initial

.gitattributes

@@ -1,35 +1,7 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+assets/user.png filter=lfs diff=lfs merge=lfs -text
+assets/assistant.png filter=lfs diff=lfs merge=lfs -text
+speakers/闫雨婷_prompt.wav filter=lfs diff=lfs merge=lfs -text
+speakers/闫雨婷RAP_prompt.wav filter=lfs diff=lfs merge=lfs -text
+speakers/闫雨婷VOCAL_prompt.wav filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+__pycache__/

Dockerfile

@@ -0,0 +1,46 @@
+FROM nvidia/cuda:12.1.0-base-ubuntu20.04
+
+ENV TZ=Asia/Shanghai
+RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime \
+    && echo $TZ > /etc/timezone
+
+RUN apt-get update \
+    && apt-get install -y build-essential \
+    && apt-get install -y wget \
+    && apt-get install -y software-properties-common curl zip unzip git-lfs awscli libssl-dev openssh-server vim \
+    && apt-get install -y net-tools iputils-ping iproute2
+
+RUN apt-get install --reinstall ca-certificates && update-ca-certificates
+
+RUN add-apt-repository -y 'ppa:deadsnakes/ppa' && apt update
+RUN apt install python3.10 python3.10-dev python3.10-distutils python3.10-venv -y \
+    && apt-get clean \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN wget -qO- https://bootstrap.pypa.io/get-pip.py | python3.10
+RUN ln -s /usr/bin/python3.10 /usr/bin/python
+RUN pip uninstall -y Pillow && pip install pillow
+
+# https://huggingface.co/docs/hub/spaces-sdks-docker#permissions
+RUN useradd -m -u 1000 user
+USER user
+
+ENV HOME="/home/user" \
+    PATH="/home/user/.local/bin:${PATH}"
+
+RUN python3.10 -m pip install pipx
+RUN pipx install poetry
+
+RUN poetry --version || { echo 'Poetry installation check failed' ; exit 1; }
+
+WORKDIR /workspace
+
+COPY --chown=user requirements.txt .
+RUN pip install -r requirements.txt
+
+COPY --chown=user . .
+
+RUN pip install gradio
+RUN chmod +x clone_hf_model.sh
+ENV HF_MODEL_PATH="/tmp/hf_model"
+CMD ["./clone_hf_model.sh", "$HF_MODEL_PATH", "&&", "python", "app.py", "--model", "$HF_MODEL_PATH"]

app.py

@@ -0,0 +1,173 @@
+import gradio as gr
+import time
+from pathlib import Path
+import torchaudio
+from stepaudio import StepAudio
+
+from funasr import AutoModel
+from funasr.utils.postprocess_utils import rich_transcription_postprocess
+
+CACHE_DIR = "/tmp/gradio/"
+system_promtp = {"role": "system", "content": "适配用户的语言，用简短口语化的文字回答"}
+
+
+class CustomAsr:
+    def __init__(self, model_name="iic/SenseVoiceSmall", device="cuda"):
+        self.model = AutoModel(
+            model=model_name,
+            vad_model="fsmn-vad",
+            vad_kwargs={"max_single_segment_time": 30000},
+            device=device,
+        )
+
+    def run(self, audio_path):
+        res = self.model.generate(
+            input=audio_path,
+            cache={},
+            language="auto",  # "zh", "en", "yue", "ja", "ko", "nospeech"
+            use_itn=True,
+            batch_size_s=60,
+            merge_vad=True,  #
+            merge_length_s=15,
+        )
+        text = rich_transcription_postprocess(res[0]["text"])
+        return text
+
+
+def add_message(chatbot, history, mic, text, asr_model):
+    if not mic and not text:
+        return chatbot, history, "Input is empty"
+
+    if text:
+        chatbot.append({"role": "user", "content": text})
+        history.append({"role": "user", "content": text})
+    elif mic and Path(mic).exists():
+        chatbot.append({"role": "user", "content": {"path": mic}})
+        # 使用用户语音的 asr 结果为了加速推理
+        text = asr_model.run(mic)
+        chatbot.append({"role": "user", "content": text})
+        history.append({"role": "user", "content": text})
+
+    print(f"{history=}")
+    return chatbot, history, None
+
+
+def reset_state():
+    """Reset the chat history."""
+    return [], [system_promtp]
+
+
+def save_tmp_audio(audio, sr):
+    import tempfile
+
+    with tempfile.NamedTemporaryFile(
+        dir=CACHE_DIR, delete=False, suffix=".wav"
+    ) as temp_audio:
+        temp_audio_path = temp_audio.name
+        torchaudio.save(temp_audio_path, audio, sr)
+
+    return temp_audio.name
+
+
+def predict(chatbot, history, audio_model):
+    """Generate a response from the model."""
+    try:
+        text, audio, sr = audio_model(history, "闫雨婷")
+        print(f"predict {text=}")
+        audio_path = save_tmp_audio(audio, sr)
+        chatbot.append({"role": "assistant", "content": {"path": audio_path}})
+        chatbot.append({"role": "assistant", "content": text})
+        history.append({"role": "assistant", "content": text})
+    except Exception as e:
+        print(e)
+        gr.Warning(f"Some error happend, retry submit")
+    return chatbot, history
+
+
+def _launch_demo(args, audio_model, asr_model):
+    with gr.Blocks(delete_cache=(86400, 86400)) as demo:
+        gr.Markdown("""<center><font size=8>Step Audio Chat</center>""")
+        chatbot = gr.Chatbot(
+            elem_id="chatbot",
+            avatar_images=["assets/user.png", "assets/assistant.png"],
+            min_height=800,
+            type="messages",
+        )
+        # 保存 chat 历史，不需要每次再重新拼格式
+        history = gr.State([system_promtp])
+        mic = gr.Audio(type="filepath")
+        text = gr.Textbox(placeholder="Enter message ...")
+
+        with gr.Row():
+            clean_btn = gr.Button("🧹 Clear History (清除历史)")
+            regen_btn = gr.Button("🤔️ Regenerate (重试)")
+            submit_btn = gr.Button("🚀 Submit")
+
+        def on_submit(chatbot, history, mic, text):
+            chatbot, history, error = add_message(
+                chatbot, history, mic, text, asr_model
+            )
+            if error:
+                gr.Warning(error)  # 显示警告消息
+                return chatbot, history, None, None
+            else:
+                chatbot, history = predict(chatbot, history, audio_model)
+                return chatbot, history, None, None
+
+        submit_btn.click(
+            fn=on_submit,
+            inputs=[chatbot, history, mic, text],
+            outputs=[chatbot, history, mic, text],
+            concurrency_limit=4,
+            concurrency_id="gpu_queue",
+        )
+        clean_btn.click(
+            reset_state,
+            outputs=[chatbot, history],
+            show_progress=True,
+        )
+
+        def regenerate(chatbot, history):
+            while chatbot and chatbot[-1]["role"] == "assistant":
+                chatbot.pop()
+            while history and history[-1]["role"] == "assistant":
+                print(f"discard {history[-1]}")
+                history.pop()
+            return predict(chatbot, history, audio_model)
+
+        regen_btn.click(
+            regenerate,
+            [chatbot, history],
+            [chatbot, history],
+            show_progress=True,
+            concurrency_id="gpu_queue",
+        )
+
+    demo.queue().launch(
+        share=False,
+        server_port=args.server_port,
+        server_name=args.server_name,
+    )
+
+
+if __name__ == "__main__":
+    from argparse import ArgumentParser
+    import os
+
+    parser = ArgumentParser()
+    parser.add_argument("--model-path", type=str, required=True, help="Model path.")
+    parser.add_argument(
+        "--server-port", type=int, default=7860, help="Demo server port."
+    )
+    parser.add_argument(
+        "--server-name", type=str, default="0.0.0.0", help="Demo server name."
+    )
+    args = parser.parse_args()
+
+    audio_model = StepAudio(
+        tokenizer_path=os.path.join(args.model_path, "Step-Audio-Tokenizer"),
+        tts_path=os.path.join(args.model_path, "Step-Audio-TTS-3B"),
+        llm_path=os.path.join(args.model_path, "Step-Audio-Chat"),
+    )
+    asr_model = CustomAsr()
+    _launch_demo(args, audio_model, asr_model)

clone_hf_model.sh

@@ -0,0 +1,57 @@
+#!/bin/bash
+
+if [ -z "$HF_USER_NAME" ]; then
+  echo "错误：环境变量 HF_USER_NAME 未设置！"
+  exit 1
+fi
+
+if [ -z "$HF_USER_TOKEN" ]; then
+  echo "错误：环境变量 HF_USER_TOKEN 未设置！"
+  exit 1
+fi
+
+# 启用Git LFS支持
+git lfs install --force
+
+# 定义需要克隆的仓库列表
+BASE_REPO_URL="https://${HF_USER_NAME}:${HF_USER_TOKEN}@huggingface.co/stepfun-ai"
+REPOSITORIES=(
+    "Step-Audio-Tokenizer"
+    "Step-Audio-TTS-3B"
+    "Step-Audio-Chat"
+)
+
+# 定义本地存放仓库的目录，默认为当前目录
+LOCAL_DIR="${1:-$(pwd)}"
+
+# 克隆函数（带无限重试机制）
+clone_with_retry() {
+    local repo_name=$1
+    local repo_url="${BASE_REPO_URL}/${repo_name}"
+    local target_dir="${LOCAL_DIR}/${repo_name}"
+
+    # 检查是否已存在目录
+    if [ -d "${target_dir}" ]; then
+        echo "目录 ${target_dir} 已存在，跳过克隆。"
+        return 0
+    fi
+
+    # 无限重试循环
+    while true; do
+        echo "正在尝试克隆 ${repo_name} 到 ${target_dir}..."
+        if git clone "${repo_url}" "${target_dir}"; then
+            echo "成功克隆 ${repo_name} 到 ${target_dir}"
+            return 0
+        else
+            echo "克隆失败, 5秒后重试..."
+            sleep 5
+        fi
+    done
+}
+
+# 遍历所有仓库进行克隆
+for repo in "${REPOSITORIES[@]}"; do
+    clone_with_retry "${repo}"
+done
+
+echo "所有仓库已成功下载！"

cosyvoice/cli/cosyvoice.py

@@ -0,0 +1,68 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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 os
+import uuid
+import time
+from tqdm import tqdm
+import torch
+import torchaudio
+from hyperpyyaml import load_hyperpyyaml
+from cosyvoice.cli.frontend import CosyVoiceFrontEnd
+from cosyvoice.cli.model import CosyVoiceModel
+
+
+class CosyVoice:
+
+    def __init__(
+        self,
+        model_dir,
+    ):
+        self.model_dir = model_dir
+        with open("{}/cosyvoice.yaml".format(model_dir), "r") as f:
+            configs = load_hyperpyyaml(f)
+        self.frontend = CosyVoiceFrontEnd(
+            configs["feat_extractor"],
+            "{}/campplus.onnx".format(model_dir),
+            "{}/speech_tokenizer_v1.onnx".format(model_dir),
+        )
+        self.model = CosyVoiceModel(configs["flow"], configs["hift"])
+        self.model.load(
+            "{}/flow.pt".format(model_dir),
+            "{}/hift.pt".format(model_dir),
+        )
+        self.model.flow = self.model.flow.to(torch.bfloat16)
+        del configs
+
+    def token_to_wav_offline(
+        self,
+        speech_token,
+        speech_feat,
+        speech_feat_len,
+        prompt_token,
+        prompt_token_len,
+        embedding,
+    ):
+        tts_mel = self.model.flow.inference(
+            token=speech_token.to(self.model.device),
+            token_len=torch.tensor([speech_token.size(1)], dtype=torch.int32).to(
+                self.model.device
+            ),
+            prompt_token=prompt_token.to(self.model.device),
+            prompt_token_len=prompt_token_len.to(self.model.device),
+            prompt_feat=speech_feat.to(self.model.device),
+            prompt_feat_len=speech_feat_len.to(self.model.device),
+            embedding=embedding.to(self.model.device),
+        )
+        tts_speech = self.model.hift.inference(mel=tts_mel.float())[0].cpu()
+        return tts_speech

cosyvoice/cli/frontend.py

@@ -0,0 +1,106 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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 onnxruntime
+import torch
+import numpy as np
+import whisper
+from typing import Callable
+import torchaudio.compliance.kaldi as kaldi
+
+
+class CosyVoiceFrontEnd:
+
+    def __init__(
+        self,
+        feat_extractor: Callable,
+        campplus_model: str,
+        speech_tokenizer_model: str,
+    ):
+        self.feat_extractor = feat_extractor
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        option = onnxruntime.SessionOptions()
+        option.graph_optimization_level = (
+            onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+        )
+        option.intra_op_num_threads = 1
+        self.campplus_session = onnxruntime.InferenceSession(
+            campplus_model, sess_options=option, providers=["CPUExecutionProvider"]
+        )
+        self.speech_tokenizer_session = onnxruntime.InferenceSession(
+            speech_tokenizer_model,
+            sess_options=option,
+            providers=[
+                (
+                    "CUDAExecutionProvider"
+                    if torch.cuda.is_available()
+                    else "CPUExecutionProvider"
+                )
+            ],
+        )
+
+    def _extract_speech_token(self, speech):
+        assert (
+            speech.shape[1] / 16000 <= 30
+        ), "do not support extract speech token for audio longer than 30s"
+        feat = whisper.log_mel_spectrogram(speech, n_mels=128)
+        speech_token = (
+            self.speech_tokenizer_session.run(
+                None,
+                {
+                    self.speech_tokenizer_session.get_inputs()[0]
+                    .name: feat.detach()
+                    .cpu()
+                    .numpy(),
+                    self.speech_tokenizer_session.get_inputs()[1].name: np.array(
+                        [feat.shape[2]], dtype=np.int32
+                    ),
+                },
+            )[0]
+            .flatten()
+            .tolist()
+        )
+        speech_token = torch.tensor([speech_token], dtype=torch.int32).to(self.device)
+        speech_token_len = torch.tensor([speech_token.shape[1]], dtype=torch.int32).to(
+            self.device
+        )
+        return speech_token, speech_token_len
+
+    def _extract_spk_embedding(self, speech):
+        feat = kaldi.fbank(speech, num_mel_bins=80, dither=0, sample_frequency=16000)
+        feat = feat - feat.mean(dim=0, keepdim=True)
+        embedding = (
+            self.campplus_session.run(
+                None,
+                {
+                    self.campplus_session.get_inputs()[0]
+                    .name: feat.unsqueeze(dim=0)
+                    .cpu()
+                    .numpy()
+                },
+            )[0]
+            .flatten()
+            .tolist()
+        )
+        embedding = torch.tensor([embedding]).to(self.device)
+        return embedding
+
+    def _extract_speech_feat(self, speech):
+        speech_feat = (
+            self.feat_extractor(speech).squeeze(dim=0).transpose(0, 1).to(self.device)
+        )
+        speech_feat = speech_feat.unsqueeze(dim=0)
+        speech_feat_len = torch.tensor([speech_feat.shape[1]], dtype=torch.int32).to(
+            self.device
+        )
+        return speech_feat, speech_feat_len

cosyvoice/cli/model.py

@@ -0,0 +1,32 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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
+
+
+class CosyVoiceModel:
+
+    def __init__(
+        self,
+        flow: torch.nn.Module,
+        hift: torch.nn.Module,
+    ):
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.flow = flow
+        self.hift = hift
+
+    def load(self, flow_model, hift_model):
+        self.flow.load_state_dict(torch.load(flow_model, map_location=self.device))
+        self.flow.to(self.device).eval()
+        self.hift.load_state_dict(torch.load(hift_model, map_location=self.device))
+        self.hift.to(self.device).eval()

cosyvoice/flow/decoder.py

@@ -0,0 +1,238 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# 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
+import torch.nn as nn
+from einops import pack, rearrange, repeat
+from cosyvoice.matcha.decoder import (
+    SinusoidalPosEmb,
+    Block1D,
+    ResnetBlock1D,
+    Downsample1D,
+    TimestepEmbedding,
+    Upsample1D,
+)
+from cosyvoice.matcha.transformer import BasicTransformerBlock
+
+
+class ConditionalDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        channels=(256, 256),
+        dropout=0.05,
+        attention_head_dim=64,
+        n_blocks=1,
+        num_mid_blocks=2,
+        num_heads=4,
+        act_fn="snake",
+    ):
+        """
+        This decoder requires an input with the same shape of the target. So, if your text content
+        is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
+        """
+        super().__init__()
+        channels = tuple(channels)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.time_embeddings = SinusoidalPosEmb(in_channels)
+        time_embed_dim = channels[0] * 4
+        self.time_mlp = TimestepEmbedding(
+            in_channels=in_channels,
+            time_embed_dim=time_embed_dim,
+            act_fn="silu",
+        )
+        self.down_blocks = nn.ModuleList([])
+        self.mid_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = in_channels
+        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
+            input_channel = output_channel
+            output_channel = channels[i]
+            is_last = i == len(channels) - 1
+            resnet = ResnetBlock1D(
+                dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            downsample = (
+                Downsample1D(output_channel)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.down_blocks.append(
+                nn.ModuleList([resnet, transformer_blocks, downsample])
+            )
+
+        for _ in range(num_mid_blocks):
+            input_channel = channels[-1]
+            out_channels = channels[-1]
+            resnet = ResnetBlock1D(
+                dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim
+            )
+
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+
+            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))
+
+        channels = channels[::-1] + (channels[0],)
+        for i in range(len(channels) - 1):
+            input_channel = channels[i] * 2
+            output_channel = channels[i + 1]
+            is_last = i == len(channels) - 2
+            resnet = ResnetBlock1D(
+                dim=input_channel,
+                dim_out=output_channel,
+                time_emb_dim=time_embed_dim,
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            upsample = (
+                Upsample1D(output_channel, use_conv_transpose=True)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
+        self.final_block = Block1D(channels[-1], channels[-1])
+        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.GroupNorm):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, mask, mu, t, spks=None, cond=None):
+        """Forward pass of the UNet1DConditional model.
+
+        Args:
+            x (torch.Tensor): shape (batch_size, in_channels, time)
+            mask (_type_): shape (batch_size, 1, time)
+            t (_type_): shape (batch_size)
+            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
+            cond (_type_, optional): placeholder for future use. Defaults to None.
+
+        Raises:
+            ValueError: _description_
+            ValueError: _description_
+
+        Returns:
+            _type_: _description_
+        """
+
+        t = self.time_embeddings(t).to(t.dtype)
+        t = self.time_mlp(t)
+
+        x = pack([x, mu], "b * t")[0]
+
+        if spks is not None:
+            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
+            x = pack([x, spks], "b * t")[0]
+        if cond is not None:
+            x = pack([x, cond], "b * t")[0]
+
+        hiddens = []
+        masks = [mask]
+        for resnet, transformer_blocks, downsample in self.down_blocks:
+            mask_down = masks[-1]
+            x = resnet(
+                x.to(torch.bfloat16), mask_down.to(torch.bfloat16), t.to(torch.bfloat16)
+            )
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            # attn_mask = torch.matmul(mask_down.transpose(1, 2).contiguous(), mask_down)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    # attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            hiddens.append(x)  # Save hidden states for skip connections
+            x = downsample(x * mask_down)
+            masks.append(mask_down[:, :, ::2])
+        masks = masks[:-1]
+        mask_mid = masks[-1]
+
+        for resnet, transformer_blocks in self.mid_blocks:
+            x = resnet(x, mask_mid, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            # attn_mask = torch.matmul(mask_mid.transpose(1, 2).contiguous(), mask_mid)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    # attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+
+        for resnet, transformer_blocks, upsample in self.up_blocks:
+            mask_up = masks.pop()
+            skip = hiddens.pop()
+            x = pack([x[:, :, : skip.shape[-1]], skip], "b * t")[0]
+            x = resnet(x, mask_up, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            # attn_mask = torch.matmul(mask_up.transpose(1, 2).contiguous(), mask_up)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    # attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            x = upsample(x * mask_up)
+        x = self.final_block(x, mask_up)
+        output = self.final_proj(x * mask_up)
+        return output * mask

cosyvoice/flow/flow.py

@@ -0,0 +1,196 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# 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 logging
+import random
+from typing import Dict, Optional
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+from omegaconf import DictConfig
+from cosyvoice.utils.mask import make_pad_mask
+import time
+
+
+class MaskedDiffWithXvec(torch.nn.Module):
+    def __init__(
+        self,
+        input_size: int = 512,
+        output_size: int = 80,
+        spk_embed_dim: int = 192,
+        output_type: str = "mel",
+        vocab_size: int = 4096,
+        input_frame_rate: int = 50,
+        only_mask_loss: bool = True,
+        encoder: torch.nn.Module = None,
+        length_regulator: torch.nn.Module = None,
+        decoder: torch.nn.Module = None,
+        decoder_conf: Dict = {
+            "in_channels": 240,
+            "out_channel": 80,
+            "spk_emb_dim": 80,
+            "n_spks": 1,
+            "cfm_params": DictConfig(
+                {
+                    "sigma_min": 1e-06,
+                    "solver": "euler",
+                    "t_scheduler": "cosine",
+                    "training_cfg_rate": 0.2,
+                    "inference_cfg_rate": 0.7,
+                    "reg_loss_type": "l1",
+                }
+            ),
+            "decoder_params": {
+                "channels": [256, 256],
+                "dropout": 0.0,
+                "attention_head_dim": 64,
+                "n_blocks": 4,
+                "num_mid_blocks": 12,
+                "num_heads": 8,
+                "act_fn": "gelu",
+            },
+        },
+        mel_feat_conf: Dict = {
+            "n_fft": 1024,
+            "num_mels": 80,
+            "sampling_rate": 22050,
+            "hop_size": 256,
+            "win_size": 1024,
+            "fmin": 0,
+            "fmax": 8000,
+        },
+    ):
+        super().__init__()
+        self.input_size = input_size
+        self.output_size = output_size
+        self.decoder_conf = decoder_conf
+        self.mel_feat_conf = mel_feat_conf
+        self.vocab_size = vocab_size
+        self.output_type = output_type
+        self.input_frame_rate = input_frame_rate
+        logging.info(f"input frame rate={self.input_frame_rate}")
+        self.input_embedding = nn.Embedding(vocab_size, input_size)
+        self.spk_embed_affine_layer = torch.nn.Linear(spk_embed_dim, output_size)
+        self.encoder = encoder
+        self.encoder_proj = torch.nn.Linear(self.encoder.output_size(), output_size)
+        self.decoder = decoder
+        self.length_regulator = length_regulator
+        self.only_mask_loss = only_mask_loss
+
+    def forward(
+        self,
+        batch: dict,
+        device: torch.device,
+    ) -> Dict[str, Optional[torch.Tensor]]:
+        token = batch["speech_token"].to(device)
+        token_len = batch["speech_token_len"].to(device)
+        feat = batch["speech_feat"].to(device)
+        feat_len = batch["speech_feat_len"].to(device)
+        embedding = batch["embedding"].to(device)
+
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        mask = (~make_pad_mask(token_len)).float().unsqueeze(-1).to(device)
+        token = self.input_embedding(torch.clamp(token, min=0)) * mask
+
+        # text encode
+        h, h_lengths = self.encoder(token, token_len)
+        h = self.encoder_proj(h)
+        h, h_lengths = self.length_regulator(h, feat_len)
+
+        # get conditions
+        conds = torch.zeros(feat.shape, device=token.device)
+        for i, j in enumerate(feat_len):
+            if random.random() < 0.5:
+                continue
+            index = random.randint(0, int(0.3 * j))
+            conds[i, :index] = feat[i, :index]
+        conds = conds.transpose(1, 2)
+
+        mask = (~make_pad_mask(feat_len)).to(h)
+        feat = F.interpolate(
+            feat.unsqueeze(dim=1), size=h.shape[1:], mode="nearest"
+        ).squeeze(dim=1)
+        loss, _ = self.decoder.compute_loss(
+            feat.transpose(1, 2).contiguous(),
+            mask.unsqueeze(1),
+            h.transpose(1, 2).contiguous(),
+            embedding,
+            cond=conds,
+        )
+        return {"loss": loss}
+
+    @torch.inference_mode()
+    def inference(
+        self,
+        token,
+        token_len,
+        prompt_token,
+        prompt_token_len,
+        prompt_feat,
+        prompt_feat_len,
+        embedding,
+    ):
+        assert token.shape[0] == 1
+        # xvec projection
+        embedding = F.normalize(embedding, dim=1)
+        embedding = self.spk_embed_affine_layer(embedding)
+
+        # concat text and prompt_text
+        token_len1, token_len2 = prompt_token.shape[1], token.shape[1]
+        # text encode
+        token, token_len = (
+            torch.concat([prompt_token, token], dim=1),
+            prompt_token_len + token_len,
+        )
+        token = self.input_embedding(torch.clamp(token, min=0))
+        h, _ = self.encoder.inference(token, token_len)
+        h = self.encoder_proj(h)
+        mel_len1, mel_len2 = prompt_feat.shape[1], int(
+            token_len2
+            / self.input_frame_rate
+            * self.mel_feat_conf["sampling_rate"]
+            / self.mel_feat_conf["hop_size"]
+        )
+
+        h, _ = self.length_regulator.inference(
+            h[:, :token_len1],
+            h[:, token_len1:],
+            mel_len1,
+            mel_len2,
+        )
+
+        # get conditions
+        conds = torch.zeros(
+            [1, mel_len1 + mel_len2, self.output_size], device=token.device
+        )
+        conds[:, :mel_len1] = prompt_feat
+        conds = conds.transpose(1, 2)
+
+        # mask = (~make_pad_mask(torch.tensor([mel_len1 + mel_len2]))).to(h)
+        mask = torch.ones(
+            [1, mel_len1 + mel_len2], device=h.device, dtype=torch.bfloat16
+        )
+        feat = self.decoder(
+            mu=h.transpose(1, 2).contiguous(),
+            mask=mask.unsqueeze(1),
+            spks=embedding,
+            cond=conds,
+            n_timesteps=10,
+        )
+        feat = feat[:, :, mel_len1:]
+        assert feat.shape[2] == mel_len2
+        return feat

