss_f0
+        return {'loss': loss, 'loss_gen': loss_gen, 'loss_fm': loss_fm, 'loss_mel': loss_mel, 'loss_tpr': loss_tpr, 'loss_f0': loss_f0}
+
+    def forward_discriminator(self, batch, device):
+        real_speech = batch['speech'].to(device)
+        # 1. calculate generator outputs
+        with torch.no_grad():
+            generated_speech, generated_f0 = self.generator(batch, device)
+        # 2. calculate discriminator outputs
+        y_d_rs, y_d_gs, fmap_rs, fmap_gs = self.discriminator(real_speech, generated_speech)
+        # 3. calculate discriminator losses, tpr losses [Optional]
+        loss_disc, _, _ = discriminator_loss(y_d_rs, y_d_gs)
+        if self.tpr_loss_weight != 0:
+            loss_tpr = tpr_loss(y_d_rs, y_d_gs, self.tpr_loss_tau)
+        else:
+            loss_tpr = torch.zeros(1).to(device)
+        loss = loss_disc + self.tpr_loss_weight * loss_tpr
+        return {'loss': loss, 'loss_disc': loss_disc, 'loss_tpr': loss_tpr}

xinference/thirdparty/cosyvoice/llm/llm.py

@@ -15,6 +15,7 @@ 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
@@ -80,7 +81,8 @@ class TransformerLM(torch.nn.Module):
     def pad_unpad_sequence(self, sos_eos_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_eos_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 = [torch.concat([sos_eos_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
@@ -104,7 +106,8 @@ class TransformerLM(torch.nn.Module):
         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 = [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
@@ -124,7 +127,8 @@ class TransformerLM(torch.nn.Module):
         speech_token = self.speech_embedding(speech_token)
 
         # 5. unpad and pad
-        lm_input, lm_input_len = self.pad_unpad_sequence(sos_eos_emb, embedding, text_token, text_token_len, task_id_emb, speech_token, speech_token_len)
+        lm_input, lm_input_len = self.pad_unpad_sequence(sos_eos_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))
@@ -194,14 +198,143 @@ class TransformerLM(torch.nn.Module):
         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=0, 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))
+            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)
+            # force continue decode first token
+            if i == 0:
+                logp[:, self.speech_token_size] = -float('inf')
             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.speech_token_size:
                 break
             # in stream mode, yield token one by one
-            yield torch.tensor([[top_ids]], dtype=torch.int64, device=device)
+            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_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(torch.nn.Module):
+    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,
+    ):
+        super().__init__()
+        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_eos = 0
+        self.task_id = 1
+        self.fill_token = 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
+
+    def sampling_ids(
+            self,
+            weighted_scores: torch.Tensor,
+            decoded_tokens: List,
+            sampling: int,
+            ignore_eos: bool = True,
+    ):
+        while True:
+            top_ids = self.sampling(weighted_scores, decoded_tokens, sampling)
+            if (not ignore_eos) or (self.speech_token_size not in top_ids):
+                break
+        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,
+    ) -> 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)
+
+        # 2. encode embedding
+        embedding = torch.zeros(1, 0, self.llm_input_size, dtype=text.dtype).to(device)
+
+        # 3. concat llm_input
+        sos_eos_emb = self.llm_embedding.weight[self.sos_eos].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_eos_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 = []
+        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).item()
+            if top_ids == self.speech_token_size:
+                break
+            if top_ids > self.speech_token_size:
+                continue
+            # 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)