cosyvoice/flow/flow_matching.py

@@ -0,0 +1,315 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# 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 time
+import torch
+import torch.nn.functional as F
+from cosyvoice.matcha.flow_matching import BASECFM
+
+
+class ConditionalCFM(BASECFM):
+    def __init__(
+        self,
+        in_channels,
+        cfm_params,
+        n_spks=1,
+        spk_emb_dim=64,
+        estimator: torch.nn.Module = None,
+    ):
+        super().__init__(
+            n_feats=in_channels,
+            cfm_params=cfm_params,
+            n_spks=n_spks,
+            spk_emb_dim=spk_emb_dim,
+        )
+        self.t_scheduler = cfm_params.t_scheduler
+        self.training_cfg_rate = cfm_params.training_cfg_rate
+        self.inference_cfg_rate = cfm_params.inference_cfg_rate
+        in_channels = in_channels + (spk_emb_dim if n_spks > 0 else 0)
+        # Just change the architecture of the estimator here
+        self.estimator = estimator
+        self.inference_graphs = {}
+        self.inference_buffers = {}
+        # self.capture_inference()
+
+    @torch.inference_mode()
+    def forward(
+        self,
+        mu,
+        mask,
+        n_timesteps,
+        temperature=1.0,
+        spks=None,
+        cond=None,
+    ):
+        """Forward diffusion
+
+        Args:
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            n_timesteps (int): number of diffusion steps
+            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+
+        Returns:
+            sample: generated mel-spectrogram
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        z = torch.randn_like(mu) * temperature
+        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device, dtype=mu.dtype)
+        if self.t_scheduler == "cosine":
+            t_span = 1 - torch.cos(t_span * 0.5 * torch.pi)
+        return self.solve_euler(
+            z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond
+        )
+
+    @torch.inference_mode()
+    def capture_inference(self, seq_len_to_capture=list(range(128, 512, 8))):
+        start_time = time.time()
+        print(
+            f"capture_inference for ConditionalCFM solve euler, seq_len_to_capture: {seq_len_to_capture}"
+        )
+        for seq_len in seq_len_to_capture:
+            static_z = torch.randn(
+                1, 80, seq_len, device=torch.device("cuda"), dtype=torch.bfloat16
+            )
+            static_t_span = torch.linspace(
+                0, 1, 11, device=torch.device("cuda"), dtype=torch.bfloat16
+            )  # only capture at 10 steps
+            static_mu = torch.randn(
+                1, 80, seq_len, device=torch.device("cuda"), dtype=torch.bfloat16
+            )
+            static_mask = torch.ones(
+                1, 1, seq_len, device=torch.device("cuda"), dtype=torch.bfloat16
+            )
+            static_spks = torch.randn(
+                1, 80, device=torch.device("cuda"), dtype=torch.bfloat16
+            )
+            static_cond = torch.randn(
+                1, 80, seq_len, device=torch.device("cuda"), dtype=torch.float32
+            )
+            static_out = torch.randn(
+                1, 80, seq_len, device=torch.device("cuda"), dtype=torch.bfloat16
+            )
+
+            self._solve_euler_impl(
+                static_z,
+                t_span=static_t_span,
+                mu=static_mu,
+                mask=static_mask,
+                spks=static_spks,
+                cond=static_cond,
+            )
+            torch.cuda.synchronize()
+
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                static_out = self._solve_euler_impl(
+                    static_z,
+                    t_span=static_t_span,
+                    mu=static_mu,
+                    mask=static_mask,
+                    spks=static_spks,
+                    cond=static_cond,
+                )
+
+        self.inference_buffers[seq_len] = {
+            "z": static_z,
+            "t_span": static_t_span,
+            "mu": static_mu,
+            "mask": static_mask,
+            "spks": static_spks,
+            "cond": static_cond,
+            "out": static_out,
+        }
+        self.inference_graphs[seq_len] = g
+        end_time = time.time()
+        print(
+            f"capture_inference for ConditionalCFM solve euler, time elapsed: {end_time - start_time}"
+        )
+
+    def solve_euler(self, x, t_span, mu, mask, spks, cond):
+        if hasattr(self, "inference_graphs") and len(self.inference_graphs) > 0:
+            curr_seq_len = x.shape[2]
+
+            available_lengths = sorted(list(self.inference_graphs.keys()))
+
+            if curr_seq_len <= max(available_lengths):
+                target_len = min(available_lengths, key=lambda x: abs(x - curr_seq_len))
+                if target_len == curr_seq_len:
+                    padded_x = x
+                    padded_mu = mu
+                    padded_mask = mask
+                    if cond is not None:
+                        padded_cond = cond
+                else:
+                    padded_x = torch.randn(
+                        (x.shape[0], x.shape[1], target_len),
+                        dtype=x.dtype,
+                        device=x.device,
+                    )
+                    padded_x[:, :, :curr_seq_len] = x
+
+                    padded_mu = torch.randn(
+                        (mu.shape[0], mu.shape[1], target_len),
+                        dtype=mu.dtype,
+                        device=mu.device,
+                    )
+                    padded_mu[:, :, :curr_seq_len] = mu
+
+                    # FIXME(ys): uses zeros and maskgroupnorm
+                    padded_mask = torch.ones(
+                        (mask.shape[0], mask.shape[1], target_len),
+                        dtype=mask.dtype,
+                        device=mask.device,
+                    )
+
+                    if cond is not None:
+                        padded_cond = torch.randn(
+                            (cond.shape[0], cond.shape[1], target_len),
+                            dtype=cond.dtype,
+                            device=cond.device,
+                        )
+                        padded_cond[:, :, :curr_seq_len] = cond
+
+                buffer = self.inference_buffers[target_len]
+                buffer["z"].copy_(padded_x)
+                buffer["t_span"].copy_(t_span)
+                buffer["mu"].copy_(padded_mu)
+                buffer["mask"].copy_(padded_mask)
+                buffer["spks"].copy_(spks)
+                if cond is not None:
+                    buffer["cond"].copy_(padded_cond)
+
+                self.inference_graphs[target_len].replay()
+
+                output = buffer["out"][:, :, :curr_seq_len]
+                return output
+
+        return self._solve_euler_impl(x, t_span, mu, mask, spks, cond)
+
+    def _solve_euler_impl(self, x, t_span, mu, mask, spks, cond):
+        """
+        Fixed euler solver for ODEs.
+        Args:
+            x (torch.Tensor): random noise
+            t_span (torch.Tensor): n_timesteps interpolated
+                shape: (n_timesteps + 1,)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+        """
+        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
+        t = t.unsqueeze(dim=0)
+
+        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
+        # Or in future might add like a return_all_steps flag
+        sol = []
+
+        for step in range(1, len(t_span)):
+            if self.inference_cfg_rate > 0:
+                x_double = torch.cat([x, x], dim=0)
+                mask_double = torch.cat([mask, mask], dim=0)
+                mu_double = torch.cat([mu, torch.zeros_like(mu)], dim=0)
+                t_double = torch.cat([t, t], dim=0)
+                spks_double = (
+                    torch.cat([spks, torch.zeros_like(spks)], dim=0)
+                    if spks is not None
+                    else None
+                )
+                cond_double = torch.cat([cond, torch.zeros_like(cond)], dim=0)
+
+                dphi_dt_double = self.forward_estimator(
+                    x_double, mask_double, mu_double, t_double, spks_double, cond_double
+                )
+
+                dphi_dt, cfg_dphi_dt = torch.chunk(dphi_dt_double, 2, dim=0)
+                dphi_dt = (
+                    1.0 + self.inference_cfg_rate
+                ) * dphi_dt - self.inference_cfg_rate * cfg_dphi_dt
+            else:
+                dphi_dt = self.forward_estimator(x, mask, mu, t, spks, cond)
+
+            x = x + dt * dphi_dt
+            t = t + dt
+            sol.append(x)
+            if step < len(t_span) - 1:
+                dt = t_span[step + 1] - t
+
+        return sol[-1]
+
+    def forward_estimator(self, x, mask, mu, t, spks, cond):
+        if isinstance(self.estimator, torch.nn.Module):
+            return self.estimator.forward(x, mask, mu, t, spks, cond)
+        else:
+            ort_inputs = {
+                "x": x.cpu().numpy(),
+                "mask": mask.cpu().numpy(),
+                "mu": mu.cpu().numpy(),
+                "t": t.cpu().numpy(),
+                "spks": spks.cpu().numpy(),
+                "cond": cond.cpu().numpy(),
+            }
+            output = self.estimator.run(None, ort_inputs)[0]
+            return torch.tensor(output, dtype=x.dtype, device=x.device)
+
+    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
+        """Computes diffusion loss
+
+        Args:
+            x1 (torch.Tensor): Target
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): target mask
+                shape: (batch_size, 1, mel_timesteps)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            spks (torch.Tensor, optional): speaker embedding. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+
+        Returns:
+            loss: conditional flow matching loss
+            y: conditional flow
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        b, _, t = mu.shape
+
+        # random timestep
+        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
+        if self.t_scheduler == "cosine":
+            t = 1 - torch.cos(t * 0.5 * torch.pi)
+        # sample noise p(x_0)
+        z = torch.randn_like(x1)
+
+        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
+        u = x1 - (1 - self.sigma_min) * z
+
+        # during training, we randomly drop condition to trade off mode coverage and sample fidelity
+        if self.training_cfg_rate > 0:
+            cfg_mask = torch.rand(b, device=x1.device) > self.training_cfg_rate
+            mu = mu * cfg_mask.view(-1, 1, 1)
+            spks = spks * cfg_mask.view(-1, 1)
+            cond = cond * cfg_mask.view(-1, 1, 1)
+
+        pred = self.estimator(y, mask, mu, t.squeeze(), spks, cond)
+        loss = F.mse_loss(pred * mask, u * mask, reduction="sum") / (
+            torch.sum(mask) * u.shape[1]
+        )
+        return loss, y

cosyvoice/flow/length_regulator.py

@@ -0,0 +1,65 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# 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.
+from typing import Tuple
+import torch.nn as nn
+import torch
+from torch.nn import functional as F
+from cosyvoice.utils.mask import make_pad_mask
+
+
+class InterpolateRegulator(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        sampling_ratios: Tuple,
+        out_channels: int = None,
+        groups: int = 1,
+    ):
+        super().__init__()
+        self.sampling_ratios = sampling_ratios
+        out_channels = out_channels or channels
+        model = nn.ModuleList([])
+        if len(sampling_ratios) > 0:
+            for _ in sampling_ratios:
+                module = nn.Conv1d(channels, channels, 3, 1, 1)
+                norm = nn.GroupNorm(groups, channels)
+                act = nn.Mish()
+                model.extend([module, norm, act])
+        model.append(nn.Conv1d(channels, out_channels, 1, 1))
+        self.model = nn.Sequential(*model)
+
+    def forward(self, x, ylens=None):
+        # x in (B, T, D)
+        mask = (~make_pad_mask(ylens)).to(x).unsqueeze(-1)
+        x = F.interpolate(
+            x.transpose(1, 2).contiguous(), size=ylens.max(), mode="linear"
+        )
+        out = self.model(x).transpose(1, 2).contiguous()
+        olens = ylens
+        return out * mask, olens
+
+    def inference(self, x1, x2, mel_len1, mel_len2):
+        # x in (B, T, D)
+        x2 = F.interpolate(
+            x2.transpose(1, 2).contiguous(), size=mel_len2, mode="linear"
+        )
+        if x1.shape[1] != 0:
+            x1 = F.interpolate(
+                x1.transpose(1, 2).contiguous(), size=mel_len1, mode="linear"
+            )
+            x = torch.concat([x1, x2], dim=2)
+        else:
+            x = x2
+        out = self.model(x).transpose(1, 2).contiguous()
+        return out, mel_len1 + mel_len2

cosyvoice/hifigan/f0_predictor.py

@@ -0,0 +1,55 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# 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
+import torch.nn as nn
+from torch.nn.utils import weight_norm
+
+
+class ConvRNNF0Predictor(nn.Module):
+    def __init__(
+        self, num_class: int = 1, in_channels: int = 80, cond_channels: int = 512
+    ):
+        super().__init__()
+
+        self.num_class = num_class
+        self.condnet = nn.Sequential(
+            weight_norm(
+                nn.Conv1d(in_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+            weight_norm(
+                nn.Conv1d(cond_channels, cond_channels, kernel_size=3, padding=1)
+            ),
+            nn.ELU(),
+        )
+        self.classifier = nn.Linear(
+            in_features=cond_channels, out_features=self.num_class
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.condnet(x)
+        x = x.transpose(1, 2)
+        return torch.abs(self.classifier(x).squeeze(-1))

cosyvoice/hifigan/generator.py

@@ -0,0 +1,566 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Kai Hu)
+#
+# 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.
+
+"""HIFI-GAN"""
+
+import typing as tp
+import time
+import numpy as np
+from scipy.signal import get_window
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import Conv1d
+from torch.nn import ConvTranspose1d
+from torch.nn.utils import remove_weight_norm
+from torch.nn.utils import weight_norm
+from torch.distributions.uniform import Uniform
+
+from cosyvoice.transformer.activation import Snake
+from cosyvoice.utils.common import get_padding
+from cosyvoice.utils.common import init_weights
+
+
+"""hifigan based generator implementation.
+
+This code is modified from https://github.com/jik876/hifi-gan
+ ,https://github.com/kan-bayashi/ParallelWaveGAN and
+ 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],
+    ):
+        super(ResBlock, self).__init__()
+        self.convs1 = nn.ModuleList()
+        self.convs2 = nn.ModuleList()
+
+        for dilation in dilations:
+            self.convs1.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation,
+                        padding=get_padding(kernel_size, dilation),
+                    )
+                )
+            )
+            self.convs2.append(
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                )
+            )
+        self.convs1.apply(init_weights)
+        self.convs2.apply(init_weights)
+        self.activations1 = nn.ModuleList(
+            [Snake(channels, alpha_logscale=False) for _ in range(len(self.convs1))]
+        )
+        self.activations2 = nn.ModuleList(
+            [Snake(channels, alpha_logscale=False) for _ in range(len(self.convs2))]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        for idx in range(len(self.convs1)):
+            xt = self.activations1[idx](x)
+            xt = self.convs1[idx](xt)
+            xt = self.activations2[idx](xt)
+            xt = self.convs2[idx](xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for idx in range(len(self.convs1)):
+            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,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(
+        self,
+        samp_rate,
+        harmonic_num=0,
+        sine_amp=0.1,
+        noise_std=0.003,
+        voiced_threshold=0,
+    ):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = (f0 > self.voiced_threshold).type(torch.float32)
+        return uv
+
+    @torch.no_grad()
+    def forward(self, f0):
+        """
+        :param f0: [B, 1, sample_len], Hz
+        :return: [B, 1, sample_len]
+        """
+
+        F_mat = torch.zeros((f0.size(0), self.harmonic_num + 1, f0.size(-1))).to(
+            f0.device
+        )
+        for i in range(self.harmonic_num + 1):
+            F_mat[:, i : i + 1, :] = f0 * (i + 1) / self.sampling_rate
+
+        theta_mat = 2 * np.pi * (torch.cumsum(F_mat, dim=-1) % 1)
+        u_dist = Uniform(low=-np.pi, high=np.pi)
+        phase_vec = u_dist.sample(
+            sample_shape=(f0.size(0), self.harmonic_num + 1, 1)
+        ).to(F_mat.device)
+        phase_vec[:, 0, :] = 0
+
+        # generate sine waveforms
+        sine_waves = self.sine_amp * torch.sin(theta_mat + phase_vec)
+
+        # generate uv signal
+        uv = self._f02uv(f0)
+
+        # noise: for unvoiced should be similar to sine_amp
+        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
+        # .       for voiced regions is self.noise_std
+        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+        noise = noise_amp * torch.randn_like(sine_waves)
+
+        # first: set the unvoiced part to 0 by uv
+        # then: additive noise
+        sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(
+        self,
+        sampling_rate,
+        upsample_scale,
+        harmonic_num=0,
+        sine_amp=0.1,
+        add_noise_std=0.003,
+        voiced_threshod=0,
+    ):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(
+            sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+        )
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x):
+        """
+        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+        F0_sampled (batchsize, length, 1)
+        Sine_source (batchsize, length, 1)
+        noise_source (batchsize, length 1)
+        """
+        # source for harmonic branch
+        with torch.no_grad():
+            sine_wavs, uv, _ = self.l_sin_gen(x.transpose(1, 2))
+            sine_wavs = sine_wavs.transpose(1, 2)
+            uv = uv.transpose(1, 2)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+
+        # source for noise branch, in the same shape as uv
+        noise = torch.randn_like(uv) * self.sine_amp / 3
+        return sine_merge, noise, uv
+
+
+class HiFTGenerator(nn.Module):
+    """
+    HiFTNet Generator: Neural Source Filter + ISTFTNet
+    https://arxiv.org/abs/2309.09493
+    """
+
+    def __init__(
+        self,
+        in_channels: int = 80,
+        base_channels: int = 512,
+        nb_harmonics: int = 8,
+        sampling_rate: int = 22050,
+        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]],
+        lrelu_slope: float = 0.1,
+        audio_limit: float = 0.99,
+        f0_predictor: torch.nn.Module = None,
+    ):
+        super(HiFTGenerator, self).__init__()
+
+        self.out_channels = 1
+        self.nb_harmonics = nb_harmonics
+        self.sampling_rate = sampling_rate
+        self.istft_params = istft_params
+        self.lrelu_slope = lrelu_slope
+        self.audio_limit = audio_limit
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.upsample_rates = upsample_rates
+        self.m_source = SourceModuleHnNSF(
+            sampling_rate=sampling_rate,
+            upsample_scale=np.prod(upsample_rates) * istft_params["hop_len"],
+            harmonic_num=nb_harmonics,
+            sine_amp=nsf_alpha,
+            add_noise_std=nsf_sigma,
+            voiced_threshod=nsf_voiced_threshold,
+        )
+        self.f0_upsamp = torch.nn.Upsample(
+            scale_factor=np.prod(upsample_rates) * istft_params["hop_len"]
+        )
+
+        self.conv_pre = weight_norm(Conv1d(in_channels, base_channels, 7, 1, padding=3))
+
+        # Up
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        base_channels // (2**i),
+                        base_channels // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        # Down
+        self.source_downs = nn.ModuleList()
+        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,
+            )
+        ):
+            if u == 1:
+                self.source_downs.append(
+                    Conv1d(
+                        istft_params["n_fft"] + 2, base_channels // (2 ** (i + 1)), 1, 1
+                    )
+                )
+            else:
+                self.source_downs.append(
+                    Conv1d(
+                        istft_params["n_fft"] + 2,
+                        base_channels // (2 ** (i + 1)),
+                        u * 2,
+                        u,
+                        padding=(u // 2),
+                    )
+                )
+
+            self.source_resblocks.append(
+                ResBlock(base_channels // (2 ** (i + 1)), k, d)
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = base_channels // (2 ** (i + 1))
+            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)
+        )
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = nn.ReflectionPad1d((1, 0))
+        self.stft_window = torch.from_numpy(
+            get_window("hann", istft_params["n_fft"], fftbins=True).astype(np.float32)
+        ).cuda()
+        self.f0_predictor = f0_predictor
+        self.inference_buffers = {}
+        self.inference_graphs = {}
+
+    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 _stft(self, x):
+        spec = torch.stft(
+            x,
+            self.istft_params["n_fft"],
+            self.istft_params["hop_len"],
+            self.istft_params["n_fft"],
+            window=self.stft_window,
+            return_complex=True,
+        )
+        spec = torch.view_as_real(spec)  # [B, F, TT, 2]
+        return spec[..., 0], spec[..., 1]
+
+    def _istft(self, magnitude, phase):
+        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,
+        )
+        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
+
+        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, self.lrelu_slope)
+            x = self.ups[i](x)
+
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            # fusion
+            si = self.source_downs[i](s_stft)
+            si = self.source_resblocks[i](si)
+            x = x + si
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        magnitude = torch.exp(x[:, : self.istft_params["n_fft"] // 2 + 1, :])
+        phase = torch.sin(
+            x[:, self.istft_params["n_fft"] // 2 + 1 :, :]
+        )  # actually, sin is redundancy
+
+        x = self._istft(magnitude, phase)
+        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
+        return x, s
+
+    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()
+
+    @torch.inference_mode()
+    def _inference_impl(self, mel: torch.Tensor, s_stft: torch.Tensor) -> torch.Tensor:
+        x = self.conv_pre(mel)
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, self.lrelu_slope)
+            x = self.ups[i](x)
+
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            # fusion
+            si = self.source_downs[i](s_stft)
+            si = self.source_resblocks[i](si)
+            x = x + si
+
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        magnitude = torch.exp(x[:, : self.istft_params["n_fft"] // 2 + 1, :])
+        phase = torch.sin(
+            x[:, self.istft_params["n_fft"] // 2 + 1 :, :]
+        )  # actually, sin is redundancy
+        # print(f"mel: {mel.shape}, magnitude: {magnitude.shape}, phase: {phase.shape}")
+        return magnitude, phase
+
+    @torch.inference_mode()
+    def inference(
+        self, mel: torch.Tensor, cache_source: torch.Tensor = torch.zeros(1, 1, 0)
+    ) -> torch.Tensor:
+        curr_seq_len = mel.shape[2]
+        f0 = self.f0_predictor(mel)
+        s = self._f02source(f0)
+        s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+        s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+        target_len = None
+        for seq_len in sorted(self.inference_buffers.keys()):
+            if curr_seq_len <= seq_len:
+                target_len = seq_len
+                break
+
+        if target_len is not None:
+            buffer = self.inference_buffers[target_len]
+
+            if curr_seq_len < target_len:
+                padded_mel = torch.zeros_like(buffer["mel"])
+                padded_mel[:, :, :curr_seq_len] = mel
+                buffer["mel"].copy_(padded_mel)
+                padded_s_stft = torch.zeros_like(buffer["s_stft"])
+                cur_s_stft_len = s_stft.shape[2]
+                padded_s_stft[:, :, :cur_s_stft_len] = s_stft
+                buffer["s_stft"].copy_(padded_s_stft)
+
+            else:
+                buffer["mel"].copy_(mel)
+                buffer["s_stft"].copy_(s_stft)
+                cur_s_stft_len = s_stft.shape[2]
+
+            self.inference_graphs[target_len].replay()
+
+            magnitude, phase = (
+                buffer["magnitude"][:, :, :cur_s_stft_len],
+                buffer["phase"][:, :, :cur_s_stft_len],
+            )
+        else:
+            magnitude, phase = self._inference_impl(mel=mel, s_stft=s_stft)
+
+        x = self._istft(magnitude, phase)
+        x = torch.clamp(x, -self.audio_limit, self.audio_limit)
+        return x, s
+
+    @torch.inference_mode()
+    def capture_inference(self, seq_len_to_capture=[64, 128, 256, 512, 1024]):
+        start_time = time.time()
+        print(
+            f"capture inference for HiFTGenerator with seq_len_to_capture: {seq_len_to_capture}"
+        )
+        for seq_len in seq_len_to_capture:
+            mel = torch.randn(
+                1, 80, seq_len, device=torch.device("cuda"), dtype=torch.float32
+            )
+            f0 = self.f0_predictor(mel)
+            s = self._f02source(f0)
+            s_stft_real, s_stft_imag = self._stft(s.squeeze(1))
+            s_stft = torch.cat([s_stft_real, s_stft_imag], dim=1)
+
+            magnitude, phase = self._inference_impl(mel=mel, s_stft=s_stft)
+            torch.cuda.synchronize()
+
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                magnitude, phase = self._inference_impl(mel=mel, s_stft=s_stft)
+            inference_buffer = {
+                "mel": mel,
+                "s_stft": s_stft,
+                "magnitude": magnitude,
+                "phase": phase,
+            }
+            self.inference_buffers[seq_len] = inference_buffer
+            self.inference_graphs[seq_len] = g
+
+        end_time = time.time()
+        print(
+            f"capture inference for HiFTGenerator with seq_len_to_capture: {seq_len_to_capture} takes {end_time - start_time} seconds"
+        )

cosyvoice/matcha/audio.py

@@ -0,0 +1,90 @@
+import numpy as np
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+from scipy.io.wavfile import read
+
+MAX_WAV_VALUE = 32768.0
+
+
+def load_wav(full_path):
+    sampling_rate, data = read(full_path)
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def mel_spectrogram(
+    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
+):
+    if torch.min(y) < -1.0:
+        print("min value is ", torch.min(y))
+    if torch.max(y) > 1.0:
+        print("max value is ", torch.max(y))
+
+    global mel_basis, hann_window  # pylint: disable=global-statement
+    if f"{str(fmax)}_{str(y.device)}" not in mel_basis:
+        mel = librosa_mel_fn(
+            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+        )
+        mel_basis[str(fmax) + "_" + str(y.device)] = (
+            torch.from_numpy(mel).float().to(y.device)
+        )
+        hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1),
+        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
+        mode="reflect",
+    )
+    y = y.squeeze(1)
+
+    spec = torch.view_as_real(
+        torch.stft(
+            y,
+            n_fft,
+            hop_length=hop_size,
+            win_length=win_size,
+            window=hann_window[str(y.device)],
+            center=center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+    )
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+    spec = torch.matmul(mel_basis[str(fmax) + "_" + str(y.device)], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec

cosyvoice/matcha/decoder.py

@@ -0,0 +1,511 @@
+import math
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from conformer import ConformerBlock
+from diffusers.models.activations import get_activation
+from einops import pack, rearrange, repeat
+
+from cosyvoice.matcha.transformer import BasicTransformerBlock
+
+
+class SinusoidalPosEmb(torch.nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+        assert self.dim % 2 == 0, "SinusoidalPosEmb requires dim to be even"
+
+    def forward(self, x, scale=1000):
+        if x.ndim < 1:
+            x = x.unsqueeze(0)
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device).float() * -emb)
+        emb = scale * x.unsqueeze(1) * emb.unsqueeze(0)
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+
+class MaskedGroupNorm(nn.GroupNorm):
+    """
+    Masked verstion of the Group normalization.
+
+    Based on: https://github.com/ptrblck/pytorch_misc/blob/20e8ea93bd458b88f921a87e2d4001a4eb753a02/batch_norm_manual.py
+
+    Receives a N-dim tensor of sequence lengths per batch element
+    along with the regular input for masking.
+
+    Check pytorch's GroupNorm implementation for argument details.
+    """
+
+    def __init__(self, num_groups, num_channels, eps=1e-5, affine=True):
+        super(MaskedGroupNorm, self).__init__(num_groups, num_channels, eps, affine)
+
+    def forward(self, inp, mask=None):
+        assert (
+            inp.shape[1] % self.num_groups == 0
+        ), "Feature size not divisible by groups"
+
+        # 计算有效长度
+        seq_lengths = mask.sum(-1, keepdim=True)  # [batch_size, 1]
+
+        # 将输入reshape为groups
+        features_per_group = inp.shape[1] // self.num_groups
+        inp_r = inp.reshape(
+            inp.shape[0], self.num_groups, features_per_group, inp.shape[-1]
+        )
+        mask_r = mask.unsqueeze(1)  # [batch_size, 1, 1, length]
+
+        # 计算masked mean和variance
+        masked_inp = inp_r * mask_r
+        n = seq_lengths * features_per_group  # 每组的有效元素数量
+        mean = masked_inp.sum([2, 3], keepdim=True) / (n.view(-1, 1, 1, 1) + 1e-5)
+        var = ((masked_inp - mean * mask_r) ** 2).sum([2, 3], keepdim=True) / (
+            n.view(-1, 1, 1, 1) + 1e-5
+        )
+
+        # 标准化
+        inp_r = (inp_r - mean) / (torch.sqrt(var + self.eps))
+        out = inp_r.reshape(inp.shape[0], self.num_channels, inp.shape[-1])
+
+        # 应用仿射变换
+        if self.affine:
+            out = out * self.weight[None, :, None] + self.bias[None, :, None]
+
+        return out
+
+
+class Block1D(torch.nn.Module):
+    def __init__(self, dim, dim_out, groups=8):
+        super().__init__()
+        self.block = torch.nn.Sequential(
+            torch.nn.Conv1d(dim, dim_out, 3, padding=1),
+            torch.nn.GroupNorm(groups, dim_out),
+            # MaskedGroupNorm(groups, dim_out),
+            nn.Mish(),
+        )
+
+    def forward(self, x, mask):
+        output = self.block(x * mask)
+        return output * mask
+        return x * mask
+
+
+class ResnetBlock1D(torch.nn.Module):
+    def __init__(self, dim, dim_out, time_emb_dim, groups=8):
+        super().__init__()
+        self.mlp = torch.nn.Sequential(
+            nn.Mish(), torch.nn.Linear(time_emb_dim, dim_out)
+        )
+
+        self.block1 = Block1D(dim, dim_out, groups=groups)
+        self.block2 = Block1D(dim_out, dim_out, groups=groups)
+
+        self.res_conv = torch.nn.Conv1d(dim, dim_out, 1)
+
+    def forward(self, x, mask, time_emb):
+        h = self.block1(x, mask)
+        h += self.mlp(time_emb).unsqueeze(-1)
+        h = self.block2(h, mask)
+        output = h + self.res_conv(x * mask)
+        return output
+
+
+class Downsample1D(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.conv = torch.nn.Conv1d(dim, dim, 3, 2, 1)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        time_embed_dim: int,
+        act_fn: str = "silu",
+        out_dim: int = None,
+        post_act_fn: Optional[str] = None,
+        cond_proj_dim=None,
+    ):
+        super().__init__()
+
+        self.linear_1 = nn.Linear(in_channels, time_embed_dim)
+
+        if cond_proj_dim is not None:
+            self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
+        else:
+            self.cond_proj = None
+
+        self.act = get_activation(act_fn)
+
+        if out_dim is not None:
+            time_embed_dim_out = out_dim
+        else:
+            time_embed_dim_out = time_embed_dim
+        self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out)
+
+        if post_act_fn is None:
+            self.post_act = None
+        else:
+            self.post_act = get_activation(post_act_fn)
+
+    def forward(self, sample, condition=None):
+        if condition is not None:
+            sample = sample + self.cond_proj(condition)
+        sample = self.linear_1(sample)
+
+        if self.act is not None:
+            sample = self.act(sample)
+
+        sample = self.linear_2(sample)
+
+        if self.post_act is not None:
+            sample = self.post_act(sample)
+        return sample
+
+
+class Upsample1D(nn.Module):
+    """A 1D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+    """
+
+    def __init__(
+        self,
+        channels,
+        use_conv=False,
+        use_conv_transpose=True,
+        out_channels=None,
+        name="conv",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        self.conv = None
+        if use_conv_transpose:
+            self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
+        elif use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, inputs):
+        assert inputs.shape[1] == self.channels
+        if self.use_conv_transpose:
+            return self.conv(inputs)
+
+        outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
+
+        if self.use_conv:
+            outputs = self.conv(outputs)
+
+        return outputs
+
+
+class ConformerWrapper(ConformerBlock):
+    def __init__(  # pylint: disable=useless-super-delegation
+        self,
+        *,
+        dim,
+        dim_head=64,
+        heads=8,
+        ff_mult=4,
+        conv_expansion_factor=2,
+        conv_kernel_size=31,
+        attn_dropout=0,
+        ff_dropout=0,
+        conv_dropout=0,
+        conv_causal=False,
+    ):
+        super().__init__(
+            dim=dim,
+            dim_head=dim_head,
+            heads=heads,
+            ff_mult=ff_mult,
+            conv_expansion_factor=conv_expansion_factor,
+            conv_kernel_size=conv_kernel_size,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            conv_dropout=conv_dropout,
+            conv_causal=conv_causal,
+        )
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        timestep=None,
+    ):
+        return super().forward(x=hidden_states, mask=attention_mask.bool())
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        channels=(256, 256),
+        dropout=0.05,
+        attention_head_dim=64,
+        n_blocks=1,
+        num_mid_blocks=2,
+        num_heads=4,
+        act_fn="snake",
+        down_block_type="transformer",
+        mid_block_type="transformer",
+        up_block_type="transformer",
+    ):
+        super().__init__()
+        channels = tuple(channels)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.time_embeddings = SinusoidalPosEmb(in_channels)
+        time_embed_dim = channels[0] * 4
+        self.time_mlp = TimestepEmbedding(
+            in_channels=in_channels,
+            time_embed_dim=time_embed_dim,
+            act_fn="silu",
+        )
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = in_channels
+        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
+            input_channel = output_channel
+            output_channel = channels[i]
+            is_last = i == len(channels) - 1
+            resnet = ResnetBlock1D(
+                dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    self.get_block(
+                        down_block_type,
+                        output_channel,
+                        attention_head_dim,
+                        num_heads,
+                        dropout,
+                        act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            downsample = (
+                Downsample1D(output_channel)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+
+            self.down_blocks.append(
+                nn.ModuleList([resnet, transformer_blocks, downsample])
+            )
+
+        for i in range(num_mid_blocks):
+            input_channel = channels[-1]
+            out_channels = channels[-1]
+
+            resnet = ResnetBlock1D(
+                dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim
+            )
+
+            transformer_blocks = nn.ModuleList(
+                [
+                    self.get_block(
+                        mid_block_type,
+                        output_channel,
+                        attention_head_dim,
+                        num_heads,
+                        dropout,
+                        act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+
+            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))
+
+        channels = channels[::-1] + (channels[0],)
+        for i in range(len(channels) - 1):
+            input_channel = channels[i]
+            output_channel = channels[i + 1]
+            is_last = i == len(channels) - 2
+
+            resnet = ResnetBlock1D(
+                dim=2 * input_channel,
+                dim_out=output_channel,
+                time_emb_dim=time_embed_dim,
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    self.get_block(
+                        up_block_type,
+                        output_channel,
+                        attention_head_dim,
+                        num_heads,
+                        dropout,
+                        act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            upsample = (
+                Upsample1D(output_channel, use_conv_transpose=True)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+
+            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
+
+        self.final_block = Block1D(channels[-1], channels[-1])
+        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
+
+        self.initialize_weights()
+        # nn.init.normal_(self.final_proj.weight)
+
+    @staticmethod
+    def get_block(block_type, dim, attention_head_dim, num_heads, dropout, act_fn):
+        if block_type == "conformer":
+            block = ConformerWrapper(
+                dim=dim,
+                dim_head=attention_head_dim,
+                heads=num_heads,
+                ff_mult=1,
+                conv_expansion_factor=2,
+                ff_dropout=dropout,
+                attn_dropout=dropout,
+                conv_dropout=dropout,
+                conv_kernel_size=31,
+            )
+        elif block_type == "transformer":
+            block = BasicTransformerBlock(
+                dim=dim,
+                num_attention_heads=num_heads,
+                attention_head_dim=attention_head_dim,
+                dropout=dropout,
+                activation_fn=act_fn,
+            )
+        else:
+            raise ValueError(f"Unknown block type {block_type}")
+
+        return block
+
+    def initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+            elif isinstance(m, nn.GroupNorm):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, mask, mu, t, spks=None, cond=None):
+        """Forward pass of the UNet1DConditional model.
+
+        Args:
+            x (torch.Tensor): shape (batch_size, in_channels, time)
+            mask (_type_): shape (batch_size, 1, time)
+            t (_type_): shape (batch_size)
+            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
+            cond (_type_, optional): placeholder for future use. Defaults to None.
+
+        Raises:
+            ValueError: _description_
+            ValueError: _description_
+
+        Returns:
+            _type_: _description_
+        """
+
+        t = self.time_embeddings(t)
+        t = self.time_mlp(t)
+
+        x = pack([x, mu], "b * t")[0]
+
+        if spks is not None:
+            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
+            x = pack([x, spks], "b * t")[0]
+
+        hiddens = []
+        masks = [mask]
+        for resnet, transformer_blocks, downsample in self.down_blocks:
+            mask_down = masks[-1]
+            x = resnet(x, mask_down, t)
+            x = rearrange(x, "b c t -> b t c")
+            mask_down = rearrange(mask_down, "b 1 t -> b t")
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=mask_down,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t")
+            mask_down = rearrange(mask_down, "b t -> b 1 t")
+            hiddens.append(x)  # Save hidden states for skip connections
+            x = downsample(x * mask_down)
+            masks.append(mask_down[:, :, ::2])
+
+        masks = masks[:-1]
+        mask_mid = masks[-1]
+
+        for resnet, transformer_blocks in self.mid_blocks:
+            x = resnet(x, mask_mid, t)
+            x = rearrange(x, "b c t -> b t c")
+            mask_mid = rearrange(mask_mid, "b 1 t -> b t")
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=mask_mid,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t")
+            mask_mid = rearrange(mask_mid, "b t -> b 1 t")
+
+        for resnet, transformer_blocks, upsample in self.up_blocks:
+            mask_up = masks.pop()
+            x = resnet(pack([x, hiddens.pop()], "b * t")[0], mask_up, t)
+            x = rearrange(x, "b c t -> b t c")
+            mask_up = rearrange(mask_up, "b 1 t -> b t")
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=mask_up,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t")
+            mask_up = rearrange(mask_up, "b t -> b 1 t")
+            x = upsample(x * mask_up)
+
+        x = self.final_block(x, mask_up)
+        output = self.final_proj(x * mask_up)
+
+        return output * mask

cosyvoice/matcha/flow_matching.py

@@ -0,0 +1,141 @@
+from abc import ABC
+
+import torch
+import torch.nn.functional as F
+
+from cosyvoice.matcha.decoder import Decoder
+
+
+class BASECFM(torch.nn.Module, ABC):
+    def __init__(
+        self,
+        n_feats,
+        cfm_params,
+        n_spks=1,
+        spk_emb_dim=128,
+    ):
+        super().__init__()
+        self.n_feats = n_feats
+        self.n_spks = n_spks
+        self.spk_emb_dim = spk_emb_dim
+        self.solver = cfm_params.solver
+        if hasattr(cfm_params, "sigma_min"):
+            self.sigma_min = cfm_params.sigma_min
+        else:
+            self.sigma_min = 1e-4
+
+        self.estimator = None
+
+    @torch.inference_mode()
+    def forward(self, mu, mask, n_timesteps, temperature=1.0, spks=None, cond=None):
+        """Forward diffusion
+
+        Args:
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            n_timesteps (int): number of diffusion steps
+            temperature (float, optional): temperature for scaling noise. Defaults to 1.0.
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+
+        Returns:
+            sample: generated mel-spectrogram
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        z = torch.randn_like(mu) * temperature
+        t_span = torch.linspace(0, 1, n_timesteps + 1, device=mu.device)
+        return self.solve_euler(
+            z, t_span=t_span, mu=mu, mask=mask, spks=spks, cond=cond
+        )
+
+    def solve_euler(self, x, t_span, mu, mask, spks, cond):
+        """
+        Fixed euler solver for ODEs.
+        Args:
+            x (torch.Tensor): random noise
+            t_span (torch.Tensor): n_timesteps interpolated
+                shape: (n_timesteps + 1,)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): output_mask
+                shape: (batch_size, 1, mel_timesteps)
+            spks (torch.Tensor, optional): speaker ids. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+            cond: Not used but kept for future purposes
+        """
+        t, _, dt = t_span[0], t_span[-1], t_span[1] - t_span[0]
+
+        # I am storing this because I can later plot it by putting a debugger here and saving it to a file
+        # Or in future might add like a return_all_steps flag
+        sol = []
+
+        for step in range(1, len(t_span)):
+            dphi_dt = self.estimator(x, mask, mu, t, spks, cond)
+
+            x = x + dt * dphi_dt
+            t = t + dt
+            sol.append(x)
+            if step < len(t_span) - 1:
+                dt = t_span[step + 1] - t
+
+        return sol[-1]
+
+    def compute_loss(self, x1, mask, mu, spks=None, cond=None):
+        """Computes diffusion loss
+
+        Args:
+            x1 (torch.Tensor): Target
+                shape: (batch_size, n_feats, mel_timesteps)
+            mask (torch.Tensor): target mask
+                shape: (batch_size, 1, mel_timesteps)
+            mu (torch.Tensor): output of encoder
+                shape: (batch_size, n_feats, mel_timesteps)
+            spks (torch.Tensor, optional): speaker embedding. Defaults to None.
+                shape: (batch_size, spk_emb_dim)
+
+        Returns:
+            loss: conditional flow matching loss
+            y: conditional flow
+                shape: (batch_size, n_feats, mel_timesteps)
+        """
+        b, _, t = mu.shape
+
+        # random timestep
+        t = torch.rand([b, 1, 1], device=mu.device, dtype=mu.dtype)
+        # sample noise p(x_0)
+        z = torch.randn_like(x1)
+
+        y = (1 - (1 - self.sigma_min) * t) * z + t * x1
+        u = x1 - (1 - self.sigma_min) * z
+
+        loss = F.mse_loss(
+            self.estimator(y, mask, mu, t.squeeze(), spks), u, reduction="sum"
+        ) / (torch.sum(mask) * u.shape[1])
+        return loss, y
+
+
+class CFM(BASECFM):
+    def __init__(
+        self,
+        in_channels,
+        out_channel,
+        cfm_params,
+        decoder_params,
+        n_spks=1,
+        spk_emb_dim=64,
+    ):
+        super().__init__(
+            n_feats=in_channels,
+            cfm_params=cfm_params,
+            n_spks=n_spks,
+            spk_emb_dim=spk_emb_dim,
+        )
+
+        in_channels = in_channels + (spk_emb_dim if n_spks > 1 else 0)
+        # Just change the architecture of the estimator here
+        self.estimator = Decoder(
+            in_channels=in_channels, out_channels=out_channel, **decoder_params
+        )

cosyvoice/matcha/transformer.py

@@ -0,0 +1,443 @@
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn as nn
+from diffusers.models.attention import (
+    GEGLU,
+    GELU,
+    AdaLayerNorm,
+    AdaLayerNormZero,
+    ApproximateGELU,
+)
+from diffusers.models.attention_processor import Attention
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+
+
+class SnakeBeta(nn.Module):
+    """
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self,
+        in_features,
+        out_features,
+        alpha=1.0,
+        alpha_trainable=True,
+        alpha_logscale=True,
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        """
+        super().__init__()
+        self.in_features = (
+            out_features if isinstance(out_features, list) else [out_features]
+        )
+        self.proj = LoRACompatibleLinear(in_features, out_features)
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # log scale alphas initialized to zeros
+            self.alpha = nn.Parameter(torch.zeros(self.in_features) * alpha)
+            self.beta = nn.Parameter(torch.zeros(self.in_features) * alpha)
+        else:  # linear scale alphas initialized to ones
+            self.alpha = nn.Parameter(torch.ones(self.in_features) * alpha)
+            self.beta = nn.Parameter(torch.ones(self.in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        """
+        x = self.proj(x)
+        if self.alpha_logscale:
+            alpha = torch.exp(self.alpha)
+            beta = torch.exp(self.beta)
+        else:
+            alpha = self.alpha
+            beta = self.beta
+
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * torch.pow(
+            torch.sin(x * alpha), 2
+        )
+
+        return x
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh")
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim)
+        elif activation_fn == "snakebeta":
+            act_fn = SnakeBeta(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(LoRACompatibleLinear(inner_dim, dim_out))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states):
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (
+            num_embeds_ada_norm is not None
+        ) and norm_type == "ada_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=(
+                    cross_attention_dim if not double_self_attention else None
+                ),
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                # scale_qk=False, # uncomment this to not to use flash attention
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+        )
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward_native(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ):
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = (
+            cross_attention_kwargs if cross_attention_kwargs is not None else {}
+        )
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=(
+                encoder_hidden_states if self.only_cross_attention else None
+            ),
+            attention_mask=(
+                encoder_attention_mask if self.only_cross_attention else attention_mask
+            ),
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [
+                    self.ff(hid_slice)
+                    for hid_slice in norm_hidden_states.chunk(
+                        num_chunks, dim=self._chunk_dim
+                    )
+                ],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        return hidden_states
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ):
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = (
+            cross_attention_kwargs if cross_attention_kwargs is not None else {}
+        )
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=(
+                encoder_hidden_states if self.only_cross_attention else None
+            ),
+            attention_mask=(
+                encoder_attention_mask if self.only_cross_attention else attention_mask
+            ),
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep)
+                if self.use_ada_layer_norm
+                else self.norm2(hidden_states)
+            )
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = (
+                norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+            )
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [
+                    self.ff(hid_slice)
+                    for hid_slice in norm_hidden_states.chunk(
+                        num_chunks, dim=self._chunk_dim
+                    )
+                ],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        return hidden_states

cosyvoice/transformer/activation.py

@@ -0,0 +1,87 @@
+# Copyright (c) 2020 Johns Hopkins University (Shinji Watanabe)
+#               2020 Northwestern Polytechnical University (Pengcheng Guo)
+#               2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+"""Swish() activation function for Conformer."""
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Swish(torch.nn.Module):
+    """Construct an Swish object."""
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Return Swish activation function."""
+        return x * torch.sigmoid(x)
+
+
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+class Snake(nn.Module):
+    """
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        """
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else:  # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        """
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

cosyvoice/transformer/attention.py

@@ -0,0 +1,322 @@
+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#               2022 Xingchen Song ([email protected])
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+"""Multi-Head Attention layer definition."""
+
+import math
+from typing import Tuple
+
+import torch
+from torch import nn
+
+
+class MultiHeadedAttention(nn.Module):
+    """Multi-Head Attention layer.
+
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, n_head: int, n_feat: int, dropout_rate: float, key_bias: bool = True
+    ):
+        """Construct an MultiHeadedAttention object."""
+        super().__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat, bias=key_bias)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.dropout = nn.Dropout(p=dropout_rate)
+
+    def forward_qkv(
+        self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Transform query, key and value.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+
+        Returns:
+            torch.Tensor: Transformed query tensor, size
+                (#batch, n_head, time1, d_k).
+            torch.Tensor: Transformed key tensor, size
+                (#batch, n_head, time2, d_k).
+            torch.Tensor: Transformed value tensor, size
+                (#batch, n_head, time2, d_k).
+
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+
+        return q, k, v
+
+    def forward_attention(
+        self,
+        value: torch.Tensor,
+        scores: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+    ) -> torch.Tensor:
+        """Compute attention context vector.
+
+        Args:
+            value (torch.Tensor): Transformed value, size
+                (#batch, n_head, time2, d_k).
+            scores (torch.Tensor): Attention score, size
+                (#batch, n_head, time1, time2).
+            mask (torch.Tensor): Mask, size (#batch, 1, time2) or
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
+
+        Returns:
+            torch.Tensor: Transformed value (#batch, time1, d_model)
+                weighted by the attention score (#batch, time1, time2).
+
+        """
+        n_batch = value.size(0)
+        # NOTE(xcsong): When will `if mask.size(2) > 0` be True?
+        #   1. onnx(16/4) [WHY? Because we feed real cache & real mask for the
+        #           1st chunk to ease the onnx export.]
+        #   2. pytorch training
+        if mask.size(2) > 0:  # time2 > 0
+            mask = mask.unsqueeze(1).eq(0)  # (batch, 1, *, time2)
+            # For last chunk, time2 might be larger than scores.size(-1)
+            mask = mask[:, :, :, : scores.size(-1)]  # (batch, 1, *, time2)
+            scores = scores.masked_fill(mask, -float("inf"))
+            attn = torch.softmax(scores, dim=-1).masked_fill(
+                mask, 0.0
+            )  # (batch, head, time1, time2)
+        # NOTE(xcsong): When will `if mask.size(2) > 0` be False?
+        #   1. onnx(16/-1, -1/-1, 16/0)
+        #   2. jit (16/-1, -1/-1, 16/0, 16/4)
+        else:
+            attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        p_attn = self.dropout(attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (
+            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+        )  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        pos_emb: torch.Tensor = torch.empty(0),
+        cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute scaled dot product attention.
+
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2).
+                1.When applying cross attention between decoder and encoder,
+                the batch padding mask for input is in (#batch, 1, T) shape.
+                2.When applying self attention of encoder,
+                the mask is in (#batch, T, T)  shape.
+                3.When applying self attention of decoder,
+                the mask is in (#batch, L, L)  shape.
+                4.If the different position in decoder see different block
+                of the encoder, such as Mocha, the passed in mask could be
+                in (#batch, L, T) shape. But there is no such case in current
+                CosyVoice.
+            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+
+        # NOTE(xcsong):
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if cache.size(0) > 0:
+            key_cache, value_cache = torch.split(cache, cache.size(-1) // 2, dim=-1)
+            k = torch.cat([key_cache, k], dim=2)
+            v = torch.cat([value_cache, v], dim=2)
+        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = torch.cat((k, v), dim=-1)
+
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        return self.forward_attention(v, scores, mask), new_cache
+
+
+class RelPositionMultiHeadedAttention(MultiHeadedAttention):
+    """Multi-Head Attention layer with relative position encoding.
+    Paper: https://arxiv.org/abs/1901.02860
+    Args:
+        n_head (int): The number of heads.
+        n_feat (int): The number of features.
+        dropout_rate (float): Dropout rate.
+    """
+
+    def __init__(
+        self, n_head: int, n_feat: int, dropout_rate: float, key_bias: bool = True
+    ):
+        """Construct an RelPositionMultiHeadedAttention object."""
+        super().__init__(n_head, n_feat, dropout_rate, key_bias)
+        # linear transformation for positional encoding
+        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        torch.nn.init.xavier_uniform_(self.pos_bias_u)
+        torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x: torch.Tensor) -> torch.Tensor:
+        """Compute relative positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, head, time1, 2*time1-1).
+            time1 means the length of query vector.
+
+        Returns:
+            torch.Tensor: Output tensor.
+
+        """
+        zero_pad = torch.zeros(
+            (x.size()[0], x.size()[1], x.size()[2], 1), device=x.device, dtype=x.dtype
+        )
+        x_padded = torch.cat([zero_pad, x], dim=-1)
+
+        x_padded = x_padded.view(x.size()[0], x.size()[1], x.size(3) + 1, x.size(2))
+        x = x_padded[:, :, 1:].view_as(x)[
+            :, :, :, : x.size(-1) // 2 + 1
+        ]  # only keep the positions from 0 to time2
+        return x
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        pos_emb: torch.Tensor = torch.empty(0),
+        cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute 'Scaled Dot Product Attention' with rel. positional encoding.
+        Args:
+            query (torch.Tensor): Query tensor (#batch, time1, size).
+            key (torch.Tensor): Key tensor (#batch, time2, size).
+            value (torch.Tensor): Value tensor (#batch, time2, size).
+            mask (torch.Tensor): Mask tensor (#batch, 1, time2) or
+                (#batch, time1, time2), (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): Positional embedding tensor
+                (#batch, time2, size).
+            cache (torch.Tensor): Cache tensor (1, head, cache_t, d_k * 2),
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+        Returns:
+            torch.Tensor: Output tensor (#batch, time1, d_model).
+            torch.Tensor: Cache tensor (1, head, cache_t + time1, d_k * 2)
+                where `cache_t == chunk_size * num_decoding_left_chunks`
+                and `head * d_k == size`
+        """
+        q, k, v = self.forward_qkv(query, key, value)
+        q = q.transpose(1, 2)  # (batch, time1, head, d_k)
+
+        # NOTE(xcsong):
+        #   when export onnx model, for 1st chunk, we feed
+        #       cache(1, head, 0, d_k * 2) (16/-1, -1/-1, 16/0 mode)
+        #       or cache(1, head, real_cache_t, d_k * 2) (16/4 mode).
+        #       In all modes, `if cache.size(0) > 0` will alwayse be `True`
+        #       and we will always do splitting and
+        #       concatnation(this will simplify onnx export). Note that
+        #       it's OK to concat & split zero-shaped tensors(see code below).
+        #   when export jit  model, for 1st chunk, we always feed
+        #       cache(0, 0, 0, 0) since jit supports dynamic if-branch.
+        # >>> a = torch.ones((1, 2, 0, 4))
+        # >>> b = torch.ones((1, 2, 3, 4))
+        # >>> c = torch.cat((a, b), dim=2)
+        # >>> torch.equal(b, c)        # True
+        # >>> d = torch.split(a, 2, dim=-1)
+        # >>> torch.equal(d[0], d[1])  # True
+        if cache.size(0) > 0:
+            key_cache, value_cache = torch.split(cache, cache.size(-1) // 2, dim=-1)
+            k = torch.cat([key_cache, k], dim=2)
+            v = torch.cat([value_cache, v], dim=2)
+        # NOTE(xcsong): We do cache slicing in encoder.forward_chunk, since it's
+        #   non-trivial to calculate `next_cache_start` here.
+        new_cache = torch.cat((k, v), dim=-1)
+
+        n_batch_pos = pos_emb.size(0)
+        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
+        p = p.transpose(1, 2)  # (batch, head, time1, d_k)
+
+        # (batch, head, time1, d_k)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        # (batch, head, time1, d_k)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        # (batch, head, time1, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
+
+        # compute matrix b and matrix d
+        # (batch, head, time1, time2)
+        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
+        # NOTE(Xiang Lyu): Keep rel_shift since espnet rel_pos_emb is used
+        if matrix_ac.shape != matrix_bd.shape:
+            matrix_bd = self.rel_shift(matrix_bd)
+
+        scores = (matrix_ac + matrix_bd) / math.sqrt(
+            self.d_k
+        )  # (batch, head, time1, time2)
+
+        return self.forward_attention(v, scores, mask), new_cache

cosyvoice/transformer/convolution.py

@@ -0,0 +1,147 @@
+# Copyright (c) 2020 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""ConvolutionModule definition."""
+
+from typing import Tuple
+
+import torch
+from torch import nn
+
+
+class ConvolutionModule(nn.Module):
+    """ConvolutionModule in Conformer model."""
+
+    def __init__(
+        self,
+        channels: int,
+        kernel_size: int = 15,
+        activation: nn.Module = nn.ReLU(),
+        norm: str = "batch_norm",
+        causal: bool = False,
+        bias: bool = True,
+    ):
+        """Construct an ConvolutionModule object.
+        Args:
+            channels (int): The number of channels of conv layers.
+            kernel_size (int): Kernel size of conv layers.
+            causal (int): Whether use causal convolution or not
+        """
+        super().__init__()
+
+        self.pointwise_conv1 = nn.Conv1d(
+            channels,
+            2 * channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        # self.lorder is used to distinguish if it's a causal convolution,
+        # if self.lorder > 0: it's a causal convolution, the input will be
+        #    padded with self.lorder frames on the left in forward.
+        # else: it's a symmetrical convolution
+        if causal:
+            padding = 0
+            self.lorder = kernel_size - 1
+        else:
+            # kernel_size should be an odd number for none causal convolution
+            assert (kernel_size - 1) % 2 == 0
+            padding = (kernel_size - 1) // 2
+            self.lorder = 0
+        self.depthwise_conv = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size,
+            stride=1,
+            padding=padding,
+            groups=channels,
+            bias=bias,
+        )
+
+        assert norm in ["batch_norm", "layer_norm"]
+        if norm == "batch_norm":
+            self.use_layer_norm = False
+            self.norm = nn.BatchNorm1d(channels)
+        else:
+            self.use_layer_norm = True
+            self.norm = nn.LayerNorm(channels)
+
+        self.pointwise_conv2 = nn.Conv1d(
+            channels,
+            channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=bias,
+        )
+        self.activation = activation
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        cache: torch.Tensor = torch.zeros((0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute convolution module.
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, channels).
+            mask_pad (torch.Tensor): used for batch padding (#batch, 1, time),
+                (0, 0, 0) means fake mask.
+            cache (torch.Tensor): left context cache, it is only
+                used in causal convolution (#batch, channels, cache_t),
+                (0, 0, 0) meas fake cache.
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, channels).
+        """
+        # exchange the temporal dimension and the feature dimension
+        x = x.transpose(1, 2)  # (#batch, channels, time)
+
+        # mask batch padding
+        if mask_pad.size(2) > 0:  # time > 0
+            x.masked_fill_(~mask_pad, 0.0)
+
+        if self.lorder > 0:
+            if cache.size(2) == 0:  # cache_t == 0
+                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
+            else:
+                assert cache.size(0) == x.size(0)  # equal batch
+                assert cache.size(1) == x.size(1)  # equal channel
+                x = torch.cat((cache, x), dim=2)
+            assert x.size(2) > self.lorder
+            new_cache = x[:, :, -self.lorder :]
+        else:
+            # It's better we just return None if no cache is required,
+            # However, for JIT export, here we just fake one tensor instead of
+            # None.
+            new_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+
+        # GLU mechanism
+        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
+        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)
+
+        # 1D Depthwise Conv
+        x = self.depthwise_conv(x)
+        if self.use_layer_norm:
+            x = x.transpose(1, 2)
+        x = self.activation(self.norm(x))
+        if self.use_layer_norm:
+            x = x.transpose(1, 2)
+        x = self.pointwise_conv2(x)
+        # mask batch padding
+        if mask_pad.size(2) > 0:  # time > 0
+            x.masked_fill_(~mask_pad, 0.0)
+
+        return x.transpose(1, 2), new_cache

cosyvoice/transformer/decoder.py

@@ -0,0 +1,418 @@
+# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Decoder definition."""
+from typing import Tuple, List, Optional
+
+import torch
+import torch.utils.checkpoint as ckpt
+import logging
+
+from cosyvoice.transformer.decoder_layer import DecoderLayer
+from cosyvoice.transformer.positionwise_feed_forward import (
+    PositionwiseFeedForward,
+)
+from cosyvoice.utils.class_utils import (
+    COSYVOICE_EMB_CLASSES,
+    COSYVOICE_ATTENTION_CLASSES,
+    COSYVOICE_ACTIVATION_CLASSES,
+)
+from cosyvoice.utils.mask import subsequent_mask, make_pad_mask
+
+
+class TransformerDecoder(torch.nn.Module):
+    """Base class of Transfomer decoder module.
+    Args:
+        vocab_size: output dim
+        encoder_output_size: dimension of attention
+        attention_heads: the number of heads of multi head attention
+        linear_units: the hidden units number of position-wise feedforward
+        num_blocks: the number of decoder blocks
+        dropout_rate: dropout rate
+        self_attention_dropout_rate: dropout rate for attention
+        input_layer: input layer type
+        use_output_layer: whether to use output layer
+        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
+        normalize_before:
+            True: use layer_norm before each sub-block of a layer.
+            False: use layer_norm after each sub-block of a layer.
+        src_attention: if false, encoder-decoder cross attention is not
+                       applied, such as CIF model
+        key_bias: whether use bias in attention.linear_k, False for whisper models.
+        gradient_checkpointing: rerunning a forward-pass segment for each
+            checkpointed segment during backward.
+        tie_word_embedding: Tie or clone module weights depending of whether we are
+            using TorchScript or not
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        encoder_output_size: int,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        self_attention_dropout_rate: float = 0.0,
+        src_attention_dropout_rate: float = 0.0,
+        input_layer: str = "embed",
+        use_output_layer: bool = True,
+        normalize_before: bool = True,
+        src_attention: bool = True,
+        key_bias: bool = True,
+        activation_type: str = "relu",
+        gradient_checkpointing: bool = False,
+        tie_word_embedding: bool = False,
+    ):
+        super().__init__()
+        attention_dim = encoder_output_size
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+
+        self.embed = torch.nn.Sequential(
+            (
+                torch.nn.Identity()
+                if input_layer == "no_pos"
+                else torch.nn.Embedding(vocab_size, attention_dim)
+            ),
+            COSYVOICE_EMB_CLASSES[input_layer](attention_dim, positional_dropout_rate),
+        )
+
+        self.normalize_before = normalize_before
+        self.after_norm = torch.nn.LayerNorm(attention_dim, eps=1e-5)
+        self.use_output_layer = use_output_layer
+        if use_output_layer:
+            self.output_layer = torch.nn.Linear(attention_dim, vocab_size)
+        else:
+            self.output_layer = torch.nn.Identity()
+        self.num_blocks = num_blocks
+        self.decoders = torch.nn.ModuleList(
+            [
+                DecoderLayer(
+                    attention_dim,
+                    COSYVOICE_ATTENTION_CLASSES["selfattn"](
+                        attention_heads,
+                        attention_dim,
+                        self_attention_dropout_rate,
+                        key_bias,
+                    ),
+                    (
+                        COSYVOICE_ATTENTION_CLASSES["selfattn"](
+                            attention_heads,
+                            attention_dim,
+                            src_attention_dropout_rate,
+                            key_bias,
+                        )
+                        if src_attention
+                        else None
+                    ),
+                    PositionwiseFeedForward(
+                        attention_dim, linear_units, dropout_rate, activation
+                    ),
+                    dropout_rate,
+                    normalize_before,
+                )
+                for _ in range(self.num_blocks)
+            ]
+        )
+
+        self.gradient_checkpointing = gradient_checkpointing
+        self.tie_word_embedding = tie_word_embedding
+
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        ys_in_pad: torch.Tensor,
+        ys_in_lens: torch.Tensor,
+        r_ys_in_pad: torch.Tensor = torch.empty(0),
+        reverse_weight: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward decoder.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoder memory mask, (batch, 1, maxlen_in)
+            ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
+            ys_in_lens: input lengths of this batch (batch)
+            r_ys_in_pad: not used in transformer decoder, in order to unify api
+                with bidirectional decoder
+            reverse_weight: not used in transformer decoder, in order to unify
+                api with bidirectional decode
+        Returns:
+            (tuple): tuple containing:
+                x: decoded token score before softmax (batch, maxlen_out,
+                    vocab_size) if use_output_layer is True,
+                torch.tensor(0.0), in order to unify api with bidirectional decoder
+                olens: (batch, )
+        NOTE(xcsong):
+            We pass the `__call__` method of the modules instead of `forward` to the
+            checkpointing API because `__call__` attaches all the hooks of the module.
+            https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2
+        """
+        tgt = ys_in_pad
+        maxlen = tgt.size(1)
+        # tgt_mask: (B, 1, L)
+        tgt_mask = ~make_pad_mask(ys_in_lens, maxlen).unsqueeze(1)
+        tgt_mask = tgt_mask.to(tgt.device)
+        # m: (1, L, L)
+        m = subsequent_mask(tgt_mask.size(-1), device=tgt_mask.device).unsqueeze(0)
+        # tgt_mask: (B, L, L)
+        tgt_mask = tgt_mask & m
+        x, _ = self.embed(tgt)
+        if self.gradient_checkpointing and self.training:
+            x = self.forward_layers_checkpointed(x, tgt_mask, memory, memory_mask)
+        else:
+            x = self.forward_layers(x, tgt_mask, memory, memory_mask)
+        if self.normalize_before:
+            x = self.after_norm(x)
+        if self.use_output_layer:
+            x = self.output_layer(x)
+        olens = tgt_mask.sum(1)
+        return x, torch.tensor(0.0), olens
+
+    def forward_layers(
+        self,
+        x: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        for layer in self.decoders:
+            x, tgt_mask, memory, memory_mask = layer(x, tgt_mask, memory, memory_mask)
+        return x
+
+    @torch.jit.unused
+    def forward_layers_checkpointed(
+        self,
+        x: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+    ) -> torch.Tensor:
+        for layer in self.decoders:
+            x, tgt_mask, memory, memory_mask = ckpt.checkpoint(
+                layer.__call__, x, tgt_mask, memory, memory_mask
+            )
+        return x
+
+    def forward_one_step(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        cache: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward one step.
+            This is only used for decoding.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoded memory mask, (batch, 1, maxlen_in)
+            tgt: input token ids, int64 (batch, maxlen_out)
+            tgt_mask: input token mask,  (batch, maxlen_out)
+                      dtype=torch.uint8 in PyTorch 1.2-
+                      dtype=torch.bool in PyTorch 1.2+ (include 1.2)
+            cache: cached output list of (batch, max_time_out-1, size)
+        Returns:
+            y, cache: NN output value and cache per `self.decoders`.
+            y.shape` is (batch, maxlen_out, token)
+        """
+        x, _ = self.embed(tgt)
+        new_cache = []
+        for i, decoder in enumerate(self.decoders):
+            if cache is None:
+                c = None
+            else:
+                c = cache[i]
+            x, tgt_mask, memory, memory_mask = decoder(
+                x, tgt_mask, memory, memory_mask, cache=c
+            )
+            new_cache.append(x)
+        if self.normalize_before:
+            y = self.after_norm(x[:, -1])
+        else:
+            y = x[:, -1]
+        if self.use_output_layer:
+            y = torch.log_softmax(self.output_layer(y), dim=-1)
+        return y, new_cache
+
+    def tie_or_clone_weights(self, jit_mode: bool = True):
+        """Tie or clone module weights (between word_emb and output_layer)
+        depending of whether we are using TorchScript or not"""
+        if not self.use_output_layer:
+            return
+        if jit_mode:
+            logging.info("clone emb.weight to output.weight")
+            self.output_layer.weight = torch.nn.Parameter(self.embed[0].weight.clone())
+        else:
+            logging.info("tie emb.weight with output.weight")
+            self.output_layer.weight = self.embed[0].weight
+
+        if getattr(self.output_layer, "bias", None) is not None:
+            self.output_layer.bias.data = torch.nn.functional.pad(
+                self.output_layer.bias.data,
+                (
+                    0,
+                    self.output_layer.weight.shape[0] - self.output_layer.bias.shape[0],
+                ),
+                "constant",
+                0,
+            )
+
+
+class BiTransformerDecoder(torch.nn.Module):
+    """Base class of Transfomer decoder module.
+    Args:
+        vocab_size: output dim
+        encoder_output_size: dimension of attention
+        attention_heads: the number of heads of multi head attention
+        linear_units: the hidden units number of position-wise feedforward
+        num_blocks: the number of decoder blocks
+        r_num_blocks: the number of right to left decoder blocks
+        dropout_rate: dropout rate
+        self_attention_dropout_rate: dropout rate for attention
+        input_layer: input layer type
+        use_output_layer: whether to use output layer
+        pos_enc_class: PositionalEncoding or ScaledPositionalEncoding
+        normalize_before:
+            True: use layer_norm before each sub-block of a layer.
+            False: use layer_norm after each sub-block of a layer.
+        key_bias: whether use bias in attention.linear_k, False for whisper models.
+    """
+
+    def __init__(
+        self,
+        vocab_size: int,
+        encoder_output_size: int,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        r_num_blocks: int = 0,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        self_attention_dropout_rate: float = 0.0,
+        src_attention_dropout_rate: float = 0.0,
+        input_layer: str = "embed",
+        use_output_layer: bool = True,
+        normalize_before: bool = True,
+        key_bias: bool = True,
+        gradient_checkpointing: bool = False,
+        tie_word_embedding: bool = False,
+    ):
+
+        super().__init__()
+        self.tie_word_embedding = tie_word_embedding
+        self.left_decoder = TransformerDecoder(
+            vocab_size,
+            encoder_output_size,
+            attention_heads,
+            linear_units,
+            num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            self_attention_dropout_rate,
+            src_attention_dropout_rate,
+            input_layer,
+            use_output_layer,
+            normalize_before,
+            key_bias=key_bias,
+            gradient_checkpointing=gradient_checkpointing,
+            tie_word_embedding=tie_word_embedding,
+        )
+
+        self.right_decoder = TransformerDecoder(
+            vocab_size,
+            encoder_output_size,
+            attention_heads,
+            linear_units,
+            r_num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            self_attention_dropout_rate,
+            src_attention_dropout_rate,
+            input_layer,
+            use_output_layer,
+            normalize_before,
+            key_bias=key_bias,
+            gradient_checkpointing=gradient_checkpointing,
+            tie_word_embedding=tie_word_embedding,
+        )
+
+    def forward(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        ys_in_pad: torch.Tensor,
+        ys_in_lens: torch.Tensor,
+        r_ys_in_pad: torch.Tensor,
+        reverse_weight: float = 0.0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Forward decoder.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoder memory mask, (batch, 1, maxlen_in)
+            ys_in_pad: padded input token ids, int64 (batch, maxlen_out)
+            ys_in_lens: input lengths of this batch (batch)
+            r_ys_in_pad: padded input token ids, int64 (batch, maxlen_out),
+                used for right to left decoder
+            reverse_weight: used for right to left decoder
+        Returns:
+            (tuple): tuple containing:
+                x: decoded token score before softmax (batch, maxlen_out,
+                    vocab_size) if use_output_layer is True,
+                r_x: x: decoded token score (right to left decoder)
+                    before softmax (batch, maxlen_out, vocab_size)
+                    if use_output_layer is True,
+                olens: (batch, )
+        """
+        l_x, _, olens = self.left_decoder(memory, memory_mask, ys_in_pad, ys_in_lens)
+        r_x = torch.tensor(0.0)
+        if reverse_weight > 0.0:
+            r_x, _, olens = self.right_decoder(
+                memory, memory_mask, r_ys_in_pad, ys_in_lens
+            )
+        return l_x, r_x, olens
+
+    def forward_one_step(
+        self,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        cache: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        """Forward one step.
+            This is only used for decoding.
+        Args:
+            memory: encoded memory, float32  (batch, maxlen_in, feat)
+            memory_mask: encoded memory mask, (batch, 1, maxlen_in)
+            tgt: input token ids, int64 (batch, maxlen_out)
+            tgt_mask: input token mask,  (batch, maxlen_out)
+                      dtype=torch.uint8 in PyTorch 1.2-
+                      dtype=torch.bool in PyTorch 1.2+ (include 1.2)
+            cache: cached output list of (batch, max_time_out-1, size)
+        Returns:
+            y, cache: NN output value and cache per `self.decoders`.
+            y.shape` is (batch, maxlen_out, token)
+        """
+        return self.left_decoder.forward_one_step(
+            memory, memory_mask, tgt, tgt_mask, cache
+        )
+
+    def tie_or_clone_weights(self, jit_mode: bool = True):
+        """Tie or clone module weights (between word_emb and output_layer)
+        depending of whether we are using TorchScript or not"""
+        self.left_decoder.tie_or_clone_weights(jit_mode)
+        self.right_decoder.tie_or_clone_weights(jit_mode)

cosyvoice/transformer/decoder_layer.py

@@ -0,0 +1,132 @@
+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# 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.
+"""Decoder self-attention layer definition."""
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+
+
+class DecoderLayer(nn.Module):
+    """Single decoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+        src_attn (torch.nn.Module): Inter-attention module instance.
+            `MultiHeadedAttention` instance can be used as the argument.
+            If `None` is passed, Inter-attention is not used, such as
+            CIF, GPT, and other decoder only model.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: to use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: nn.Module,
+        src_attn: Optional[nn.Module],
+        feed_forward: nn.Module,
+        dropout_rate: float,
+        normalize_before: bool = True,
+    ):
+        """Construct an DecoderLayer object."""
+        super().__init__()
+        self.size = size
+        self.self_attn = self_attn
+        self.src_attn = src_attn
+        self.feed_forward = feed_forward
+        self.norm1 = nn.LayerNorm(size, eps=1e-5)
+        self.norm2 = nn.LayerNorm(size, eps=1e-5)
+        self.norm3 = nn.LayerNorm(size, eps=1e-5)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        tgt: torch.Tensor,
+        tgt_mask: torch.Tensor,
+        memory: torch.Tensor,
+        memory_mask: torch.Tensor,
+        cache: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute decoded features.
+
+        Args:
+            tgt (torch.Tensor): Input tensor (#batch, maxlen_out, size).
+            tgt_mask (torch.Tensor): Mask for input tensor
+                (#batch, maxlen_out).
+            memory (torch.Tensor): Encoded memory
+                (#batch, maxlen_in, size).
+            memory_mask (torch.Tensor): Encoded memory mask
+                (#batch, maxlen_in).
+            cache (torch.Tensor): cached tensors.
+                (#batch, maxlen_out - 1, size).
+
+        Returns:
+            torch.Tensor: Output tensor (#batch, maxlen_out, size).
+            torch.Tensor: Mask for output tensor (#batch, maxlen_out).
+            torch.Tensor: Encoded memory (#batch, maxlen_in, size).
+            torch.Tensor: Encoded memory mask (#batch, maxlen_in).
+
+        """
+        residual = tgt
+        if self.normalize_before:
+            tgt = self.norm1(tgt)
+
+        if cache is None:
+            tgt_q = tgt
+            tgt_q_mask = tgt_mask
+        else:
+            # compute only the last frame query keeping dim: max_time_out -> 1
+            assert cache.shape == (
+                tgt.shape[0],
+                tgt.shape[1] - 1,
+                self.size,
+            ), "{cache.shape} == {(tgt.shape[0], tgt.shape[1] - 1, self.size)}"
+            tgt_q = tgt[:, -1:, :]
+            residual = residual[:, -1:, :]
+            tgt_q_mask = tgt_mask[:, -1:, :]
+
+        x = residual + self.dropout(self.self_attn(tgt_q, tgt, tgt, tgt_q_mask)[0])
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        if self.src_attn is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm2(x)
+            x = residual + self.dropout(
+                self.src_attn(x, memory, memory, memory_mask)[0]
+            )
+            if not self.normalize_before:
+                x = self.norm2(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm3(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm3(x)
+
+        if cache is not None:
+            x = torch.cat([cache, x], dim=1)
+
+        return x, tgt_mask, memory, memory_mask

cosyvoice/transformer/embedding.py

@@ -0,0 +1,293 @@
+# Copyright (c) 2020 Mobvoi Inc. (authors: Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Positonal Encoding Module."""
+
+import math
+from typing import Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+
+
+class PositionalEncoding(torch.nn.Module):
+    """Positional encoding.
+
+    :param int d_model: embedding dim
+    :param float dropout_rate: dropout rate
+    :param int max_len: maximum input length
+
+    PE(pos, 2i)   = sin(pos/(10000^(2i/dmodel)))
+    PE(pos, 2i+1) = cos(pos/(10000^(2i/dmodel)))
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        dropout_rate: float,
+        max_len: int = 5000,
+        reverse: bool = False,
+    ):
+        """Construct an PositionalEncoding object."""
+        super().__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.max_len = max_len
+
+        self.pe = torch.zeros(self.max_len, self.d_model)
+        position = torch.arange(0, self.max_len, dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        self.pe[:, 0::2] = torch.sin(position * div_term)
+        self.pe[:, 1::2] = torch.cos(position * div_term)
+        self.pe = self.pe.unsqueeze(0)
+
+    def forward(
+        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input. Its shape is (batch, time, ...)
+            offset (int, torch.tensor): position offset
+
+        Returns:
+            torch.Tensor: Encoded tensor. Its shape is (batch, time, ...)
+            torch.Tensor: for compatibility to RelPositionalEncoding
+        """
+
+        self.pe = self.pe.to(x.device)
+        pos_emb = self.position_encoding(offset, x.size(1), False)
+        x = x * self.xscale + pos_emb
+        return self.dropout(x), self.dropout(pos_emb)
+
+    def position_encoding(
+        self, offset: Union[int, torch.Tensor], size: int, apply_dropout: bool = True
+    ) -> torch.Tensor:
+        """For getting encoding in a streaming fashion
+
+        Attention!!!!!
+        we apply dropout only once at the whole utterance level in a none
+        streaming way, but will call this function several times with
+        increasing input size in a streaming scenario, so the dropout will
+        be applied several times.
+
+        Args:
+            offset (int or torch.tensor): start offset
+            size (int): required size of position encoding
+
+        Returns:
+            torch.Tensor: Corresponding encoding
+        """
+        # How to subscript a Union type:
+        #   https://github.com/pytorch/pytorch/issues/69434
+        if isinstance(offset, int):
+            assert offset + size <= self.max_len
+            pos_emb = self.pe[:, offset : offset + size]
+        elif isinstance(offset, torch.Tensor) and offset.dim() == 0:  # scalar
+            assert offset + size <= self.max_len
+            pos_emb = self.pe[:, offset : offset + size]
+        else:  # for batched streaming decoding on GPU
+            assert torch.max(offset) + size <= self.max_len
+            index = offset.unsqueeze(1) + torch.arange(0, size).to(
+                offset.device
+            )  # B X T
+            flag = index > 0
+            # remove negative offset
+            index = index * flag
+            pos_emb = F.embedding(index, self.pe[0])  # B X T X d_model
+
+        if apply_dropout:
+            pos_emb = self.dropout(pos_emb)
+        return pos_emb
+
+
+class RelPositionalEncoding(PositionalEncoding):
+    """Relative positional encoding module.
+    See : Appendix B in https://arxiv.org/abs/1901.02860
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000):
+        """Initialize class."""
+        super().__init__(d_model, dropout_rate, max_len, reverse=True)
+
+    def forward(
+        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Compute positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+            torch.Tensor: Positional embedding tensor (1, time, `*`).
+        """
+        self.pe = self.pe.to(x.device)
+        x = x * self.xscale
+        pos_emb = self.position_encoding(offset, x.size(1), False)
+        return self.dropout(x), self.dropout(pos_emb)
+
+
+class WhisperPositionalEncoding(PositionalEncoding):
+    """Sinusoids position encoding used in openai-whisper.encoder"""
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 1500):
+        super().__init__(d_model, dropout_rate, max_len)
+        self.xscale = 1.0
+        log_timescale_increment = np.log(10000) / (d_model // 2 - 1)
+        inv_timescales = torch.exp(
+            -log_timescale_increment * torch.arange(d_model // 2)
+        )
+        scaled_time = (
+            torch.arange(max_len)[:, np.newaxis] * inv_timescales[np.newaxis, :]
+        )
+        pe = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], dim=1)
+        delattr(self, "pe")
+        self.register_buffer("pe", pe.unsqueeze(0))
+
+
+class LearnablePositionalEncoding(PositionalEncoding):
+    """Learnable position encoding used in openai-whisper.decoder"""
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 448):
+        super().__init__(d_model, dropout_rate, max_len)
+        # NOTE(xcsong): overwrite self.pe & self.xscale
+        self.pe = torch.nn.Parameter(torch.empty(1, max_len, d_model))
+        self.xscale = 1.0
+
+
+class NoPositionalEncoding(torch.nn.Module):
+    """No position encoding"""
+
+    def __init__(self, d_model: int, dropout_rate: float):
+        super().__init__()
+        self.d_model = d_model
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+
+    def forward(
+        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Just return zero vector for interface compatibility"""
+        pos_emb = torch.zeros(1, x.size(1), self.d_model).to(x.device)
+        return self.dropout(x), pos_emb
+
+    def position_encoding(
+        self, offset: Union[int, torch.Tensor], size: int
+    ) -> torch.Tensor:
+        return torch.zeros(1, size, self.d_model)
+
+
+class EspnetRelPositionalEncoding(torch.nn.Module):
+    """Relative positional encoding module (new implementation).
+
+    Details can be found in https://github.com/espnet/espnet/pull/2816.
+
+    See : Appendix B in https://arxiv.org/abs/1901.02860
+
+    Args:
+        d_model (int): Embedding dimension.
+        dropout_rate (float): Dropout rate.
+        max_len (int): Maximum input length.
+
+    """
+
+    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000):
+        """Construct an PositionalEncoding object."""
+        super(EspnetRelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = torch.nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x: torch.Tensor):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in https://arxiv.org/abs/1901.02860
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(
+        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Add positional encoding.
+
+        Args:
+            x (torch.Tensor): Input tensor (batch, time, `*`).
+
+        Returns:
+            torch.Tensor: Encoded tensor (batch, time, `*`).
+
+        """
+        self.extend_pe(x)
+        x = x * self.xscale
+        pos_emb = self.position_encoding(size=x.size(1), offset=offset)
+        return self.dropout(x), self.dropout(pos_emb)
+
+    def position_encoding(
+        self, offset: Union[int, torch.Tensor], size: int
+    ) -> torch.Tensor:
+        """For getting encoding in a streaming fashion
+
+        Attention!!!!!
+        we apply dropout only once at the whole utterance level in a none
+        streaming way, but will call this function several times with
+        increasing input size in a streaming scenario, so the dropout will
+        be applied several times.
+
+        Args:
+            offset (int or torch.tensor): start offset
+            size (int): required size of position encoding
+
+        Returns:
+            torch.Tensor: Corresponding encoding
+        """
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2 - size + 1 : self.pe.size(1) // 2 + size,
+        ]
+        return pos_emb

cosyvoice/transformer/encoder.py

@@ -0,0 +1,633 @@
+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2022 Xingchen Song ([email protected])
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Encoder definition."""
+from typing import Tuple
+import time
+
+import torch
+import torch.utils.checkpoint as ckpt
+import torch.nn.functional as F
+
+from cosyvoice.transformer.convolution import ConvolutionModule
+from cosyvoice.transformer.encoder_layer import (
+    TransformerEncoderLayer,
+)
+from cosyvoice.transformer.encoder_layer import (
+    ConformerEncoderLayer,
+)
+from cosyvoice.transformer.positionwise_feed_forward import (
+    PositionwiseFeedForward,
+)
+from cosyvoice.utils.class_utils import (
+    COSYVOICE_EMB_CLASSES,
+    COSYVOICE_SUBSAMPLE_CLASSES,
+    COSYVOICE_ATTENTION_CLASSES,
+    COSYVOICE_ACTIVATION_CLASSES,
+)
+from cosyvoice.utils.mask import make_pad_mask
+from cosyvoice.utils.mask import add_optional_chunk_mask
+
+
+class BaseEncoder(torch.nn.Module):
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "abs_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        gradient_checkpointing: bool = False,
+    ):
+        """
+        Args:
+            input_size (int): input dim
+            output_size (int): dimension of attention
+            attention_heads (int): the number of heads of multi head attention
+            linear_units (int): the hidden units number of position-wise feed
+                forward
+            num_blocks (int): the number of decoder blocks
+            dropout_rate (float): dropout rate
+            attention_dropout_rate (float): dropout rate in attention
+            positional_dropout_rate (float): dropout rate after adding
+                positional encoding
+            input_layer (str): input layer type.
+                optional [linear, conv2d, conv2d6, conv2d8]
+            pos_enc_layer_type (str): Encoder positional encoding layer type.
+                opitonal [abs_pos, scaled_abs_pos, rel_pos, no_pos]
+            normalize_before (bool):
+                True: use layer_norm before each sub-block of a layer.
+                False: use layer_norm after each sub-block of a layer.
+            static_chunk_size (int): chunk size for static chunk training and
+                decoding
+            use_dynamic_chunk (bool): whether use dynamic chunk size for
+                training or not, You can only use fixed chunk(chunk_size > 0)
+                or dyanmic chunk size(use_dynamic_chunk = True)
+            global_cmvn (Optional[torch.nn.Module]): Optional GlobalCMVN module
+            use_dynamic_left_chunk (bool): whether use dynamic left chunk in
+                dynamic chunk training
+            key_bias: whether use bias in attention.linear_k, False for whisper models.
+            gradient_checkpointing: rerunning a forward-pass segment for each
+                checkpointed segment during backward.
+        """
+        super().__init__()
+        self._output_size = output_size
+
+        self.global_cmvn = global_cmvn
+        self.embed = COSYVOICE_SUBSAMPLE_CLASSES[input_layer](
+            input_size,
+            output_size,
+            dropout_rate,
+            COSYVOICE_EMB_CLASSES[pos_enc_layer_type](
+                output_size, positional_dropout_rate
+            ),
+        )
+
+        self.normalize_before = normalize_before
+        self.after_norm = torch.nn.LayerNorm(output_size, eps=1e-5)
+        self.static_chunk_size = static_chunk_size
+        self.use_dynamic_chunk = use_dynamic_chunk
+        self.use_dynamic_left_chunk = use_dynamic_left_chunk
+        self.gradient_checkpointing = gradient_checkpointing
+
+    def output_size(self) -> int:
+        return self._output_size
+
+    def forward(
+        self,
+        xs: torch.Tensor,
+        xs_lens: torch.Tensor,
+        decoding_chunk_size: int = 0,
+        num_decoding_left_chunks: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Embed positions in tensor.
+
+        Args:
+            xs: padded input tensor (B, T, D)
+            xs_lens: input length (B)
+            decoding_chunk_size: decoding chunk size for dynamic chunk
+                0: default for training, use random dynamic chunk.
+                <0: for decoding, use full chunk.
+                >0: for decoding, use fixed chunk size as set.
+            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
+        Returns:
+            encoder output tensor xs, and subsampled masks
+            xs: padded output tensor (B, T' ~= T/subsample_rate, D)
+            masks: torch.Tensor batch padding mask after subsample
+                (B, 1, T' ~= T/subsample_rate)
+        NOTE(xcsong):
+            We pass the `__call__` method of the modules instead of `forward` to the
+            checkpointing API because `__call__` attaches all the hooks of the module.
+            https://discuss.pytorch.org/t/any-different-between-model-input-and-model-forward-input/3690/2
+        """
+        T = xs.size(1)
+        masks = ~make_pad_mask(xs_lens, T).unsqueeze(1)  # (B, 1, T)
+        if self.global_cmvn is not None:
+            xs = self.global_cmvn(xs)
+        xs, pos_emb, masks = self.embed(xs, masks)
+        mask_pad = masks  # (B, 1, T/subsample_rate)
+        chunk_masks = add_optional_chunk_mask(
+            xs,
+            masks,
+            self.use_dynamic_chunk,
+            self.use_dynamic_left_chunk,
+            decoding_chunk_size,
+            self.static_chunk_size,
+            num_decoding_left_chunks,
+        )
+        print(f"chunk_masks shape: {chunk_masks.shape}")
+        if self.gradient_checkpointing and self.training:
+            xs = self.forward_layers_checkpointed(xs, chunk_masks, pos_emb, mask_pad)
+        else:
+            xs = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad)
+        if self.normalize_before:
+            xs = self.after_norm(xs)
+        # Here we assume the mask is not changed in encoder layers, so just
+        # return the masks before encoder layers, and the masks will be used
+        # for cross attention with decoder later
+        return xs, masks
+
+    def forward_layers(
+        self,
+        xs: torch.Tensor,
+        chunk_masks: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor,
+    ) -> torch.Tensor:
+        for layer in self.encoders:
+            xs, chunk_masks, _, _ = layer(xs, chunk_masks, pos_emb, mask_pad)
+        return xs
+
+    @torch.jit.unused
+    def forward_layers_checkpointed(
+        self,
+        xs: torch.Tensor,
+        chunk_masks: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor,
+    ) -> torch.Tensor:
+        for layer in self.encoders:
+            xs, chunk_masks, _, _ = ckpt.checkpoint(
+                layer.__call__, xs, chunk_masks, pos_emb, mask_pad
+            )
+        return xs
+
+    @torch.jit.export
+    def forward_chunk(
+        self,
+        xs: torch.Tensor,
+        offset: int,
+        required_cache_size: int,
+        att_cache: torch.Tensor = torch.zeros(0, 0, 0, 0),
+        cnn_cache: torch.Tensor = torch.zeros(0, 0, 0, 0),
+        att_mask: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """ Forward just one chunk
+
+        Args:
+            xs (torch.Tensor): chunk input, with shape (b=1, time, mel-dim),
+                where `time == (chunk_size - 1) * subsample_rate + \
+                        subsample.right_context + 1`
+            offset (int): current offset in encoder output time stamp
+            required_cache_size (int): cache size required for next chunk
+                compuation
+                >=0: actual cache size
+                <0: means all history cache is required
+            att_cache (torch.Tensor): cache tensor for KEY & VALUE in
+                transformer/conformer attention, with shape
+                (elayers, head, cache_t1, d_k * 2), where
+                `head * d_k == hidden-dim` and
+                `cache_t1 == chunk_size * num_decoding_left_chunks`.
+            cnn_cache (torch.Tensor): cache tensor for cnn_module in conformer,
+                (elayers, b=1, hidden-dim, cache_t2), where
+                `cache_t2 == cnn.lorder - 1`
+
+        Returns:
+            torch.Tensor: output of current input xs,
+                with shape (b=1, chunk_size, hidden-dim).
+            torch.Tensor: new attention cache required for next chunk, with
+                dynamic shape (elayers, head, ?, d_k * 2)
+                depending on required_cache_size.
+            torch.Tensor: new conformer cnn cache required for next chunk, with
+                same shape as the original cnn_cache.
+
+        """
+        assert xs.size(0) == 1
+        # tmp_masks is just for interface compatibility
+        tmp_masks = torch.ones(1, xs.size(1), device=xs.device, dtype=torch.bool)
+        tmp_masks = tmp_masks.unsqueeze(1)
+        if self.global_cmvn is not None:
+            xs = self.global_cmvn(xs)
+        # NOTE(xcsong): Before embed, shape(xs) is (b=1, time, mel-dim)
+        xs, pos_emb, _ = self.embed(xs, tmp_masks, offset)
+        # NOTE(xcsong): After  embed, shape(xs) is (b=1, chunk_size, hidden-dim)
+        elayers, cache_t1 = att_cache.size(0), att_cache.size(2)
+        chunk_size = xs.size(1)
+        attention_key_size = cache_t1 + chunk_size
+        pos_emb = self.embed.position_encoding(
+            offset=offset - cache_t1, size=attention_key_size
+        )
+        if required_cache_size < 0:
+            next_cache_start = 0
+        elif required_cache_size == 0:
+            next_cache_start = attention_key_size
+        else:
+            next_cache_start = max(attention_key_size - required_cache_size, 0)
+        r_att_cache = []
+        r_cnn_cache = []
+        for i, layer in enumerate(self.encoders):
+            # NOTE(xcsong): Before layer.forward
+            #   shape(att_cache[i:i + 1]) is (1, head, cache_t1, d_k * 2),
+            #   shape(cnn_cache[i])       is (b=1, hidden-dim, cache_t2)
+            xs, _, new_att_cache, new_cnn_cache = layer(
+                xs,
+                att_mask,
+                pos_emb,
+                att_cache=att_cache[i : i + 1] if elayers > 0 else att_cache,
+                cnn_cache=cnn_cache[i] if cnn_cache.size(0) > 0 else cnn_cache,
+            )
+            # NOTE(xcsong): After layer.forward
+            #   shape(new_att_cache) is (1, head, attention_key_size, d_k * 2),
+            #   shape(new_cnn_cache) is (b=1, hidden-dim, cache_t2)
+            r_att_cache.append(new_att_cache[:, :, next_cache_start:, :])
+            r_cnn_cache.append(new_cnn_cache.unsqueeze(0))
+        if self.normalize_before:
+            xs = self.after_norm(xs)
+
+        # NOTE(xcsong): shape(r_att_cache) is (elayers, head, ?, d_k * 2),
+        #   ? may be larger than cache_t1, it depends on required_cache_size
+        r_att_cache = torch.cat(r_att_cache, dim=0)
+        # NOTE(xcsong): shape(r_cnn_cache) is (e, b=1, hidden-dim, cache_t2)
+        r_cnn_cache = torch.cat(r_cnn_cache, dim=0)
+
+        return (xs, r_att_cache, r_cnn_cache)
+
+    @torch.jit.unused
+    def forward_chunk_by_chunk(
+        self,
+        xs: torch.Tensor,
+        decoding_chunk_size: int,
+        num_decoding_left_chunks: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Forward input chunk by chunk with chunk_size like a streaming
+            fashion
+
+        Here we should pay special attention to computation cache in the
+        streaming style forward chunk by chunk. Three things should be taken
+        into account for computation in the current network:
+            1. transformer/conformer encoder layers output cache
+            2. convolution in conformer
+            3. convolution in subsampling
+
+        However, we don't implement subsampling cache for:
+            1. We can control subsampling module to output the right result by
+               overlapping input instead of cache left context, even though it
+               wastes some computation, but subsampling only takes a very
+               small fraction of computation in the whole model.
+            2. Typically, there are several covolution layers with subsampling
+               in subsampling module, it is tricky and complicated to do cache
+               with different convolution layers with different subsampling
+               rate.
+            3. Currently, nn.Sequential is used to stack all the convolution
+               layers in subsampling, we need to rewrite it to make it work
+               with cache, which is not preferred.
+        Args:
+            xs (torch.Tensor): (1, max_len, dim)
+            chunk_size (int): decoding chunk size
+        """
+        assert decoding_chunk_size > 0
+        # The model is trained by static or dynamic chunk
+        assert self.static_chunk_size > 0 or self.use_dynamic_chunk
+        subsampling = self.embed.subsampling_rate
+        context = self.embed.right_context + 1  # Add current frame
+        stride = subsampling * decoding_chunk_size
+        decoding_window = (decoding_chunk_size - 1) * subsampling + context
+        num_frames = xs.size(1)
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0), device=xs.device)
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0), device=xs.device)
+        outputs = []
+        offset = 0
+        required_cache_size = decoding_chunk_size * num_decoding_left_chunks
+
+        # Feed forward overlap input step by step
+        for cur in range(0, num_frames - context + 1, stride):
+            end = min(cur + decoding_window, num_frames)
+            chunk_xs = xs[:, cur:end, :]
+            (y, att_cache, cnn_cache) = self.forward_chunk(
+                chunk_xs, offset, required_cache_size, att_cache, cnn_cache
+            )
+            outputs.append(y)
+            offset += y.size(1)
+        ys = torch.cat(outputs, 1)
+        masks = torch.ones((1, 1, ys.size(1)), device=ys.device, dtype=torch.bool)
+        return ys, masks
+
+
+class TransformerEncoder(BaseEncoder):
+    """Transformer encoder module."""
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "abs_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        key_bias: bool = True,
+        selfattention_layer_type: str = "selfattn",
+        activation_type: str = "relu",
+        gradient_checkpointing: bool = False,
+    ):
+        """Construct TransformerEncoder
+
+        See Encoder for the meaning of each parameter.
+        """
+        super().__init__(
+            input_size,
+            output_size,
+            attention_heads,
+            linear_units,
+            num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            attention_dropout_rate,
+            input_layer,
+            pos_enc_layer_type,
+            normalize_before,
+            static_chunk_size,
+            use_dynamic_chunk,
+            global_cmvn,
+            use_dynamic_left_chunk,
+            gradient_checkpointing,
+        )
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+        self.encoders = torch.nn.ModuleList(
+            [
+                TransformerEncoderLayer(
+                    output_size,
+                    COSYVOICE_ATTENTION_CLASSES[selfattention_layer_type](
+                        attention_heads, output_size, attention_dropout_rate, key_bias
+                    ),
+                    PositionwiseFeedForward(
+                        output_size, linear_units, dropout_rate, activation
+                    ),
+                    dropout_rate,
+                    normalize_before,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+
+
+class ConformerEncoder(BaseEncoder):
+    """Conformer encoder module."""
+
+    def __init__(
+        self,
+        input_size: int,
+        output_size: int = 256,
+        attention_heads: int = 4,
+        linear_units: int = 2048,
+        num_blocks: int = 6,
+        dropout_rate: float = 0.1,
+        positional_dropout_rate: float = 0.1,
+        attention_dropout_rate: float = 0.0,
+        input_layer: str = "conv2d",
+        pos_enc_layer_type: str = "rel_pos",
+        normalize_before: bool = True,
+        static_chunk_size: int = 0,
+        use_dynamic_chunk: bool = False,
+        global_cmvn: torch.nn.Module = None,
+        use_dynamic_left_chunk: bool = False,
+        positionwise_conv_kernel_size: int = 1,
+        macaron_style: bool = True,
+        selfattention_layer_type: str = "rel_selfattn",
+        activation_type: str = "swish",
+        use_cnn_module: bool = True,
+        cnn_module_kernel: int = 15,
+        causal: bool = False,
+        cnn_module_norm: str = "batch_norm",
+        key_bias: bool = True,
+        gradient_checkpointing: bool = False,
+    ):
+        """Construct ConformerEncoder
+
+        Args:
+            input_size to use_dynamic_chunk, see in BaseEncoder
+            positionwise_conv_kernel_size (int): Kernel size of positionwise
+                conv1d layer.
+            macaron_style (bool): Whether to use macaron style for
+                positionwise layer.
+            selfattention_layer_type (str): Encoder attention layer type,
+                the parameter has no effect now, it's just for configure
+                compatibility.
+            activation_type (str): Encoder activation function type.
+            use_cnn_module (bool): Whether to use convolution module.
+            cnn_module_kernel (int): Kernel size of convolution module.
+            causal (bool): whether to use causal convolution or not.
+            key_bias: whether use bias in attention.linear_k, False for whisper models.
+        """
+        super().__init__(
+            input_size,
+            output_size,
+            attention_heads,
+            linear_units,
+            num_blocks,
+            dropout_rate,
+            positional_dropout_rate,
+            attention_dropout_rate,
+            input_layer,
+            pos_enc_layer_type,
+            normalize_before,
+            static_chunk_size,
+            use_dynamic_chunk,
+            global_cmvn,
+            use_dynamic_left_chunk,
+            gradient_checkpointing,
+        )
+        activation = COSYVOICE_ACTIVATION_CLASSES[activation_type]()
+
+        # self-attention module definition
+        encoder_selfattn_layer_args = (
+            attention_heads,
+            output_size,
+            attention_dropout_rate,
+            key_bias,
+        )
+        # feed-forward module definition
+        positionwise_layer_args = (
+            output_size,
+            linear_units,
+            dropout_rate,
+            activation,
+        )
+        # convolution module definition
+        convolution_layer_args = (
+            output_size,
+            cnn_module_kernel,
+            activation,
+            cnn_module_norm,
+            causal,
+        )
+
+        self.encoders = torch.nn.ModuleList(
+            [
+                ConformerEncoderLayer(
+                    output_size,
+                    COSYVOICE_ATTENTION_CLASSES[selfattention_layer_type](
+                        *encoder_selfattn_layer_args
+                    ),
+                    PositionwiseFeedForward(*positionwise_layer_args),
+                    (
+                        PositionwiseFeedForward(*positionwise_layer_args)
+                        if macaron_style
+                        else None
+                    ),
+                    (
+                        ConvolutionModule(*convolution_layer_args)
+                        if use_cnn_module
+                        else None
+                    ),
+                    dropout_rate,
+                    normalize_before,
+                )
+                for _ in range(num_blocks)
+            ]
+        )
+        self.inference_buffers = {}
+        self.inference_graphs = {}
+
+    @torch.inference_mode()
+    def capture_inference(self, seq_len_to_capture=[128, 256, 512, 1024]):
+        device = next(self.parameters()).device
+        start_time = time.time()
+        print(
+            f"Start capture_inference for ConformerEncoder, seq_len_to_capture: {seq_len_to_capture}"
+        )
+
+        for seq_len in seq_len_to_capture:
+            xs = torch.randn(
+                1, seq_len, self._output_size, device=device, dtype=torch.bfloat16
+            )
+            xs_lens = torch.tensor([seq_len], device=device, dtype=torch.int32)
+            decoding_chunk_size = 0
+            num_decoding_left_chunks = -1
+
+            T = xs.size(1)
+            masks = ~make_pad_mask(xs_lens, T).unsqueeze(1)  # (B, 1, T)
+            if self.global_cmvn is not None:
+                xs = self.global_cmvn(xs)
+            xs, pos_emb, masks = self.embed(xs, masks)
+            mask_pad = masks  # (B, 1, T/subsample_rate)
+            chunk_masks = add_optional_chunk_mask(
+                xs,
+                masks,
+                self.use_dynamic_chunk,
+                self.use_dynamic_left_chunk,
+                decoding_chunk_size,
+                self.static_chunk_size,
+                num_decoding_left_chunks,
+            )
+
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                out = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad)
+
+            self.inference_graphs[seq_len] = g
+            self.inference_buffers[seq_len] = {
+                "xs": xs,
+                "chunk_masks": chunk_masks,
+                "pos_emb": pos_emb,
+                "mask_pad": mask_pad,
+                "out": out,
+            }
+        end_time = time.time()
+        print(
+            f"Finish capture_inference for ConformerEncoder, time elapsed: {end_time - start_time}"
+        )
+
+    @torch.inference_mode()
+    def inference(self, xs: torch.Tensor, xs_lens: torch.Tensor):
+        curr_seq_len = xs.shape[1]
+        target_len = None
+
+        for seq_len in sorted(self.inference_graphs.keys()):
+            if seq_len >= curr_seq_len:
+                target_len = seq_len
+                break
+
+        if target_len is not None:
+            xs = F.pad(xs, (0, 0, 0, target_len - curr_seq_len), "constant", 0)
+
+        decoding_chunk_size = 0
+        num_decoding_left_chunks = -1
+
+        T = xs.size(1)
+        masks = ~make_pad_mask(xs_lens, T).unsqueeze(1)  # (B, 1, T)
+        if self.global_cmvn is not None:
+            xs = self.global_cmvn(xs)
+        xs, pos_emb, masks = self.embed(xs, masks)
+        mask_pad = masks  # (B, 1, T/subsample_rate)
+        chunk_masks = add_optional_chunk_mask(
+            xs,
+            masks,
+            self.use_dynamic_chunk,
+            self.use_dynamic_left_chunk,
+            decoding_chunk_size,
+            self.static_chunk_size,
+            num_decoding_left_chunks,
+        )
+
+        if target_len is not None:
+            buffer = self.inference_buffers[target_len]
+            buffer["xs"].copy_(xs)
+            buffer["chunk_masks"].copy_(chunk_masks)
+            buffer["pos_emb"].copy_(pos_emb)
+            buffer["mask_pad"].copy_(mask_pad)
+
+            self.inference_graphs[target_len].replay()
+
+            out = buffer["out"][:, :curr_seq_len, :]
+        else:
+            out = self.forward_layers(xs, chunk_masks, pos_emb, mask_pad)
+
+        if self.normalize_before:
+            out = self.after_norm(out)
+        return out, masks

cosyvoice/transformer/encoder_layer.py

@@ -0,0 +1,237 @@
+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2022 Xingchen Song ([email protected])
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Encoder self-attention layer definition."""
+
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+
+
+class TransformerEncoderLayer(nn.Module):
+    """Encoder layer module.
+
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+            instance can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward`, instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: to use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: torch.nn.Module,
+        feed_forward: torch.nn.Module,
+        dropout_rate: float,
+        normalize_before: bool = True,
+    ):
+        """Construct an EncoderLayer object."""
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.norm1 = nn.LayerNorm(size, eps=1e-5)
+        self.norm2 = nn.LayerNorm(size, eps=1e-5)
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute encoded features.
+
+        Args:
+            x (torch.Tensor): (#batch, time, size)
+            mask (torch.Tensor): Mask tensor for the input (#batch, time，time),
+                (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): just for interface compatibility
+                to ConformerEncoderLayer
+            mask_pad (torch.Tensor): does not used in transformer layer,
+                just for unified api with conformer.
+            att_cache (torch.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
+            cnn_cache (torch.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2), not used here, it's for interface
+                compatibility to ConformerEncoderLayer.
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time, time).
+            torch.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+            torch.Tensor: cnn_cahce tensor (#batch=1, size, cache_t2).
+
+        """
+        residual = x
+        if self.normalize_before:
+            x = self.norm1(x)
+        x_att, new_att_cache = self.self_attn(
+            x, x, x, mask, pos_emb=pos_emb, cache=att_cache
+        )
+        x = residual + self.dropout(x_att)
+        if not self.normalize_before:
+            x = self.norm1(x)
+
+        residual = x
+        if self.normalize_before:
+            x = self.norm2(x)
+        x = residual + self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm2(x)
+
+        fake_cnn_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+        return x, mask, new_att_cache, fake_cnn_cache
+
+
+class ConformerEncoderLayer(nn.Module):
+    """Encoder layer module.
+    Args:
+        size (int): Input dimension.
+        self_attn (torch.nn.Module): Self-attention module instance.
+            `MultiHeadedAttention` or `RelPositionMultiHeadedAttention`
+            instance can be used as the argument.
+        feed_forward (torch.nn.Module): Feed-forward module instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        feed_forward_macaron (torch.nn.Module): Additional feed-forward module
+             instance.
+            `PositionwiseFeedForward` instance can be used as the argument.
+        conv_module (torch.nn.Module): Convolution module instance.
+            `ConvlutionModule` instance can be used as the argument.
+        dropout_rate (float): Dropout rate.
+        normalize_before (bool):
+            True: use layer_norm before each sub-block.
+            False: use layer_norm after each sub-block.
+    """
+
+    def __init__(
+        self,
+        size: int,
+        self_attn: torch.nn.Module,
+        feed_forward: Optional[nn.Module] = None,
+        feed_forward_macaron: Optional[nn.Module] = None,
+        conv_module: Optional[nn.Module] = None,
+        dropout_rate: float = 0.1,
+        normalize_before: bool = True,
+    ):
+        """Construct an EncoderLayer object."""
+        super().__init__()
+        self.self_attn = self_attn
+        self.feed_forward = feed_forward
+        self.feed_forward_macaron = feed_forward_macaron
+        self.conv_module = conv_module
+        self.norm_ff = nn.LayerNorm(size, eps=1e-5)  # for the FNN module
+        self.norm_mha = nn.LayerNorm(size, eps=1e-5)  # for the MHA module
+        if feed_forward_macaron is not None:
+            self.norm_ff_macaron = nn.LayerNorm(size, eps=1e-5)
+            self.ff_scale = 0.5
+        else:
+            self.ff_scale = 1.0
+        if self.conv_module is not None:
+            self.norm_conv = nn.LayerNorm(size, eps=1e-5)  # for the CNN module
+            self.norm_final = nn.LayerNorm(
+                size, eps=1e-5
+            )  # for the final output of the block
+        self.dropout = nn.Dropout(dropout_rate)
+        self.size = size
+        self.normalize_before = normalize_before
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: torch.Tensor,
+        pos_emb: torch.Tensor,
+        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
+        att_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+        cnn_cache: torch.Tensor = torch.zeros((0, 0, 0, 0)),
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Compute encoded features.
+
+        Args:
+            x (torch.Tensor): (#batch, time, size)
+            mask (torch.Tensor): Mask tensor for the input (#batch, time，time),
+                (0, 0, 0) means fake mask.
+            pos_emb (torch.Tensor): positional encoding, must not be None
+                for ConformerEncoderLayer.
+            mask_pad (torch.Tensor): batch padding mask used for conv module.
+                (#batch, 1，time), (0, 0, 0) means fake mask.
+            att_cache (torch.Tensor): Cache tensor of the KEY & VALUE
+                (#batch=1, head, cache_t1, d_k * 2), head * d_k == size.
+            cnn_cache (torch.Tensor): Convolution cache in conformer layer
+                (#batch=1, size, cache_t2)
+        Returns:
+            torch.Tensor: Output tensor (#batch, time, size).
+            torch.Tensor: Mask tensor (#batch, time, time).
+            torch.Tensor: att_cache tensor,
+                (#batch=1, head, cache_t1 + time, d_k * 2).
+            torch.Tensor: cnn_cahce tensor (#batch, size, cache_t2).
+        """
+
+        # whether to use macaron style
+        if self.feed_forward_macaron is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_ff_macaron(x)
+            x = residual + self.ff_scale * self.dropout(self.feed_forward_macaron(x))
+            if not self.normalize_before:
+                x = self.norm_ff_macaron(x)
+
+        # multi-headed self-attention module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_mha(x)
+        x_att, new_att_cache = self.self_attn(x, x, x, mask, pos_emb, att_cache)
+        x = residual + self.dropout(x_att)
+        if not self.normalize_before:
+            x = self.norm_mha(x)
+
+        # convolution module
+        # Fake new cnn cache here, and then change it in conv_module
+        new_cnn_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)
+        if self.conv_module is not None:
+            residual = x
+            if self.normalize_before:
+                x = self.norm_conv(x)
+            x, new_cnn_cache = self.conv_module(x, mask_pad, cnn_cache)
+            x = residual + self.dropout(x)
+
+            if not self.normalize_before:
+                x = self.norm_conv(x)
+
+        # feed forward module
+        residual = x
+        if self.normalize_before:
+            x = self.norm_ff(x)
+
+        x = residual + self.ff_scale * self.dropout(self.feed_forward(x))
+        if not self.normalize_before:
+            x = self.norm_ff(x)
+
+        if self.conv_module is not None:
+            x = self.norm_final(x)
+
+        return x, mask, new_att_cache, new_cnn_cache

cosyvoice/transformer/label_smoothing_loss.py

@@ -0,0 +1,98 @@
+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# 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.
+"""Label smoothing module."""
+
+import torch
+from torch import nn
+
+
+class LabelSmoothingLoss(nn.Module):
+    """Label-smoothing loss.
+
+    In a standard CE loss, the label's data distribution is:
+    [0,1,2] ->
+    [
+        [1.0, 0.0, 0.0],
+        [0.0, 1.0, 0.0],
+        [0.0, 0.0, 1.0],
+    ]
+
+    In the smoothing version CE Loss,some probabilities
+    are taken from the true label prob (1.0) and are divided
+    among other labels.
+
+    e.g.
+    smoothing=0.1
+    [0,1,2] ->
+    [
+        [0.9, 0.05, 0.05],
+        [0.05, 0.9, 0.05],
+        [0.05, 0.05, 0.9],
+    ]
+
+    Args:
+        size (int): the number of class
+        padding_idx (int): padding class id which will be ignored for loss
+        smoothing (float): smoothing rate (0.0 means the conventional CE)
+        normalize_length (bool):
+            normalize loss by sequence length if True
+            normalize loss by batch size if False
+    """
+
+    def __init__(
+        self,
+        size: int,
+        padding_idx: int,
+        smoothing: float,
+        normalize_length: bool = False,
+    ):
+        """Construct an LabelSmoothingLoss object."""
+        super(LabelSmoothingLoss, self).__init__()
+        self.criterion = nn.KLDivLoss(reduction="none")
+        self.padding_idx = padding_idx
+        self.confidence = 1.0 - smoothing
+        self.smoothing = smoothing
+        self.size = size
+        self.normalize_length = normalize_length
+
+    def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
+        """Compute loss between x and target.
+
+        The model outputs and data labels tensors are flatten to
+        (batch*seqlen, class) shape and a mask is applied to the
+        padding part which should not be calculated for loss.
+
+        Args:
+            x (torch.Tensor): prediction (batch, seqlen, class)
+            target (torch.Tensor):
+                target signal masked with self.padding_id (batch, seqlen)
+        Returns:
+            loss (torch.Tensor) : The KL loss, scalar float value
+        """
+        assert x.size(2) == self.size
+        batch_size = x.size(0)
+        x = x.view(-1, self.size)
+        target = target.view(-1)
+        # use zeros_like instead of torch.no_grad() for true_dist,
+        # since no_grad() can not be exported by JIT
+        true_dist = torch.zeros_like(x)
+        true_dist.fill_(self.smoothing / (self.size - 1))
+        ignore = target == self.padding_idx  # (B,)
+        total = len(target) - ignore.sum().item()
+        target = target.masked_fill(ignore, 0)  # avoid -1 index
+        true_dist.scatter_(1, target.unsqueeze(1), self.confidence)
+        kl = self.criterion(torch.log_softmax(x, dim=1), true_dist)
+        denom = total if self.normalize_length else batch_size
+        return kl.masked_fill(ignore.unsqueeze(1), 0).sum() / denom

cosyvoice/transformer/positionwise_feed_forward.py

@@ -0,0 +1,116 @@
+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#
+# 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.
+"""Positionwise feed forward layer definition."""
+
+import torch
+
+
+class PositionwiseFeedForward(torch.nn.Module):
+    """Positionwise feed forward layer.
+
+    FeedForward are appied on each position of the sequence.
+    The output dim is same with the input dim.
+
+    Args:
+        idim (int): Input dimenstion.
+        hidden_units (int): The number of hidden units.
+        dropout_rate (float): Dropout rate.
+        activation (torch.nn.Module): Activation function
+    """
+
+    def __init__(
+        self,
+        idim: int,
+        hidden_units: int,
+        dropout_rate: float,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+    ):
+        """Construct a PositionwiseFeedForward object."""
+        super(PositionwiseFeedForward, self).__init__()
+        self.w_1 = torch.nn.Linear(idim, hidden_units)
+        self.activation = activation
+        self.dropout = torch.nn.Dropout(dropout_rate)
+        self.w_2 = torch.nn.Linear(hidden_units, idim)
+
+    def forward(self, xs: torch.Tensor) -> torch.Tensor:
+        """Forward function.
+
+        Args:
+            xs: input tensor (B, L, D)
+        Returns:
+            output tensor, (B, L, D)
+        """
+        return self.w_2(self.dropout(self.activation(self.w_1(xs))))
+
+
+class MoEFFNLayer(torch.nn.Module):
+    """
+    Mixture of expert with Positionwise feed forward layer
+    See also figure 1 in https://arxiv.org/pdf/2305.15663.pdf
+    The output dim is same with the input dim.
+
+    Modified from https://github.com/Lightning-AI/lit-gpt/pull/823
+                  https://github.com/mistralai/mistral-src/blob/b46d6/moe_one_file_ref.py#L203-L219
+    Args:
+        n_expert: number of expert.
+        n_expert_per_token: The actual number of experts used for each frame
+        idim (int): Input dimenstion.
+        hidden_units (int): The number of hidden units.
+        dropout_rate (float): Dropout rate.
+        activation (torch.nn.Module): Activation function
+    """
+
+    def __init__(
+        self,
+        n_expert: int,
+        n_expert_per_token: int,
+        idim: int,
+        hidden_units: int,
+        dropout_rate: float,
+        activation: torch.nn.Module = torch.nn.ReLU(),
+    ):
+        super(MoEFFNLayer, self).__init__()
+        self.gate = torch.nn.Linear(idim, n_expert, bias=False)
+        self.experts = torch.nn.ModuleList(
+            PositionwiseFeedForward(idim, hidden_units, dropout_rate, activation)
+            for _ in range(n_expert)
+        )
+        self.n_expert_per_token = n_expert_per_token
+
+    def forward(self, xs: torch.Tensor) -> torch.Tensor:
+        """Foward function.
+        Args:
+            xs: input tensor (B, L, D)
+        Returns:
+            output tensor, (B, L, D)
+
+        """
+        B, L, D = xs.size()  # batch size, sequence length, embedding dimension (idim)
+        xs = xs.view(-1, D)  # (B*L, D)
+        router = self.gate(xs)  # (B*L, n_expert)
+        logits, indices = torch.topk(
+            router, self.n_expert_per_token
+        )  # probs:(B*L, n_expert), indices: (B*L, n_expert)
+        weights = torch.nn.functional.softmax(logits, dim=1, dtype=torch.float).to(
+            dtype=xs.dtype
+        )  # (B*L, n_expert_per_token)
+        output = torch.zeros_like(xs)  # (B*L, D)
+        for i, expert in enumerate(self.experts):
+            mask = indices == i
+            batch_idx, ith_expert = torch.where(mask)
+            output[batch_idx] += weights[batch_idx, ith_expert, None] * expert(
+                xs[batch_idx]
+            )
+        return output.view(B, L, D)

cosyvoice/transformer/subsampling.py

@@ -0,0 +1,391 @@
+# Copyright (c) 2021 Mobvoi Inc (Binbin Zhang, Di Wu)
+#               2024 Alibaba Inc (Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Subsampling layer definition."""
+
+from typing import Tuple, Union
+
+import torch
+
+
+class BaseSubsampling(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def position_encoding(
+        self, offset: Union[int, torch.Tensor], size: int
+    ) -> torch.Tensor:
+        return self.pos_enc.position_encoding(offset, size)
+
+
+class EmbedinigNoSubsampling(BaseSubsampling):
+    """Embedding input without subsampling"""
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        super().__init__()
+        self.embed = torch.nn.Embedding(idim, odim)
+        self.pos_enc = pos_enc_class
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.embed(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask
+
+
+class LinearNoSubsampling(BaseSubsampling):
+    """Linear transform the input without subsampling
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an linear object."""
+        super().__init__()
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(idim, odim),
+            torch.nn.LayerNorm(odim, eps=1e-5),
+            torch.nn.Dropout(dropout_rate),
+        )
+        self.pos_enc = pos_enc_class
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.out(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask
+
+
+class Conv1dSubsampling2(BaseSubsampling):
+    """Convolutional 1D subsampling (to 1/2 length).
+       It is designed for Whisper, ref:
+       https://github.com/openai/whisper/blob/main/whisper/model.py
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an Conv1dSubsampling2 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv1d(idim, odim, kernel_size=3, padding=1),
+            torch.nn.GELU(),
+            torch.nn.Conv1d(odim, odim, kernel_size=3, stride=2, padding=1),
+            torch.nn.GELU(),
+        )
+        self.pos_enc = pos_enc_class
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) * frame_rate_of_this_layer
+        self.subsampling_rate = 2
+        # 4 = (3 - 1) * 1 + (3 - 1) * 1
+        self.right_context = 4
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 2.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 2.
+            torch.Tensor: positional encoding
+
+        """
+        time = x.size(1)
+        x = x.transpose(1, 2)  # (b, f, t)
+        x = self.conv(x)
+        x = x.transpose(1, 2)  # (b, t, f)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, (time + 1) % 2 :: 2]
+
+
+class Conv2dSubsampling4(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/4 length).
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an Conv2dSubsampling4 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+        )
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
+        )
+        self.pos_enc = pos_enc_class
+        # The right context for every conv layer is computed by:
+        # (kernel_size - 1) * frame_rate_of_this_layer
+        self.subsampling_rate = 4
+        # 6 = (3 - 1) * 1 + (3 - 1) * 2
+        self.right_context = 6
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 4.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 4.
+            torch.Tensor: positional encoding
+
+        """
+        x = x.unsqueeze(1)  # (b, c=1, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2]
+
+
+class Conv2dSubsampling6(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/6 length).
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+        pos_enc (torch.nn.Module): Custom position encoding layer.
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an Conv2dSubsampling6 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 5, 3),
+            torch.nn.ReLU(),
+        )
+        self.linear = torch.nn.Linear(odim * (((idim - 1) // 2 - 2) // 3), odim)
+        self.pos_enc = pos_enc_class
+        # 10 = (3 - 1) * 1 + (5 - 1) * 2
+        self.subsampling_rate = 6
+        self.right_context = 10
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 6.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 6.
+            torch.Tensor: positional encoding
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 4::3]
+
+
+class Conv2dSubsampling8(BaseSubsampling):
+    """Convolutional 2D subsampling (to 1/8 length).
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an Conv2dSubsampling8 object."""
+        super().__init__()
+        self.conv = torch.nn.Sequential(
+            torch.nn.Conv2d(1, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+            torch.nn.Conv2d(odim, odim, 3, 2),
+            torch.nn.ReLU(),
+        )
+        self.linear = torch.nn.Linear(
+            odim * ((((idim - 1) // 2 - 1) // 2 - 1) // 2), odim
+        )
+        self.pos_enc = pos_enc_class
+        self.subsampling_rate = 8
+        # 14 = (3 - 1) * 1 + (3 - 1) * 2 + (3 - 1) * 4
+        self.right_context = 14
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Subsample x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: Subsampled tensor (#batch, time', odim),
+                where time' = time // 8.
+            torch.Tensor: Subsampled mask (#batch, 1, time'),
+                where time' = time // 8.
+            torch.Tensor: positional encoding
+        """
+        x = x.unsqueeze(1)  # (b, c, t, f)
+        x = self.conv(x)
+        b, c, t, f = x.size()
+        x = self.linear(x.transpose(1, 2).contiguous().view(b, t, c * f))
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask[:, :, 2::2][:, :, 2::2][:, :, 2::2]
+
+
+class LegacyLinearNoSubsampling(BaseSubsampling):
+    """Linear transform the input without subsampling
+
+    Args:
+        idim (int): Input dimension.
+        odim (int): Output dimension.
+        dropout_rate (float): Dropout rate.
+
+    """
+
+    def __init__(
+        self, idim: int, odim: int, dropout_rate: float, pos_enc_class: torch.nn.Module
+    ):
+        """Construct an linear object."""
+        super().__init__()
+        self.out = torch.nn.Sequential(
+            torch.nn.Linear(idim, odim),
+            torch.nn.LayerNorm(odim, eps=1e-5),
+            torch.nn.Dropout(dropout_rate),
+            torch.nn.ReLU(),
+        )
+        self.pos_enc = pos_enc_class
+        self.right_context = 0
+        self.subsampling_rate = 1
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        offset: Union[int, torch.Tensor] = 0,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Input x.
+
+        Args:
+            x (torch.Tensor): Input tensor (#batch, time, idim).
+            x_mask (torch.Tensor): Input mask (#batch, 1, time).
+
+        Returns:
+            torch.Tensor: linear input tensor (#batch, time', odim),
+                where time' = time .
+            torch.Tensor: linear input mask (#batch, 1, time'),
+                where time' = time .
+
+        """
+        x = self.out(x)
+        x, pos_emb = self.pos_enc(x, offset)
+        return x, pos_emb, x_mask

cosyvoice/utils/audio.py

@@ -0,0 +1,90 @@
+import numpy as np
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+from scipy.io.wavfile import read
+
+MAX_WAV_VALUE = 32768.0
+
+
+def load_wav(full_path):
+    sampling_rate, data = read(full_path)
+    return data, sampling_rate
+
+
+def dynamic_range_compression(x, C=1, clip_val=1e-5):
+    return np.log(np.clip(x, a_min=clip_val, a_max=None) * C)
+
+
+def dynamic_range_decompression(x, C=1):
+    return np.exp(x) / C
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def mel_spectrogram(
+    y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
+):
+    # if torch.min(y) < -1.0:
+    #     print("min value is ", torch.min(y))
+    # if torch.max(y) > 1.0:
+    #     print("max value is ", torch.max(y))
+
+    global mel_basis, hann_window  # pylint: disable=global-statement
+    if f"{str(fmax)}_{str(y.device)}" not in mel_basis:
+        mel = librosa_mel_fn(
+            sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+        )
+        mel_basis[str(fmax) + "_" + str(y.device)] = (
+            torch.from_numpy(mel).float().to(y.device)
+        )
+        hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
+
+    y = torch.nn.functional.pad(
+        y.unsqueeze(1),
+        (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
+        mode="reflect",
+    )
+    y = y.squeeze(1)
+
+    spec = torch.view_as_real(
+        torch.stft(
+            y,
+            n_fft,
+            hop_length=hop_size,
+            win_length=win_size,
+            window=hann_window[str(y.device)],
+            center=center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+    )
+
+    spec = torch.sqrt(spec.pow(2).sum(-1) + (1e-9))
+
+    spec = torch.matmul(mel_basis[str(fmax) + "_" + str(y.device)], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec

cosyvoice/utils/class_utils.py

@@ -0,0 +1,78 @@
+# Copyright [2023-11-28] <[email protected], Xingchen Song>
+#            2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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
+
+from cosyvoice.transformer.activation import Swish
+from cosyvoice.transformer.subsampling import (
+    LinearNoSubsampling,
+    EmbedinigNoSubsampling,
+    Conv1dSubsampling2,
+    Conv2dSubsampling4,
+    Conv2dSubsampling6,
+    Conv2dSubsampling8,
+)
+from cosyvoice.transformer.embedding import (
+    PositionalEncoding,
+    RelPositionalEncoding,
+    WhisperPositionalEncoding,
+    LearnablePositionalEncoding,
+    NoPositionalEncoding,
+)
+from cosyvoice.transformer.attention import (
+    MultiHeadedAttention,
+    RelPositionMultiHeadedAttention,
+)
+from cosyvoice.transformer.embedding import (
+    EspnetRelPositionalEncoding,
+)
+from cosyvoice.transformer.subsampling import (
+    LegacyLinearNoSubsampling,
+)
+
+
+COSYVOICE_ACTIVATION_CLASSES = {
+    "hardtanh": torch.nn.Hardtanh,
+    "tanh": torch.nn.Tanh,
+    "relu": torch.nn.ReLU,
+    "selu": torch.nn.SELU,
+    "swish": getattr(torch.nn, "SiLU", Swish),
+    "gelu": torch.nn.GELU,
+}
+
+COSYVOICE_SUBSAMPLE_CLASSES = {
+    "linear": LinearNoSubsampling,
+    "linear_legacy": LegacyLinearNoSubsampling,
+    "embed": EmbedinigNoSubsampling,
+    "conv1d2": Conv1dSubsampling2,
+    "conv2d": Conv2dSubsampling4,
+    "conv2d6": Conv2dSubsampling6,
+    "conv2d8": Conv2dSubsampling8,
+    "paraformer_dummy": torch.nn.Identity,
+}
+
+COSYVOICE_EMB_CLASSES = {
+    "embed": PositionalEncoding,
+    "abs_pos": PositionalEncoding,
+    "rel_pos": RelPositionalEncoding,
+    "rel_pos_espnet": EspnetRelPositionalEncoding,
+    "no_pos": NoPositionalEncoding,
+    "abs_pos_whisper": WhisperPositionalEncoding,
+    "embed_learnable_pe": LearnablePositionalEncoding,
+}
+
+COSYVOICE_ATTENTION_CLASSES = {
+    "selfattn": MultiHeadedAttention,
+    "rel_selfattn": RelPositionMultiHeadedAttention,
+}

cosyvoice/utils/common.py

@@ -0,0 +1,169 @@
+# Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+"""Unility functions for Transformer."""
+
+import random
+from typing import List
+
+import numpy as np
+import torch
+
+IGNORE_ID = -1
+
+
+def pad_list(xs: List[torch.Tensor], pad_value: int):
+    """Perform padding for the list of tensors.
+
+    Args:
+        xs (List): List of Tensors [(T_1, `*`), (T_2, `*`), ..., (T_B, `*`)].
+        pad_value (float): Value for padding.
+
+    Returns:
+        Tensor: Padded tensor (B, Tmax, `*`).
+
+    Examples:
+        >>> x = [torch.ones(4), torch.ones(2), torch.ones(1)]
+        >>> x
+        [tensor([1., 1., 1., 1.]), tensor([1., 1.]), tensor([1.])]
+        >>> pad_list(x, 0)
+        tensor([[1., 1., 1., 1.],
+                [1., 1., 0., 0.],
+                [1., 0., 0., 0.]])
+
+    """
+    max_len = max([len(item) for item in xs])
+    batchs = len(xs)
+    ndim = xs[0].ndim
+    if ndim == 1:
+        pad_res = torch.zeros(batchs, max_len, dtype=xs[0].dtype, device=xs[0].device)
+    elif ndim == 2:
+        pad_res = torch.zeros(
+            batchs, max_len, xs[0].shape[1], dtype=xs[0].dtype, device=xs[0].device
+        )
+    elif ndim == 3:
+        pad_res = torch.zeros(
+            batchs,
+            max_len,
+            xs[0].shape[1],
+            xs[0].shape[2],
+            dtype=xs[0].dtype,
+            device=xs[0].device,
+        )
+    else:
+        raise ValueError(f"Unsupported ndim: {ndim}")
+    pad_res.fill_(pad_value)
+    for i in range(batchs):
+        pad_res[i, : len(xs[i])] = xs[i]
+    return pad_res
+
+
+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 get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+# Repetition Aware Sampling in VALL-E 2
+def ras_sampling(
+    weighted_scores,
+    decoded_tokens,
+    sampling,
+    top_p=0.8,
+    top_k=25,
+    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, sampling)
+    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, sampling):
+    top_ids = weighted_scores.softmax(dim=0).multinomial(1, replacement=True)
+    return top_ids
+
+
+def fade_in_out(fade_in_mel, fade_out_mel, window):
+    device = fade_in_mel.device
+    fade_in_mel, fade_out_mel = fade_in_mel.cpu(), fade_out_mel.cpu()
+    mel_overlap_len = int(window.shape[0] / 2)
+    fade_in_mel[..., :mel_overlap_len] = (
+        fade_in_mel[..., :mel_overlap_len] * window[:mel_overlap_len]
+        + fade_out_mel[..., -mel_overlap_len:] * window[mel_overlap_len:]
+    )
+    return fade_in_mel.to(device)
+
+
+def set_all_random_seed(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)

cosyvoice/utils/executor.py

@@ -0,0 +1,151 @@
+# Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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 logging
+from contextlib import nullcontext
+import os
+
+import torch
+import torch.distributed as dist
+
+from cosyvoice.utils.train_utils import (
+    update_parameter_and_lr,
+    log_per_step,
+    log_per_save,
+    batch_forward,
+    batch_backward,
+    save_model,
+    cosyvoice_join,
+)
+
+
+class Executor:
+
+    def __init__(self):
+        self.step = 0
+        self.epoch = 0
+        self.rank = int(os.environ.get("RANK", 0))
+        self.device = torch.device("cuda:{}".format(self.rank))
+
+    def train_one_epoc(
+        self,
+        model,
+        optimizer,
+        scheduler,
+        train_data_loader,
+        cv_data_loader,
+        writer,
+        info_dict,
+        group_join,
+    ):
+        """Train one epoch"""
+
+        lr = optimizer.param_groups[0]["lr"]
+        logging.info(
+            "Epoch {} TRAIN info lr {} rank {}".format(self.epoch, lr, self.rank)
+        )
+        logging.info(
+            "using accumulate grad, new batch size is {} times"
+            " larger than before".format(info_dict["accum_grad"])
+        )
+        # A context manager to be used in conjunction with an instance of
+        # torch.nn.parallel.DistributedDataParallel to be able to train
+        # with uneven inputs across participating processes.
+        model.train()
+        model_context = (
+            model.join if info_dict["train_engine"] == "torch_ddp" else nullcontext
+        )
+        with model_context():
+            for batch_idx, batch_dict in enumerate(train_data_loader):
+                info_dict["tag"] = "TRAIN"
+                info_dict["step"] = self.step
+                info_dict["epoch"] = self.epoch
+                info_dict["batch_idx"] = batch_idx
+                if cosyvoice_join(group_join, info_dict):
+                    break
+
+                # Disable gradient synchronizations across DDP processes.
+                # Within this context, gradients will be accumulated on module
+                # variables, which will later be synchronized.
+                if (
+                    info_dict["train_engine"] == "torch_ddp"
+                    and (batch_idx + 1) % info_dict["accum_grad"] != 0
+                ):
+                    context = model.no_sync
+                # Used for single gpu training and DDP gradient synchronization
+                # processes.
+                else:
+                    context = nullcontext
+
+                with context():
+                    info_dict = batch_forward(model, batch_dict, info_dict)
+                    info_dict = batch_backward(model, info_dict)
+
+                info_dict = update_parameter_and_lr(
+                    model, optimizer, scheduler, info_dict
+                )
+                log_per_step(writer, info_dict)
+                # NOTE specify save_per_step in cosyvoice.yaml if you want to enable step save
+                if (
+                    info_dict["save_per_step"] > 0
+                    and (self.step + 1) % info_dict["save_per_step"] == 0
+                    and (batch_idx + 1) % info_dict["accum_grad"] == 0
+                ):
+                    dist.barrier()
+                    self.cv(
+                        model, cv_data_loader, writer, info_dict, on_batch_end=False
+                    )
+                    model.train()
+                if (batch_idx + 1) % info_dict["accum_grad"] == 0:
+                    self.step += 1
+        dist.barrier()
+        self.cv(model, cv_data_loader, writer, info_dict, on_batch_end=True)
+
+    @torch.inference_mode()
+    def cv(self, model, cv_data_loader, writer, info_dict, on_batch_end=True):
+        """Cross validation on"""
+        logging.info(
+            "Epoch {} Step {} on_batch_end {} CV rank {}".format(
+                self.epoch, self.step + 1, on_batch_end, self.rank
+            )
+        )
+        model.eval()
+        total_num_utts, total_loss_dict = 0, {}  # avoid division by 0
+        for batch_idx, batch_dict in enumerate(cv_data_loader):
+            info_dict["tag"] = "CV"
+            info_dict["step"] = self.step
+            info_dict["epoch"] = self.epoch
+            info_dict["batch_idx"] = batch_idx
+
+            num_utts = len(batch_dict["utts"])
+            total_num_utts += num_utts
+
+            info_dict = batch_forward(model, batch_dict, info_dict)
+
+            for k, v in info_dict["loss_dict"].items():
+                if k not in total_loss_dict:
+                    total_loss_dict[k] = []
+                total_loss_dict[k].append(v.item() * num_utts)
+            log_per_step(None, info_dict)
+        for k, v in total_loss_dict.items():
+            total_loss_dict[k] = sum(v) / total_num_utts
+        info_dict["loss_dict"] = total_loss_dict
+        log_per_save(writer, info_dict)
+        model_name = (
+            "epoch_{}_whole".format(self.epoch)
+            if on_batch_end
+            else "epoch_{}_step_{}".format(self.epoch, self.step + 1)
+        )
+        save_model(model, model_name, info_dict)

cosyvoice/utils/file_utils.py

@@ -0,0 +1,49 @@
+# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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 json
+import torchaudio
+import logging
+
+logging.getLogger("matplotlib").setLevel(logging.WARNING)
+logging.basicConfig(level=logging.DEBUG, format="%(asctime)s %(levelname)s %(message)s")
+
+
+def read_lists(list_file):
+    lists = []
+    with open(list_file, "r", encoding="utf8") as fin:
+        for line in fin:
+            lists.append(line.strip())
+    return lists
+
+
+def read_json_lists(list_file):
+    lists = read_lists(list_file)
+    results = {}
+    for fn in lists:
+        with open(fn, "r", encoding="utf8") as fin:
+            results.update(json.load(fin))
+    return results
+
+
+def load_wav(wav, target_sr):
+    speech, sample_rate = torchaudio.load(wav)
+    speech = speech.mean(dim=0, keepdim=True)
+    if sample_rate != target_sr:
+        # assert sample_rate > target_sr, 'wav sample rate {} must be greater than {}'.format(sample_rate, target_sr)
+        speech = torchaudio.transforms.Resample(
+            orig_freq=sample_rate, new_freq=target_sr
+        )(speech)
+    return speech

cosyvoice/utils/frontend_utils.py

@@ -0,0 +1,142 @@
+# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Zhihao Du)
+#
+# 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 re
+
+chinese_char_pattern = re.compile(r"[\u4e00-\u9fff]+")
+
+
+# whether contain chinese character
+def contains_chinese(text):
+    return bool(chinese_char_pattern.search(text))
+
+
+# replace special symbol
+def replace_corner_mark(text):
+    text = text.replace("²", "平方")
+    text = text.replace("³", "立方")
+    return text
+
+
+# remove meaningless symbol
+def remove_bracket(text):
+    text = text.replace("（", "").replace("）", "")
+    text = text.replace("【", "").replace("】", "")
+    text = text.replace("`", "").replace("`", "")
+    text = text.replace("——", " ")
+    return text
+
+
+# spell Arabic numerals
+def spell_out_number(text: str, inflect_parser):
+    new_text = []
+    st = None
+    for i, c in enumerate(text):
+        if not c.isdigit():
+            if st is not None:
+                num_str = inflect_parser.number_to_words(text[st:i])
+                new_text.append(num_str)
+                st = None
+            new_text.append(c)
+        else:
+            if st is None:
+                st = i
+    if st is not None and st < len(text):
+        num_str = inflect_parser.number_to_words(text[st:])
+        new_text.append(num_str)
+    return "".join(new_text)
+
+
+# split paragrah logic：
+# 1. per sentence max len token_max_n, min len token_min_n, merge if last sentence len less than merge_len
+# 2. cal sentence len according to lang
+# 3. split sentence according to puncatation
+def split_paragraph(
+    text: str,
+    tokenize,
+    lang="zh",
+    token_max_n=80,
+    token_min_n=60,
+    merge_len=20,
+    comma_split=False,
+):
+    def calc_utt_length(_text: str):
+        if lang == "zh":
+            return len(_text)
+        else:
+            return len(tokenize(_text))
+
+    def should_merge(_text: str):
+        if lang == "zh":
+            return len(_text) < merge_len
+        else:
+            return len(tokenize(_text)) < merge_len
+
+    if lang == "zh":
+        pounc = ["。", "？", "！", "；", "：", "、", ".", "?", "!", ";"]
+    else:
+        pounc = [".", "?", "!", ";", ":"]
+    if comma_split:
+        pounc.extend(["，", ","])
+
+    if text[-1] not in pounc:
+        if lang == "zh":
+            text += "。"
+        else:
+            text += "."
+
+    st = 0
+    utts = []
+    for i, c in enumerate(text):
+        if c in pounc:
+            if len(text[st:i]) > 0:
+                utts.append(text[st:i] + c)
+            if i + 1 < len(text) and text[i + 1] in ['"', "”"]:
+                tmp = utts.pop(-1)
+                utts.append(tmp + text[i + 1])
+                st = i + 2
+            else:
+                st = i + 1
+
+    final_utts = []
+    cur_utt = ""
+    for utt in utts:
+        if (
+            calc_utt_length(cur_utt + utt) > token_max_n
+            and calc_utt_length(cur_utt) > token_min_n
+        ):
+            final_utts.append(cur_utt)
+            cur_utt = ""
+        cur_utt = cur_utt + utt
+    if len(cur_utt) > 0:
+        if should_merge(cur_utt) and len(final_utts) != 0:
+            final_utts[-1] = final_utts[-1] + cur_utt
+        else:
+            final_utts.append(cur_utt)
+
+    return final_utts
+
+
+# remove blank between chinese character
+def replace_blank(text: str):
+    out_str = []
+    for i, c in enumerate(text):
+        if c == " ":
+            if (text[i + 1].isascii() and text[i + 1] != " ") and (
+                text[i - 1].isascii() and text[i - 1] != " "
+            ):
+                out_str.append(c)
+        else:
+            out_str.append(c)
+    return "".join(out_str)

cosyvoice/utils/mask.py

@@ -0,0 +1,226 @@
+# Copyright (c) 2019 Shigeki Karita
+#               2020 Mobvoi Inc (Binbin Zhang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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
+
+'''
+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]]
+    """
+    ret = torch.ones(size, size, device=device, dtype=torch.bool)
+    return torch.tril(ret)
+'''
+
+
+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

cosyvoice/utils/scheduler.py

@@ -0,0 +1,761 @@
+# Copyright (c) 2020 Mobvoi Inc (Binbin Zhang)
+#               2022 Ximalaya Inc (Yuguang Yang)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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.
+# Modified from ESPnet(https://github.com/espnet/espnet)
+#               NeMo(https://github.com/NVIDIA/NeMo)
+
+from typing import Union
+
+import math
+import warnings
+import torch
+from torch.optim.lr_scheduler import _LRScheduler
+
+
+class WarmupLR(_LRScheduler):
+    """The WarmupLR scheduler
+
+    This scheduler is almost same as NoamLR Scheduler except for following
+    difference:
+
+    NoamLR:
+        lr = optimizer.lr * model_size ** -0.5
+             * min(step ** -0.5, step * warmup_step ** -1.5)
+    WarmupLR:
+        lr = optimizer.lr * warmup_step ** 0.5
+             * min(step ** -0.5, step * warmup_step ** -1.5)
+
+    Note that the maximum lr equals to optimizer.lr in this scheduler.
+
+    """
+
+    def __init__(
+        self,
+        optimizer: torch.optim.Optimizer,
+        warmup_steps: Union[int, float] = 25000,
+        last_epoch: int = -1,
+    ):
+        self.warmup_steps = warmup_steps
+
+        # __init__() must be invoked before setting field
+        # because step() is also invoked in __init__()
+        super().__init__(optimizer, last_epoch)
+
+    def __repr__(self):
+        return f"{self.__class__.__name__}(warmup_steps={self.warmup_steps})"
+
+    def get_lr(self):
+        step_num = self.last_epoch + 1
+        if self.warmup_steps == 0:
+            return [lr * step_num**-0.5 for lr in self.base_lrs]
+        else:
+            return [
+                lr
+                * self.warmup_steps**0.5
+                * min(step_num**-0.5, step_num * self.warmup_steps**-1.5)
+                for lr in self.base_lrs
+            ]
+
+    def set_step(self, step: int):
+        self.last_epoch = step
+
+
+class WarmupPolicy(_LRScheduler):
+    """Adds warmup kwargs and warmup logic to lr policy.
+    All arguments should be passed as kwargs for clarity,
+    Args:
+        warmup_steps: Number of training steps in warmup stage
+        warmup_ratio: Ratio of warmup steps to total steps
+        max_steps: Total number of steps while training or `None` for
+            infinite training
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        warmup_steps=None,
+        warmup_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
+    ):
+        assert not (
+            warmup_steps is not None and warmup_ratio is not None
+        ), "Either use particular number of step or ratio"
+        assert (
+            warmup_ratio is None or max_steps is not None
+        ), "If there is a ratio, there should be a total steps"
+
+        # It is necessary to assign all attributes *before* __init__,
+        # as class is wrapped by an inner class.
+        self.max_steps = max_steps
+        if warmup_steps is not None:
+            self.warmup_steps = warmup_steps
+        elif warmup_ratio is not None:
+            self.warmup_steps = int(warmup_ratio * max_steps)
+        else:
+            self.warmup_steps = 0
+
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn(
+                "To get the last learning rate computed "
+                "by the scheduler, please use `get_last_lr()`.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        step = self.last_epoch
+
+        if step <= self.warmup_steps and self.warmup_steps > 0:
+            return self._get_warmup_lr(step)
+
+        if step > self.max_steps:
+            return [self.min_lr for _ in self.base_lrs]
+
+        return self._get_lr(step)
+
+    def _get_warmup_lr(self, step):
+        lr_val = (step + 1) / (self.warmup_steps + 1)
+        return [initial_lr * lr_val for initial_lr in self.base_lrs]
+
+    def _get_lr(self, step):
+        """Simple const lr policy"""
+        return self.base_lrs
+
+
+class SquareRootConstantPolicy(_LRScheduler):
+    """Adds warmup kwargs and warmup logic to lr policy.
+    All arguments should be passed as kwargs for clarity,
+    Args:
+        warmup_steps: Number of training steps in warmup stage
+        warmup_ratio: Ratio of warmup steps to total steps
+        max_steps: Total number of steps while training or `None` for
+            infinite training
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        constant_steps=None,
+        constant_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
+    ):
+        assert not (
+            constant_steps is not None and constant_ratio is not None
+        ), "Either use particular number of step or ratio"
+        assert (
+            constant_ratio is None or max_steps is not None
+        ), "If there is a ratio, there should be a total steps"
+
+        # It is necessary to assign all attributes *before* __init__,
+        # as class is wrapped by an inner class.
+        self.max_steps = max_steps
+        if constant_steps is not None:
+            self.constant_steps = constant_steps
+        elif constant_ratio is not None:
+            self.constant_steps = int(constant_ratio * max_steps)
+        else:
+            self.constant_steps = 0
+
+        self.constant_lr = 1 / (constant_steps**0.5)
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn(
+                "To get the last learning rate computed "
+                "by the scheduler, please use `get_last_lr()`.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        step = self.last_epoch
+
+        if step <= self.constant_steps:
+            return [self.constant_lr for _ in self.base_lrs]
+
+        if step > self.max_steps:
+            return [self.min_lr for _ in self.base_lrs]
+
+        return self._get_lr(step)
+
+    def _get_lr(self, step):
+        """Simple const lr policy"""
+        return self.base_lrs
+
+
+class WarmupHoldPolicy(WarmupPolicy):
+    """Variant of WarmupPolicy which maintains high
+       learning rate for a defined number of steps.
+    All arguments should be passed as kwargs for clarity,
+    Args:
+        warmup_steps: Number of training steps in warmup stage
+        warmup_ratio: Ratio of warmup steps to total steps
+        hold_steps: Number of training steps to
+                    hold the learning rate after warm up
+        hold_ratio: Ratio of hold steps to total steps
+        max_steps: Total number of steps while training or `None` for
+            infinite training
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        warmup_steps=None,
+        warmup_ratio=None,
+        hold_steps=None,
+        hold_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
+    ):
+        assert not (
+            hold_steps is not None and hold_ratio is not None
+        ), "Either use particular number of step or ratio"
+        assert (
+            hold_ratio is None or max_steps is not None
+        ), "If there is a ratio, there should be a total steps"
+
+        self.min_lr = min_lr
+        self._last_warmup_lr = 0.0
+
+        # Necessary to duplicate as class attributes are hidden in inner class
+        self.max_steps = max_steps
+        if warmup_steps is not None:
+            self.warmup_steps = warmup_steps
+        elif warmup_ratio is not None:
+            self.warmup_steps = int(warmup_ratio * max_steps)
+        else:
+            self.warmup_steps = 0
+
+        if hold_steps is not None:
+            self.hold_steps = hold_steps + self.warmup_steps
+        elif hold_ratio is not None:
+            self.hold_steps = int(hold_ratio * max_steps) + self.warmup_steps
+        else:
+            self.hold_steps = 0
+
+        super().__init__(
+            optimizer,
+            warmup_steps=warmup_steps,
+            warmup_ratio=warmup_ratio,
+            max_steps=max_steps,
+            last_epoch=last_epoch,
+            min_lr=min_lr,
+        )
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn(
+                "To get the last learning rate computed by the scheduler,"
+                " "
+                "please use `get_last_lr()`.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        step = self.last_epoch
+
+        # Warmup phase
+        if step <= self.warmup_steps and self.warmup_steps > 0:
+            return self._get_warmup_lr(step)
+
+        # Hold phase
+        if (step >= self.warmup_steps) and (step < self.hold_steps):
+            return self.base_lrs
+
+        if step > self.max_steps:
+            return [self.min_lr for _ in self.base_lrs]
+
+        return self._get_lr(step)
+
+
+class WarmupAnnealHoldPolicy(_LRScheduler):
+    """Adds warmup kwargs and warmup logic to lr policy.
+    All arguments should be passed as kwargs for clarity,
+    Args:
+        warmup_steps: Number of training steps in warmup stage
+        warmup_ratio: Ratio of warmup steps to total steps
+        max_steps: Total number of steps while training or `None` for
+            infinite training
+        min_lr: Minimum lr to hold the learning rate after decay at.
+        constant_steps: Number of steps to keep lr constant at.
+        constant_ratio: Ratio of steps to keep lr constant.
+    """
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        warmup_steps=None,
+        warmup_ratio=None,
+        constant_steps=None,
+        constant_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
+    ):
+        assert not (
+            warmup_steps is not None and warmup_ratio is not None
+        ), "Either use particular number of step or ratio"
+        assert not (
+            constant_steps is not None and constant_ratio is not None
+        ), "Either use constant_steps or constant_ratio"
+        assert (
+            warmup_ratio is None or max_steps is not None
+        ), "If there is a ratio, there should be a total steps"
+
+        # It is necessary to assign all attributes *before* __init__,
+        # as class is wrapped by an inner class.
+        self.max_steps = max_steps
+
+        if warmup_steps is not None:
+            self.warmup_steps = warmup_steps
+        elif warmup_ratio is not None:
+            self.warmup_steps = int(warmup_ratio * max_steps)
+        else:
+            self.warmup_steps = 0
+
+        if constant_steps is not None:
+            self.constant_steps = constant_steps
+        elif constant_ratio is not None:
+            self.constant_steps = int(constant_ratio * max_steps)
+        else:
+            self.constant_steps = 0
+
+        self.decay_steps = max_steps - (self.constant_steps + self.warmup_steps)
+
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn(
+                "To get the last learning rate computed "
+                "by the scheduler, please use `get_last_lr()`.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        step = self.last_epoch
+
+        # Warmup steps
+        if self.warmup_steps > 0 and step <= self.warmup_steps:
+            return self._get_warmup_lr(step)
+
+        # Constant steps after warmup and decay
+        if (
+            self.constant_steps > 0
+            and (self.warmup_steps + self.decay_steps) < step <= self.max_steps
+        ):
+            return self._get_constant_lr(step)
+
+        # Min lr after max steps of updates
+        if step > self.max_steps:
+            return [self.min_lr for _ in self.base_lrs]
+
+        return self._get_lr(step)
+
+    def _get_warmup_lr(self, step):
+        lr_val = (step + 1) / (self.warmup_steps + 1)
+        return [initial_lr * lr_val for initial_lr in self.base_lrs]
+
+    def _get_constant_lr(self, step):
+        return [self.min_lr for _ in self.base_lrs]
+
+    def _get_lr(self, step):
+        """Simple const lr policy"""
+        return self.base_lrs
+
+
+def _squareroot_annealing(initial_lr, step, max_steps, min_lr):
+    mult = ((max_steps - step) / max_steps) ** 0.5
+    out_lr = initial_lr * mult
+    out_lr = max(out_lr, min_lr)
+    return out_lr
+
+
+def _square_annealing(initial_lr, step, max_steps, min_lr):
+    mult = ((max_steps - step) / max_steps) ** 2
+    out_lr = initial_lr * mult
+    out_lr = max(out_lr, min_lr)
+    return out_lr
+
+
+def _cosine_annealing(initial_lr, step, max_steps, min_lr):
+    mult = 0.5 * (1 + math.cos(math.pi * step / max_steps))
+    out_lr = (initial_lr - min_lr) * mult + min_lr
+    return out_lr
+
+
+def _linear_warmup_with_cosine_annealing(
+    max_lr, warmup_steps, step, decay_steps, min_lr
+):
+    assert max_lr > min_lr
+    # Use linear warmup for the initial part.
+    if warmup_steps > 0 and step <= warmup_steps:
+        return max_lr * float(step) / float(warmup_steps)
+
+    # For any steps larger than `decay_steps`, use `min_lr`.
+    if step > warmup_steps + decay_steps:
+        return min_lr
+
+    # If we are done with the warmup period, use the decay style.
+    num_steps_ = step - warmup_steps
+    decay_steps_ = decay_steps
+    decay_ratio = float(num_steps_) / float(decay_steps_)
+    assert decay_ratio >= 0.0
+    assert decay_ratio <= 1.0
+    delta_lr = max_lr - min_lr
+
+    coeff = 0.5 * (math.cos(math.pi * decay_ratio) + 1.0)
+
+    return min_lr + coeff * delta_lr
+
+
+def _poly_decay(initial_lr, step, decay_steps, power, min_lr, cycle):
+    if cycle:
+        multiplier = 1.0 if step == 0 else math.ceil(step / decay_steps)
+        decay_steps *= multiplier
+    else:
+        step = min(step, decay_steps)
+    p = step / decay_steps
+    lr = (initial_lr - min_lr) * math.pow(1.0 - p, power)
+    lr += min_lr
+    return lr
+
+
+def _noam_hold_annealing(
+    initial_lr, step, warmup_steps, hold_steps, decay_rate, min_lr
+):
+    # hold_steps = total number of steps
+    # to hold the LR, not the warmup + hold steps.
+    T_warmup_decay = max(1, warmup_steps**decay_rate)
+    T_hold_decay = max(1, (step - hold_steps) ** decay_rate)
+    lr = (initial_lr * T_warmup_decay) / T_hold_decay
+    lr = max(lr, min_lr)
+    return lr
+
+
+class SquareAnnealing(WarmupPolicy):
+
+    def __init__(self, optimizer, *, max_steps, min_lr=1e-5, last_epoch=-1, **kwargs):
+        super().__init__(
+            optimizer=optimizer,
+            max_steps=max_steps,
+            last_epoch=last_epoch,
+            min_lr=min_lr,
+            **kwargs,
+        )
+
+    def _get_lr(self, step):
+        new_lrs = [
+            _square_annealing(
+                initial_lr=initial_lr,
+                step=step - self.warmup_steps,
+                max_steps=self.max_steps - self.warmup_steps,
+                min_lr=self.min_lr,
+            )
+            for initial_lr in self.base_lrs
+        ]
+        return new_lrs
+
+
+class SquareRootAnnealing(WarmupPolicy):
+
+    def __init__(self, optimizer, *, max_steps, min_lr=0, last_epoch=-1, **kwargs):
+        super().__init__(
+            optimizer=optimizer,
+            max_steps=max_steps,
+            last_epoch=last_epoch,
+            min_lr=min_lr,
+            **kwargs,
+        )
+
+    def _get_lr(self, step):
+        new_lrs = [
+            _squareroot_annealing(
+                initial_lr=initial_lr,
+                step=step,
+                max_steps=self.max_steps,
+                min_lr=self.min_lr,
+            )
+            for initial_lr in self.base_lrs
+        ]
+        return new_lrs
+
+
+class CosineAnnealing(WarmupAnnealHoldPolicy):
+
+    def __init__(self, optimizer, *, max_steps, min_lr=0, last_epoch=-1, **kwargs):
+        super().__init__(
+            optimizer=optimizer,
+            max_steps=max_steps,
+            last_epoch=last_epoch,
+            min_lr=min_lr,
+            **kwargs,
+        )
+
+    def _get_lr(self, step):
+        for initial_lr in self.base_lrs:
+            if initial_lr < self.min_lr:
+                raise ValueError(
+                    f"{self} received an initial learning rate "
+                    f"that was lower than the minimum learning rate."
+                )
+
+        if self.constant_steps is None or self.constant_steps == 0:
+            new_lrs = [
+                _cosine_annealing(
+                    initial_lr=initial_lr,
+                    step=step - self.warmup_steps,
+                    max_steps=self.max_steps - self.warmup_steps,
+                    min_lr=self.min_lr,
+                )
+                for initial_lr in self.base_lrs
+            ]
+        else:
+            new_lrs = self._get_linear_warmup_with_cosine_annealing_lr(step)
+        return new_lrs
+
+    def _get_warmup_lr(self, step):
+        if self.constant_steps is None or self.constant_steps == 0:
+            return super()._get_warmup_lr(step)
+        else:
+            # Use linear warmup for the initial part.
+            return self._get_linear_warmup_with_cosine_annealing_lr(step)
+
+    def _get_constant_lr(self, step):
+        # Only called when `constant_steps` > 0.
+        return self._get_linear_warmup_with_cosine_annealing_lr(step)
+
+    def _get_linear_warmup_with_cosine_annealing_lr(self, step):
+        # Cosine Schedule for Megatron LM,
+        # slightly different warmup schedule + constant LR at the end.
+        new_lrs = [
+            _linear_warmup_with_cosine_annealing(
+                max_lr=self.base_lrs[0],
+                warmup_steps=self.warmup_steps,
+                step=step,
+                decay_steps=self.decay_steps,
+                min_lr=self.min_lr,
+            )
+            for _ in self.base_lrs
+        ]
+        return new_lrs
+
+
+class NoamAnnealing(_LRScheduler):
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        d_model,
+        warmup_steps=None,
+        warmup_ratio=None,
+        max_steps=None,
+        min_lr=0.0,
+        last_epoch=-1,
+    ):
+        self._normalize = d_model ** (-0.5)
+        assert not (
+            warmup_steps is not None and warmup_ratio is not None
+        ), "Either use particular number of step or ratio"
+        assert (
+            warmup_ratio is None or max_steps is not None
+        ), "If there is a ratio, there should be a total steps"
+
+        # It is necessary to assign all attributes *before* __init__,
+        # as class is wrapped by an inner class.
+        self.max_steps = max_steps
+        if warmup_steps is not None:
+            self.warmup_steps = warmup_steps
+        elif warmup_ratio is not None:
+            self.warmup_steps = int(warmup_ratio * max_steps)
+        else:
+            self.warmup_steps = 0
+
+        self.min_lr = min_lr
+        super().__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        if not self._get_lr_called_within_step:
+            warnings.warn(
+                "To get the last learning rate computed "
+                "by the scheduler, please use `get_last_lr()`.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+        step = max(1, self.last_epoch)
+
+        for initial_lr in self.base_lrs:
+            if initial_lr < self.min_lr:
+                raise ValueError(
+                    f"{self} received an initial learning rate "
+                    f"that was lower than the minimum learning rate."
+                )
+
+        new_lrs = [
+            self._noam_annealing(initial_lr=initial_lr, step=step)
+            for initial_lr in self.base_lrs
+        ]
+        return new_lrs
+
+    def _noam_annealing(self, initial_lr, step):
+        if self.warmup_steps > 0:
+            mult = self._normalize * min(
+                step ** (-0.5), step * (self.warmup_steps ** (-1.5))
+            )
+        else:
+            mult = self._normalize * step ** (-0.5)
+
+        out_lr = initial_lr * mult
+        if step > self.warmup_steps:
+            out_lr = max(out_lr, self.min_lr)
+        return out_lr
+
+
+class NoamHoldAnnealing(WarmupHoldPolicy):
+
+    def __init__(
+        self,
+        optimizer,
+        *,
+        max_steps,
+        decay_rate=0.5,
+        min_lr=0.0,
+        last_epoch=-1,
+        **kwargs,
+    ):
+        """
+        From Nemo:
+        Implementation of the Noam Hold Annealing policy
+        from the SqueezeFormer paper.
+
+        Unlike NoamAnnealing, the peak learning rate
+        can be explicitly set for this scheduler.
+        The schedule first performs linear warmup,
+        then holds the peak LR, then decays with some schedule for
+        the remainder of the steps.
+        Therefore the min-lr is still dependent
+        on the hyper parameters selected.
+
+        It's schedule is determined by three factors-
+
+        Warmup Steps: Initial stage, where linear warmup
+            occurs uptil the peak LR is reached. Unlike NoamAnnealing,
+            the peak LR is explicitly stated here instead of a scaling factor.
+
+        Hold Steps: Intermediate stage, where the peak LR
+            is maintained for some number of steps. In this region,
+            the high peak LR allows the model to converge faster
+            if training is stable. However the high LR
+            may also cause instability during training.
+            Should usually be a significant fraction of training
+            steps (around 30-40% of the entire training steps).
+
+        Decay Steps: Final stage, where the LR rapidly decays
+            with some scaling rate (set by decay rate).
+            To attain Noam decay, use 0.5,
+            for Squeezeformer recommended decay, use 1.0.
+            The fast decay after prolonged high LR during
+            hold phase allows for rapid convergence.
+
+        References:
+            - [Squeezeformer:
+            An Efficient Transformer for Automatic Speech Recognition]
+            (https://arxiv.org/abs/2206.00888)
+
+        Args:
+            optimizer: Pytorch compatible Optimizer object.
+            warmup_steps: Number of training steps in warmup stage
+            warmup_ratio: Ratio of warmup steps to total steps
+            hold_steps: Number of training steps to
+                        hold the learning rate after warm up
+            hold_ratio: Ratio of hold steps to total steps
+            max_steps: Total number of steps while training or `None` for
+                infinite training
+            decay_rate: Float value describing the polynomial decay
+                        after the hold period. Default value
+                        of 0.5 corresponds to Noam decay.
+            min_lr: Minimum learning rate.
+        """
+        self.decay_rate = decay_rate
+        super().__init__(
+            optimizer=optimizer,
+            max_steps=max_steps,
+            last_epoch=last_epoch,
+            min_lr=min_lr,
+            **kwargs,
+        )
+
+    def _get_lr(self, step):
+        if self.warmup_steps is None or self.warmup_steps == 0:
+            raise ValueError("Noam scheduler cannot be used without warmup steps")
+
+        if self.hold_steps > 0:
+            hold_steps = self.hold_steps - self.warmup_steps
+        else:
+            hold_steps = 0
+
+        new_lrs = [
+            _noam_hold_annealing(
+                initial_lr,
+                step=step,
+                warmup_steps=self.warmup_steps,
+                hold_steps=hold_steps,
+                decay_rate=self.decay_rate,
+                min_lr=self.min_lr,
+            )
+            for initial_lr in self.base_lrs
+        ]
+        return new_lrs
+
+    def set_step(self, step: int):
+        self.last_epoch = step
+
+
+class ConstantLR(_LRScheduler):
+    """The ConstantLR scheduler
+
+    This scheduler keeps a constant lr
+
+    """
+
+    def __init__(
+        self,
+        optimizer: torch.optim.Optimizer,
+    ):
+        # __init__() must be invoked before setting field
+        # because step() is also invoked in __init__()
+        super().__init__(optimizer)
+
+    def get_lr(self):
+        return self.base_lrs
+
+    def set_step(self, step: int):
+        self.last_epoch = step

cosyvoice/utils/train_utils.py

@@ -0,0 +1,350 @@
+# Copyright (c) 2021 Mobvoi Inc. (authors: Binbin Zhang)
+#               2023 Horizon Inc. (authors: Xingchen Song)
+#               2024 Alibaba Inc (authors: Xiang Lyu)
+#
+# 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.
+
+from contextlib import nullcontext
+import logging
+import os
+import torch
+import json
+import re
+import datetime
+import yaml
+
+import deepspeed
+import torch.optim as optim
+import torch.distributed as dist
+
+from torch.utils.tensorboard import SummaryWriter
+from torch.utils.data import DataLoader
+from torch.nn.utils import clip_grad_norm_
+
+from deepspeed.runtime.zero.stage_1_and_2 import (
+    estimate_zero2_model_states_mem_needs_all_live,
+)
+
+from cosyvoice.dataset.dataset import Dataset
+from cosyvoice.utils.scheduler import (
+    WarmupLR,
+    NoamHoldAnnealing,
+    ConstantLR,
+)
+
+
+def init_distributed(args):
+    world_size = int(os.environ.get("WORLD_SIZE", 1))
+    local_rank = int(os.environ.get("LOCAL_RANK", 0))
+    rank = int(os.environ.get("RANK", 0))
+    logging.info(
+        "training on multiple gpus, this gpu {}".format(local_rank)
+        + ", rank {}, world_size {}".format(rank, world_size)
+    )
+    if args.train_engine == "torch_ddp":
+        torch.cuda.set_device(local_rank)
+        dist.init_process_group(args.dist_backend)
+    else:
+        deepspeed.init_distributed(dist_backend=args.dist_backend)
+    return world_size, local_rank, rank
+
+
+def init_dataset_and_dataloader(args, configs):
+    train_dataset = Dataset(
+        args.train_data,
+        data_pipeline=configs["data_pipeline"],
+        mode="train",
+        shuffle=True,
+        partition=True,
+    )
+    cv_dataset = Dataset(
+        args.cv_data,
+        data_pipeline=configs["data_pipeline"],
+        mode="train",
+        shuffle=False,
+        partition=False,
+    )
+
+    # do not use persistent_workers=True, as whisper tokenizer opens tiktoken file each time when the for loop starts
+    train_data_loader = DataLoader(
+        train_dataset,
+        batch_size=None,
+        pin_memory=args.pin_memory,
+        num_workers=args.num_workers,
+        prefetch_factor=args.prefetch,
+    )
+    cv_data_loader = DataLoader(
+        cv_dataset,
+        batch_size=None,
+        pin_memory=args.pin_memory,
+        num_workers=args.num_workers,
+        prefetch_factor=args.prefetch,
+    )
+    return train_dataset, cv_dataset, train_data_loader, cv_data_loader
+
+
+def check_modify_and_save_config(args, configs):
+    if args.train_engine == "torch_ddp":
+        configs["train_conf"]["dtype"] = "fp32"
+    else:
+        with open(args.deepspeed_config, "r") as fin:
+            ds_configs = json.load(fin)
+        if "fp16" in ds_configs and ds_configs["fp16"]["enabled"]:
+            configs["train_conf"]["dtype"] = "fp16"
+        elif "bf16" in ds_configs and ds_configs["bf16"]["enabled"]:
+            configs["train_conf"]["dtype"] = "bf16"
+        else:
+            configs["train_conf"]["dtype"] = "fp32"
+        assert ds_configs["train_micro_batch_size_per_gpu"] == 1
+        # if use deepspeed, override ddp config
+        configs["train_conf"]["save_per_step"] = int(
+            configs["train_conf"]["save_per_step"]
+            * configs["train_conf"]["accum_grad"]
+            / ds_configs["gradient_accumulation_steps"]
+        )
+        configs["train_conf"]["accum_grad"] = ds_configs["gradient_accumulation_steps"]
+        configs["train_conf"]["grad_clip"] = ds_configs["gradient_clipping"]
+        configs["train_conf"]["log_interval"] = ds_configs["steps_per_print"]
+    return configs
+
+
+def wrap_cuda_model(args, model):
+    local_world_size = int(os.environ.get("LOCAL_WORLD_SIZE", 1))
+    world_size = int(os.environ.get("WORLD_SIZE", 1))
+    if args.train_engine == "torch_ddp":  # native pytorch ddp
+        assert torch.cuda.is_available()
+        model.cuda()
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, find_unused_parameters=True
+        )
+    else:
+        if int(os.environ.get("RANK", 0)) == 0:
+            logging.info("Estimating model states memory needs (zero2)...")
+            estimate_zero2_model_states_mem_needs_all_live(
+                model,
+                num_gpus_per_node=local_world_size,
+                num_nodes=world_size // local_world_size,
+            )
+    return model
+
+
+def init_optimizer_and_scheduler(args, configs, model):
+    if configs["train_conf"]["optim"] == "adam":
+        optimizer = optim.Adam(
+            model.parameters(), **configs["train_conf"]["optim_conf"]
+        )
+    elif configs["train_conf"]["optim"] == "adamw":
+        optimizer = optim.AdamW(
+            model.parameters(), **configs["train_conf"]["optim_conf"]
+        )
+    else:
+        raise ValueError("unknown optimizer: " + configs["train_conf"])
+
+    if configs["train_conf"]["scheduler"] == "warmuplr":
+        scheduler_type = WarmupLR
+        scheduler = WarmupLR(optimizer, **configs["train_conf"]["scheduler_conf"])
+    elif configs["train_conf"]["scheduler"] == "NoamHoldAnnealing":
+        scheduler_type = NoamHoldAnnealing
+        scheduler = NoamHoldAnnealing(
+            optimizer, **configs["train_conf"]["scheduler_conf"]
+        )
+    elif configs["train_conf"]["scheduler"] == "constantlr":
+        scheduler_type = ConstantLR
+        scheduler = ConstantLR(optimizer)
+    else:
+        raise ValueError("unknown scheduler: " + configs["train_conf"])
+
+    # use deepspeed optimizer for speedup
+    if args.train_engine == "deepspeed":
+
+        def scheduler(opt):
+            return scheduler_type(opt, **configs["train_conf"]["scheduler_conf"])
+
+        model, optimizer, _, scheduler = deepspeed.initialize(
+            args=args,
+            model=model,
+            optimizer=None,
+            lr_scheduler=scheduler,
+            model_parameters=model.parameters(),
+        )
+
+    return model, optimizer, scheduler
+
+
+def init_summarywriter(args):
+    writer = None
+    if int(os.environ.get("RANK", 0)) == 0:
+        os.makedirs(args.model_dir, exist_ok=True)
+        writer = SummaryWriter(args.tensorboard_dir)
+    return writer
+
+
+def save_model(model, model_name, info_dict):
+    rank = int(os.environ.get("RANK", 0))
+    model_dir = info_dict["model_dir"]
+    save_model_path = os.path.join(model_dir, "{}.pt".format(model_name))
+
+    if info_dict["train_engine"] == "torch_ddp":
+        if rank == 0:
+            torch.save(model.module.state_dict(), save_model_path)
+    else:
+        with torch.no_grad():
+            model.save_checkpoint(
+                save_dir=model_dir, tag=model_name, client_state=info_dict
+            )
+    if rank == 0:
+        info_path = re.sub(".pt$", ".yaml", save_model_path)
+        info_dict["save_time"] = datetime.datetime.now().strftime("%d/%m/%Y %H:%M:%S")
+        with open(info_path, "w") as fout:
+            data = yaml.dump(info_dict)
+            fout.write(data)
+        logging.info(
+            "[Rank {}] Checkpoint: save to checkpoint {}".format(rank, save_model_path)
+        )
+
+
+def cosyvoice_join(group_join, info_dict):
+    world_size = int(os.environ.get("WORLD_SIZE", 1))
+    local_rank = int(os.environ.get("LOCAL_RANK", 0))
+    rank = int(os.environ.get("RANK", 0))
+
+    if info_dict["batch_idx"] != 0:
+        # we try to join all rank in both ddp and deepspeed mode, in case different rank has different lr
+        try:
+            dist.monitored_barrier(
+                group=group_join, timeout=group_join.options._timeout
+            )
+            return False
+        except RuntimeError as e:
+            logging.info(
+                "Detected uneven workload distribution: {}\n".format(e)
+                + "Break current worker to manually join all workers, "
+                + "world_size {}, current rank {}, current local_rank {}\n".format(
+                    world_size, rank, local_rank
+                )
+            )
+            return True
+    else:
+        return False
+
+
+def batch_forward(model, batch, info_dict):
+    device = int(os.environ.get("LOCAL_RANK", 0))
+
+    dtype = info_dict["dtype"]
+    if dtype == "fp16":
+        dtype = torch.float16
+    elif dtype == "bf16":
+        dtype = torch.bfloat16
+    else:  # fp32
+        dtype = torch.float32
+
+    if info_dict["train_engine"] == "torch_ddp":
+        autocast = nullcontext()
+    else:
+        autocast = torch.cuda.amp.autocast(
+            enabled=True, dtype=dtype, cache_enabled=False
+        )
+
+    with autocast:
+        info_dict["loss_dict"] = model(batch, device)
+    return info_dict
+
+
+def batch_backward(model, info_dict):
+    if info_dict["train_engine"] == "deepspeed":
+        scaled_loss = model.backward(info_dict["loss_dict"]["loss"])
+    else:
+        scaled_loss = info_dict["loss_dict"]["loss"] / info_dict["accum_grad"]
+        scaled_loss.backward()
+
+    info_dict["loss_dict"]["loss"] = scaled_loss
+    return info_dict
+
+
+def update_parameter_and_lr(model, optimizer, scheduler, info_dict):
+    grad_norm = 0.0
+    if info_dict["train_engine"] == "deepspeed":
+        info_dict["is_gradient_accumulation_boundary"] = (
+            model.is_gradient_accumulation_boundary()
+        )
+        model.step()
+        grad_norm = model.get_global_grad_norm()
+    elif (info_dict["batch_idx"] + 1) % info_dict["accum_grad"] == 0:
+        grad_norm = clip_grad_norm_(model.parameters(), info_dict["grad_clip"])
+        if torch.isfinite(grad_norm):
+            optimizer.step()
+        optimizer.zero_grad()
+        scheduler.step()
+    info_dict["lr"] = optimizer.param_groups[0]["lr"]
+    info_dict["grad_norm"] = grad_norm
+    return info_dict
+
+
+def log_per_step(writer, info_dict):
+    tag = info_dict["tag"]
+    epoch = info_dict.get("epoch", 0)
+    step = info_dict["step"]
+    batch_idx = info_dict["batch_idx"]
+    loss_dict = info_dict["loss_dict"]
+    rank = int(os.environ.get("RANK", 0))
+
+    # only rank 0 write to tensorboard to avoid multi-process write
+    if writer is not None:
+        if (
+            info_dict["train_engine"] == "deepspeed"
+            and info_dict["is_gradient_accumulation_boundary"] is True
+        ) or (
+            info_dict["train_engine"] == "torch_ddp"
+            and (info_dict["batch_idx"] + 1) % info_dict["accum_grad"] == 0
+        ):
+            for k in ["epoch", "lr", "grad_norm"]:
+                writer.add_scalar("{}/{}".format(tag, k), info_dict[k], step + 1)
+            for k, v in loss_dict.items():
+                writer.add_scalar("{}/{}".format(tag, k), v, step + 1)
+
+    # TRAIN & CV, Shell log (stdout)
+    if (info_dict["batch_idx"] + 1) % info_dict["log_interval"] == 0:
+        log_str = "{} Batch {}/{} ".format(tag, epoch, batch_idx + 1)
+        for name, value in loss_dict.items():
+            log_str += "{} {:.6f} ".format(name, value)
+        if tag == "TRAIN":
+            log_str += "lr {:.8f} grad_norm {:.6f}".format(
+                info_dict["lr"], info_dict["grad_norm"]
+            )
+        log_str += " rank {}".format(rank)
+        logging.debug(log_str)
+
+
+def log_per_save(writer, info_dict):
+    tag = info_dict["tag"]
+    epoch = info_dict["epoch"]
+    step = info_dict["step"]
+    loss_dict = info_dict["loss_dict"]
+    lr = info_dict["lr"]
+    rank = int(os.environ.get("RANK", 0))
+    logging.info(
+        "Epoch {} Step {} CV info lr {} {} rank {}".format(
+            epoch,
+            step + 1,
+            lr,
+            rank,
+            " ".join(["{}_{}".format(k, v) for k, v in loss_dict.items()]),
+        )
+    )
+
+    if writer is not None:
+        for k in ["epoch", "lr"]:
+            writer.add_scalar("{}/{}".format(tag, k), info_dict[k], step + 1)
+        for k, v in loss_dict.items():
+            writer.add_scalar("{}/{}".format(tag, k), v, step + 1)

funasr_detach/__init__.py

@@ -0,0 +1,38 @@
+"""Initialize funasr package."""
+
+import os
+import pkgutil
+import importlib
+
+dirname = os.path.dirname(__file__)
+version_file = os.path.join(dirname, "version.txt")
+with open(version_file, "r") as f:
+    __version__ = f.read().strip()
+
+
+import importlib
+import pkgutil
+
+
+def import_submodules(package, recursive=True):
+    if isinstance(package, str):
+        package = importlib.import_module(package)
+    results = {}
+    for loader, name, is_pkg in pkgutil.walk_packages(
+        package.__path__, package.__name__ + "."
+    ):
+        try:
+            results[name] = importlib.import_module(name)
+        except Exception as e:
+            # 如果想要看到导入错误的具体信息，可以取消注释下面的行
+            # print(f"Failed to import {name}: {e}")
+            pass
+        if recursive and is_pkg:
+            results.update(import_submodules(name))
+    return results
+
+
+import_submodules(__name__)
+
+from funasr_detach.auto.auto_model import AutoModel
+from funasr_detach.auto.auto_frontend import AutoFrontend

funasr_detach/auto/auto_frontend.py

@@ -0,0 +1,90 @@
+import time
+import logging
+from tqdm import tqdm
+
+from funasr_detach.register import tables
+from funasr_detach.download.download_from_hub import download_model
+from funasr_detach.utils.load_utils import load_audio_text_image_video, extract_fbank
+from funasr_detach.auto.auto_model import prepare_data_iterator
+from funasr_detach.auto.auto_model import prepare_data_iterator
+
+
+class AutoFrontend:
+    def __init__(self, **kwargs):
+        assert "model" in kwargs
+        if "model_conf" not in kwargs:
+            logging.info(
+                "download models from model hub: {}".format(
+                    kwargs.get("model_hub", "ms")
+                )
+            )
+            kwargs = download_model(**kwargs)
+
+        # build frontend
+        frontend = kwargs.get("frontend", None)
+        if frontend is not None:
+            frontend_class = tables.frontend_classes.get(frontend)
+            frontend = frontend_class(**kwargs["frontend_conf"])
+
+        self.frontend = frontend
+        if "frontend" in kwargs:
+            del kwargs["frontend"]
+        self.kwargs = kwargs
+
+    def __call__(self, input, input_len=None, kwargs=None, **cfg):
+
+        kwargs = self.kwargs if kwargs is None else kwargs
+        kwargs.update(cfg)
+
+        key_list, data_list = prepare_data_iterator(input, input_len=input_len)
+        batch_size = kwargs.get("batch_size", 1)
+        device = kwargs.get("device", "cpu")
+        if device == "cpu":
+            batch_size = 1
+
+        meta_data = {}
+
+        result_list = []
+        num_samples = len(data_list)
+        pbar = tqdm(colour="blue", total=num_samples + 1, dynamic_ncols=True)
+
+        time0 = time.perf_counter()
+        for beg_idx in range(0, num_samples, batch_size):
+            end_idx = min(num_samples, beg_idx + batch_size)
+            data_batch = data_list[beg_idx:end_idx]
+            key_batch = key_list[beg_idx:end_idx]
+
+            # extract fbank feats
+            time1 = time.perf_counter()
+            audio_sample_list = load_audio_text_image_video(
+                data_batch, fs=self.frontend.fs, audio_fs=kwargs.get("fs", 16000)
+            )
+            time2 = time.perf_counter()
+            meta_data["load_data"] = f"{time2 - time1:0.3f}"
+            speech, speech_lengths = extract_fbank(
+                audio_sample_list,
+                data_type=kwargs.get("data_type", "sound"),
+                frontend=self.frontend,
+                **kwargs,
+            )
+            time3 = time.perf_counter()
+            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
+            meta_data["batch_data_time"] = (
+                speech_lengths.sum().item()
+                * self.frontend.frame_shift
+                * self.frontend.lfr_n
+                / 1000
+            )
+
+            speech.to(device=device), speech_lengths.to(device=device)
+            batch = {"input": speech, "input_len": speech_lengths, "key": key_batch}
+            result_list.append(batch)
+
+            pbar.update(1)
+            description = f"{meta_data}, "
+            pbar.set_description(description)
+
+        time_end = time.perf_counter()
+        pbar.set_description(f"time escaped total: {time_end - time0:0.3f}")
+
+        return result_list

funasr_detach/auto/auto_model.py

@@ -0,0 +1,573 @@
+import json
+import time
+import copy
+import torch
+import random
+import string
+import logging
+import os.path
+import numpy as np
+from tqdm import tqdm
+
+from funasr_detach.register import tables
+from funasr_detach.utils.load_utils import load_bytes
+from funasr_detach.download.file import download_from_url
+from funasr_detach.download.download_from_hub import download_model
+from funasr_detach.utils.vad_utils import slice_padding_audio_samples
+from funasr_detach.train_utils.set_all_random_seed import set_all_random_seed
+from funasr_detach.train_utils.load_pretrained_model import load_pretrained_model
+from funasr_detach.utils.load_utils import load_audio_text_image_video
+from funasr_detach.utils.timestamp_tools import timestamp_sentence
+from funasr_detach.models.campplus.utils import sv_chunk, postprocess, distribute_spk
+
+try:
+    from funasr_detach.models.campplus.cluster_backend import ClusterBackend
+except:
+    print("If you want to use the speaker diarization, please `pip install hdbscan`")
+
+
+def prepare_data_iterator(data_in, input_len=None, data_type=None, key=None):
+    """
+
+    :param input:
+    :param input_len:
+    :param data_type:
+    :param frontend:
+    :return:
+    """
+    data_list = []
+    key_list = []
+    filelist = [".scp", ".txt", ".json", ".jsonl"]
+
+    chars = string.ascii_letters + string.digits
+    if isinstance(data_in, str) and data_in.startswith("http"):  # url
+        data_in = download_from_url(data_in)
+    if isinstance(data_in, str) and os.path.exists(
+        data_in
+    ):  # wav_path; filelist: wav.scp, file.jsonl;text.txt;
+        _, file_extension = os.path.splitext(data_in)
+        file_extension = file_extension.lower()
+        if file_extension in filelist:  # filelist: wav.scp, file.jsonl;text.txt;
+            with open(data_in, encoding="utf-8") as fin:
+                for line in fin:
+                    key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+                    if data_in.endswith(
+                        ".jsonl"
+                    ):  # file.jsonl: json.dumps({"source": data})
+                        lines = json.loads(line.strip())
+                        data = lines["source"]
+                        key = data["key"] if "key" in data else key
+                    else:  # filelist, wav.scp, text.txt: id \t data or data
+                        lines = line.strip().split(maxsplit=1)
+                        data = lines[1] if len(lines) > 1 else lines[0]
+                        key = lines[0] if len(lines) > 1 else key
+
+                    data_list.append(data)
+                    key_list.append(key)
+        else:
+            key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+            data_list = [data_in]
+            key_list = [key]
+    elif isinstance(data_in, (list, tuple)):
+        if data_type is not None and isinstance(
+            data_type, (list, tuple)
+        ):  # mutiple inputs
+            data_list_tmp = []
+            for data_in_i, data_type_i in zip(data_in, data_type):
+                key_list, data_list_i = prepare_data_iterator(
+                    data_in=data_in_i, data_type=data_type_i
+                )
+                data_list_tmp.append(data_list_i)
+            data_list = []
+            for item in zip(*data_list_tmp):
+                data_list.append(item)
+        else:
+            # [audio sample point, fbank, text]
+            data_list = data_in
+            key_list = [
+                "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+                for _ in range(len(data_in))
+            ]
+    else:  # raw text; audio sample point, fbank; bytes
+        if isinstance(data_in, bytes):  # audio bytes
+            data_in = load_bytes(data_in)
+        if key is None:
+            key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
+        data_list = [data_in]
+        key_list = [key]
+
+    return key_list, data_list
+
+
+class AutoModel:
+
+    def __init__(self, **kwargs):
+        if not kwargs.get("disable_log", False):
+            tables.print()
+
+        model, kwargs = self.build_model(**kwargs)
+
+        # if vad_model is not None, build vad model else None
+        vad_model = kwargs.get("vad_model", None)
+        vad_kwargs = kwargs.get("vad_model_revision", None)
+        if vad_model is not None:
+            logging.info("Building VAD model.")
+            vad_kwargs = {
+                "model": vad_model,
+                "model_revision": vad_kwargs,
+                "device": kwargs["device"],
+            }
+            vad_model, vad_kwargs = self.build_model(**vad_kwargs)
+
+        # if punc_model is not None, build punc model else None
+        punc_model = kwargs.get("punc_model", None)
+        punc_kwargs = kwargs.get("punc_model_revision", None)
+        if punc_model is not None:
+            logging.info("Building punc model.")
+            punc_kwargs = {
+                "model": punc_model,
+                "model_revision": punc_kwargs,
+                "device": kwargs["device"],
+            }
+            punc_model, punc_kwargs = self.build_model(**punc_kwargs)
+
+        # if spk_model is not None, build spk model else None
+        spk_model = kwargs.get("spk_model", None)
+        spk_kwargs = kwargs.get("spk_model_revision", None)
+        if spk_model is not None:
+            logging.info("Building SPK model.")
+            spk_kwargs = {
+                "model": spk_model,
+                "model_revision": spk_kwargs,
+                "device": kwargs["device"],
+            }
+            spk_model, spk_kwargs = self.build_model(**spk_kwargs)
+            self.cb_model = ClusterBackend().to(kwargs["device"])
+            spk_mode = kwargs.get("spk_mode", "punc_segment")
+            if spk_mode not in ["default", "vad_segment", "punc_segment"]:
+                logging.error(
+                    "spk_mode should be one of default, vad_segment and punc_segment."
+                )
+            self.spk_mode = spk_mode
+
+        self.kwargs = kwargs
+        self.model = model
+        self.vad_model = vad_model
+        self.vad_kwargs = vad_kwargs
+        self.punc_model = punc_model
+        self.punc_kwargs = punc_kwargs
+        self.spk_model = spk_model
+        self.spk_kwargs = spk_kwargs
+        self.model_path = kwargs.get("model_path")
+
+    def build_model(self, **kwargs):
+        assert "model" in kwargs
+        if "model_conf" not in kwargs:
+            logging.info(
+                "download models from model hub: {}".format(
+                    kwargs.get("model_hub", "ms")
+                )
+            )
+            kwargs = download_model(**kwargs)
+
+        set_all_random_seed(kwargs.get("seed", 0))
+
+        device = kwargs.get("device", "cuda")
+        if not torch.cuda.is_available() or kwargs.get("ngpu", 1) == 0:
+            device = "cpu"
+            kwargs["batch_size"] = 1
+        kwargs["device"] = device
+
+        if kwargs.get("ncpu", None):
+            torch.set_num_threads(kwargs.get("ncpu"))
+
+        # build tokenizer
+        tokenizer = kwargs.get("tokenizer", None)
+        if tokenizer is not None:
+            tokenizer_class = tables.tokenizer_classes.get(tokenizer)
+            tokenizer = tokenizer_class(**kwargs["tokenizer_conf"])
+            kwargs["tokenizer"] = tokenizer
+            kwargs["token_list"] = tokenizer.token_list
+            vocab_size = len(tokenizer.token_list)
+        else:
+            vocab_size = -1
+
+        # build frontend
+        frontend = kwargs.get("frontend", None)
+        if frontend is not None:
+            frontend_class = tables.frontend_classes.get(frontend)
+            frontend = frontend_class(**kwargs["frontend_conf"])
+            kwargs["frontend"] = frontend
+            kwargs["input_size"] = frontend.output_size()
+
+        # build model
+        model_class = tables.model_classes.get(kwargs["model"])
+        model = model_class(**kwargs, **kwargs["model_conf"], vocab_size=vocab_size)
+
+        model.to(device)
+
+        # init_param
+        init_param = kwargs.get("init_param", None)
+        if init_param is not None:
+            logging.info(f"Loading pretrained params from {init_param}")
+            load_pretrained_model(
+                model=model,
+                path=init_param,
+                ignore_init_mismatch=kwargs.get("ignore_init_mismatch", False),
+                oss_bucket=kwargs.get("oss_bucket", None),
+                scope_map=kwargs.get("scope_map", None),
+                excludes=kwargs.get("excludes", None),
+            )
+
+        return model, kwargs
+
+    def __call__(self, *args, **cfg):
+        kwargs = self.kwargs
+        kwargs.update(cfg)
+        res = self.model(*args, kwargs)
+        return res
+
+    def generate(self, input, input_len=None, **cfg):
+        if self.vad_model is None:
+            return self.inference(input, input_len=input_len, **cfg)
+
+        else:
+            return self.inference_with_vad(input, input_len=input_len, **cfg)
+
+    def inference(
+        self, input, input_len=None, model=None, kwargs=None, key=None, **cfg
+    ):
+        kwargs = self.kwargs if kwargs is None else kwargs
+        kwargs.update(cfg)
+        model = self.model if model is None else model
+        model = model.cuda()
+        model.eval()
+
+        batch_size = kwargs.get("batch_size", 1)
+        # if kwargs.get("device", "cpu") == "cpu":
+        #     batch_size = 1
+
+        key_list, data_list = prepare_data_iterator(
+            input, input_len=input_len, data_type=kwargs.get("data_type", None), key=key
+        )
+
+        speed_stats = {}
+        asr_result_list = []
+        num_samples = len(data_list)
+        disable_pbar = kwargs.get("disable_pbar", False)
+        pbar = (
+            tqdm(colour="blue", total=num_samples, dynamic_ncols=True)
+            if not disable_pbar
+            else None
+        )
+        time_speech_total = 0.0
+        time_escape_total = 0.0
+        for beg_idx in range(0, num_samples, batch_size):
+            end_idx = min(num_samples, beg_idx + batch_size)
+            data_batch = data_list[beg_idx:end_idx]
+            key_batch = key_list[beg_idx:end_idx]
+            batch = {"data_in": data_batch, "key": key_batch}
+            if (end_idx - beg_idx) == 1 and kwargs.get(
+                "data_type", None
+            ) == "fbank":  # fbank
+                batch["data_in"] = data_batch[0]
+                batch["data_lengths"] = input_len
+
+            time1 = time.perf_counter()
+            with torch.no_grad():
+                results, meta_data = model.inference(**batch, **kwargs)
+            time2 = time.perf_counter()
+
+            asr_result_list.extend(results)
+
+            # batch_data_time = time_per_frame_s * data_batch_i["speech_lengths"].sum().item()
+            batch_data_time = meta_data.get("batch_data_time", -1)
+            time_escape = time2 - time1
+            speed_stats["load_data"] = meta_data.get("load_data", 0.0)
+            speed_stats["extract_feat"] = meta_data.get("extract_feat", 0.0)
+            speed_stats["forward"] = f"{time_escape:0.3f}"
+            speed_stats["batch_size"] = f"{len(results)}"
+            speed_stats["time_cost"] = f"{(time_escape)}"
+            speed_stats["rtf"] = f"{(time_escape) / batch_data_time:0.3f}"
+            description = f"{speed_stats}, "
+            if pbar:
+                pbar.update(1)
+                pbar.set_description(description)
+            time_speech_total += batch_data_time
+            time_escape_total += time_escape
+
+        if pbar:
+            # pbar.update(1)
+            pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
+        torch.cuda.empty_cache()
+        return asr_result_list
+
+    def inference_with_vad(self, input, input_len=None, **cfg):
+
+        # step.1: compute the vad model
+        self.vad_kwargs.update(cfg)
+        beg_vad = time.time()
+        res = self.inference(
+            input,
+            input_len=input_len,
+            model=self.vad_model,
+            kwargs=self.vad_kwargs,
+            **cfg,
+        )
+        end_vad = time.time()
+        print(f"time cost vad: {end_vad - beg_vad:0.3f}")
+
+        # step.2 compute asr model
+        model = self.model
+        kwargs = self.kwargs
+        kwargs.update(cfg)
+        batch_size = int(kwargs.get("batch_size_s", 300)) * 1000
+        batch_size_threshold_ms = int(kwargs.get("batch_size_threshold_s", 60)) * 1000
+        kwargs["batch_size"] = batch_size
+
+        key_list, data_list = prepare_data_iterator(
+            input, input_len=input_len, data_type=kwargs.get("data_type", None)
+        )
+        results_ret_list = []
+        time_speech_total_all_samples = 1e-6
+
+        beg_total = time.time()
+        pbar_total = tqdm(colour="red", total=len(res), dynamic_ncols=True)
+        for i in range(len(res)):
+            key = res[i]["key"]
+            vadsegments = res[i]["value"]
+            input_i = data_list[i]
+            speech = load_audio_text_image_video(
+                input_i, fs=kwargs["frontend"].fs, audio_fs=kwargs.get("fs", 16000)
+            )
+            speech_lengths = len(speech)
+            n = len(vadsegments)
+            data_with_index = [(vadsegments[i], i) for i in range(n)]
+            sorted_data = sorted(data_with_index, key=lambda x: x[0][1] - x[0][0])
+            results_sorted = []
+
+            if not len(sorted_data):
+                logging.info("decoding, utt: {}, empty speech".format(key))
+                continue
+
+            if len(sorted_data) > 0 and len(sorted_data[0]) > 0:
+                batch_size = max(
+                    batch_size, sorted_data[0][0][1] - sorted_data[0][0][0]
+                )
+
+            batch_size_ms_cum = 0
+            beg_idx = 0
+            beg_asr_total = time.time()
+            time_speech_total_per_sample = speech_lengths / 16000
+            time_speech_total_all_samples += time_speech_total_per_sample
+
+            all_segments = []
+            for j, _ in enumerate(range(0, n)):
+                # pbar_sample.update(1)
+                batch_size_ms_cum += sorted_data[j][0][1] - sorted_data[j][0][0]
+                if (
+                    j < n - 1
+                    and (
+                        batch_size_ms_cum
+                        + sorted_data[j + 1][0][1]
+                        - sorted_data[j + 1][0][0]
+                    )
+                    < batch_size
+                    and (sorted_data[j + 1][0][1] - sorted_data[j + 1][0][0])
+                    < batch_size_threshold_ms
+                ):
+                    continue
+                batch_size_ms_cum = 0
+                end_idx = j + 1
+                speech_j, speech_lengths_j = slice_padding_audio_samples(
+                    speech, speech_lengths, sorted_data[beg_idx:end_idx]
+                )
+                results = self.inference(
+                    speech_j,
+                    input_len=None,
+                    model=model,
+                    kwargs=kwargs,
+                    disable_pbar=True,
+                    **cfg,
+                )
+                if self.spk_model is not None:
+                    # compose vad segments: [[start_time_sec, end_time_sec, speech], [...]]
+                    for _b in range(len(speech_j)):
+                        vad_segments = [
+                            [
+                                sorted_data[beg_idx:end_idx][_b][0][0] / 1000.0,
+                                sorted_data[beg_idx:end_idx][_b][0][1] / 1000.0,
+                                np.array(speech_j[_b]),
+                            ]
+                        ]
+                        segments = sv_chunk(vad_segments)
+                        all_segments.extend(segments)
+                        speech_b = [i[2] for i in segments]
+                        spk_res = self.inference(
+                            speech_b,
+                            input_len=None,
+                            model=self.spk_model,
+                            kwargs=kwargs,
+                            disable_pbar=True,
+                            **cfg,
+                        )
+                        results[_b]["spk_embedding"] = spk_res[0]["spk_embedding"]
+                beg_idx = end_idx
+                if len(results) < 1:
+                    continue
+                results_sorted.extend(results)
+
+            restored_data = [0] * n
+            for j in range(n):
+                index = sorted_data[j][1]
+                restored_data[index] = results_sorted[j]
+            result = {}
+
+            # results combine for texts, timestamps, speaker embeddings and others
+            # TODO: rewrite for clean code
+            for j in range(n):
+                for k, v in restored_data[j].items():
+                    if k.startswith("timestamp"):
+                        if k not in result:
+                            result[k] = []
+                        for t in restored_data[j][k]:
+                            t[0] += vadsegments[j][0]
+                            t[1] += vadsegments[j][0]
+                        result[k].extend(restored_data[j][k])
+                    elif k == "spk_embedding":
+                        if k not in result:
+                            result[k] = restored_data[j][k]
+                        else:
+                            result[k] = torch.cat(
+                                [result[k], restored_data[j][k]], dim=0
+                            )
+                    elif "text" in k:
+                        if k not in result:
+                            result[k] = restored_data[j][k]
+                        else:
+                            result[k] += " " + restored_data[j][k]
+                    else:
+                        if k not in result:
+                            result[k] = restored_data[j][k]
+                        else:
+                            result[k] += restored_data[j][k]
+
+            return_raw_text = kwargs.get("return_raw_text", False)
+            # step.3 compute punc model
+            if self.punc_model is not None:
+                self.punc_kwargs.update(cfg)
+                punc_res = self.inference(
+                    result["text"],
+                    model=self.punc_model,
+                    kwargs=self.punc_kwargs,
+                    disable_pbar=True,
+                    **cfg,
+                )
+                raw_text = copy.copy(result["text"])
+                if return_raw_text:
+                    result["raw_text"] = raw_text
+                result["text"] = punc_res[0]["text"]
+            else:
+                raw_text = None
+
+            # speaker embedding cluster after resorted
+            if self.spk_model is not None and kwargs.get("return_spk_res", True):
+                if raw_text is None:
+                    logging.error("Missing punc_model, which is required by spk_model.")
+                all_segments = sorted(all_segments, key=lambda x: x[0])
+                spk_embedding = result["spk_embedding"]
+                labels = self.cb_model(
+                    spk_embedding.cpu(), oracle_num=kwargs.get("preset_spk_num", None)
+                )
+                # del result['spk_embedding']
+                sv_output = postprocess(all_segments, None, labels, spk_embedding.cpu())
+                if self.spk_mode == "vad_segment":  # recover sentence_list
+                    sentence_list = []
+                    for res, vadsegment in zip(restored_data, vadsegments):
+                        if "timestamp" not in res:
+                            logging.error(
+                                "Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+                                           and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
+                                           can predict timestamp, and speaker diarization relies on timestamps."
+                            )
+                        sentence_list.append(
+                            {
+                                "start": vadsegment[0],
+                                "end": vadsegment[1],
+                                "sentence": res["text"],
+                                "timestamp": res["timestamp"],
+                            }
+                        )
+                elif self.spk_mode == "punc_segment":
+                    if "timestamp" not in result:
+                        logging.error(
+                            "Only 'iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch' \
+                                       and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
+                                       can predict timestamp, and speaker diarization relies on timestamps."
+                        )
+                    sentence_list = timestamp_sentence(
+                        punc_res[0]["punc_array"],
+                        result["timestamp"],
+                        raw_text,
+                        return_raw_text=return_raw_text,
+                    )
+                distribute_spk(sentence_list, sv_output)
+                result["sentence_info"] = sentence_list
+            elif kwargs.get("sentence_timestamp", False):
+                sentence_list = timestamp_sentence(
+                    punc_res[0]["punc_array"],
+                    result["timestamp"],
+                    raw_text,
+                    return_raw_text=return_raw_text,
+                )
+                result["sentence_info"] = sentence_list
+            if "spk_embedding" in result:
+                del result["spk_embedding"]
+
+            result["key"] = key
+            results_ret_list.append(result)
+            end_asr_total = time.time()
+            time_escape_total_per_sample = end_asr_total - beg_asr_total
+            pbar_total.update(1)
+            pbar_total.set_description(
+                f"rtf_avg: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
+                f"time_speech: {time_speech_total_per_sample: 0.3f}, "
+                f"time_escape: {time_escape_total_per_sample:0.3f}"
+            )
+
+        return results_ret_list
+
+    def infer_encoder(
+        self, input, input_len=None, model=None, kwargs=None, key=None, **cfg
+    ):
+        kwargs = self.kwargs if kwargs is None else kwargs
+        kwargs.update(cfg)
+        model = self.model if model is None else model
+        model = model.cuda()
+        model.eval()
+
+        batch_size = kwargs.get("batch_size", 1)
+
+        key_list, data_list = prepare_data_iterator(
+            input, input_len=input_len, data_type=kwargs.get("data_type", None), key=key
+        )
+
+        asr_result_list = []
+        num_samples = len(data_list)
+        for beg_idx in range(0, num_samples, batch_size):
+            end_idx = min(num_samples, beg_idx + batch_size)
+            data_batch = data_list[beg_idx:end_idx]
+            key_batch = key_list[beg_idx:end_idx]
+            batch = {"data_in": data_batch, "key": key_batch}
+            if (end_idx - beg_idx) == 1 and kwargs.get(
+                "data_type", None
+            ) == "fbank":  # fbank
+                batch["data_in"] = data_batch[0]
+                batch["data_lengths"] = input_len
+
+            with torch.no_grad():
+                results, meta_data, cache = model.infer_encoder(**batch, **kwargs)
+            asr_result_list.extend(results)
+
+        torch.cuda.empty_cache()
+        return asr_result_list, cache

funasr_detach/auto/auto_tokenizer.py

@@ -0,0 +1,7 @@
+class AutoTokenizer:
+    """
+    Undo
+    """
+
+    def __init__(self):
+        pass