minicpmo-utils 0.0.5__tar.gz

minicpmo_utils-0.0.5/PKG-INFO

@@ -0,0 +1,116 @@
+Metadata-Version: 2.4
+Name: minicpmo-utils
+Version: 0.0.5
+Summary: Unified utilities package for MiniCPM-o: includes cosyvoice + stepaudio2 and extensible utils.
+Author: MiniCPM-o Utils Maintainers
+License: Apache-2.0
+Keywords: minicpmo,audio,tts,utils,cosyvoice,stepaudio2
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+Requires-Dist: numpy
+Requires-Dist: pillow==10.4.0
+Requires-Dist: librosa==0.9.0
+Requires-Dist: decord==0.6.0
+Requires-Dist: moviepy==2.1.2
+Requires-Dist: numba==0.61.2
+Provides-Extra: tts
+Requires-Dist: torch>=2.3.0; extra == "tts"
+Requires-Dist: torchaudio>=2.3.0; extra == "tts"
+Requires-Dist: transformers<4.53.0,>=4.51.0; extra == "tts"
+Requires-Dist: onnxruntime<=1.21.0,>=1.18.0; extra == "tts"
+Requires-Dist: onnx; extra == "tts"
+Requires-Dist: hyperpyyaml; extra == "tts"
+Requires-Dist: openai-whisper==20231117; extra == "tts"
+Requires-Dist: tqdm; extra == "tts"
+Requires-Dist: tiktoken; extra == "tts"
+Requires-Dist: inflect; extra == "tts"
+Requires-Dist: omegaconf>=2.0.6; extra == "tts"
+Requires-Dist: conformer==0.3.2; extra == "tts"
+Requires-Dist: einops==0.8.1; extra == "tts"
+Requires-Dist: hydra-core; extra == "tts"
+Requires-Dist: lightning==2.2.4; extra == "tts"
+Requires-Dist: rich; extra == "tts"
+Requires-Dist: gdown==5.2.0; extra == "tts"
+Requires-Dist: matplotlib; extra == "tts"
+Requires-Dist: wget; extra == "tts"
+Requires-Dist: pyarrow; extra == "tts"
+Requires-Dist: pyworld; extra == "tts"
+Requires-Dist: scipy; extra == "tts"
+Requires-Dist: pyyaml; extra == "tts"
+Requires-Dist: regex; extra == "tts"
+Requires-Dist: soundfile==0.12.1; extra == "tts"
+Requires-Dist: diffusers==0.29.0; extra == "tts"
+Provides-Extra: streaming
+Requires-Dist: minicpmo-utils[tts]; extra == "streaming"
+Provides-Extra: streaming-flash
+Requires-Dist: minicpmo-utils[streaming]; extra == "streaming-flash"
+Requires-Dist: flash-attn>=2.6.0; sys_platform == "linux" and extra == "streaming-flash"
+Requires-Dist: triton>=2.3.0; sys_platform == "linux" and extra == "streaming-flash"
+Requires-Dist: safetensors; extra == "streaming-flash"
+Requires-Dist: pynvml; extra == "streaming-flash"
+Requires-Dist: xxhash; extra == "streaming-flash"
+Provides-Extra: gpu
+Requires-Dist: onnxruntime-gpu<=1.23.2,>=1.18.0; sys_platform == "linux" and extra == "gpu"
+Provides-Extra: all
+Requires-Dist: minicpmo-utils[gpu,streaming,tts]; extra == "all"
+
+## minicpmo-utils
+
+一个统一安装的工具包（一个 PyPI 分发包），把仓库里的 `cosyvoice` 与 `stepaudio2` 一起打进同一个 wheel，并预留 `minicpmo` 作为后续扩展 utils 的统一入口。
+
+### 安装方式
+
+- 从源码本地安装（开发态，可编辑，默认只装公共依赖）：
+```bash
+cd minicpmo-utils
+pip install -e .
+```
+
+- 如果只想安装 cosyvoice 相关依赖（TTS）：
+```bash
+pip install -e .[tts]
+```
+
+- 如果只想安装 stepaudio2 / streaming 相关依赖：
+```bash
+pip install -e .[streaming]
+```
+
+- 同时安装 cosyvoice + stepaudio2 相关依赖：
+```bash
+pip install -e .[tts,streaming]
+```
+
+- 构建并安装 wheel（推荐分发）：
+```bash
+cd minicpmo-utils
+python -m build        # 生成 dist/*.whl
+pip install \"dist/minicpmo_utils-0.1.0-py3-none-any.whl[tts,streaming]\"
+```
+
+### 导入方式
+
+包会暴露以下顶层模块，安装后可直接使用：
+- `import cosyvoice`
+- `import stepaudio2`
+- `import matcha`
+- `import minicpmo`
+
+也支持通过统一入口导入子包：
+```python
+from minicpmo import cosyvoice, stepaudio2, matcha
+```
+
+以及通过统一的 utils 入口使用通用工具函数，例如：
+
+```python
+from minicpmo.utils import get_video_frame_audio_segments
+```
+

minicpmo_utils-0.0.5/README.md

@@ -0,0 +1,53 @@
+## minicpmo-utils
+
+一个统一安装的工具包（一个 PyPI 分发包），把仓库里的 `cosyvoice` 与 `stepaudio2` 一起打进同一个 wheel，并预留 `minicpmo` 作为后续扩展 utils 的统一入口。
+
+### 安装方式
+
+- 从源码本地安装（开发态，可编辑，默认只装公共依赖）：
+```bash
+cd minicpmo-utils
+pip install -e .
+```
+
+- 如果只想安装 cosyvoice 相关依赖（TTS）：
+```bash
+pip install -e .[tts]
+```
+
+- 如果只想安装 stepaudio2 / streaming 相关依赖：
+```bash
+pip install -e .[streaming]
+```
+
+- 同时安装 cosyvoice + stepaudio2 相关依赖：
+```bash
+pip install -e .[tts,streaming]
+```
+
+- 构建并安装 wheel（推荐分发）：
+```bash
+cd minicpmo-utils
+python -m build        # 生成 dist/*.whl
+pip install \"dist/minicpmo_utils-0.1.0-py3-none-any.whl[tts,streaming]\"
+```
+
+### 导入方式
+
+包会暴露以下顶层模块，安装后可直接使用：
+- `import cosyvoice`
+- `import stepaudio2`
+- `import matcha`
+- `import minicpmo`
+
+也支持通过统一入口导入子包：
+```python
+from minicpmo import cosyvoice, stepaudio2, matcha
+```
+
+以及通过统一的 utils 入口使用通用工具函数，例如：
+
+```python
+from minicpmo.utils import get_video_frame_audio_segments
+```
+

minicpmo_utils-0.0.5/pyproject.toml

@@ -0,0 +1,108 @@
+[build-system]
+requires = ["setuptools>=69", "wheel"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "minicpmo-utils"
+version = "0.0.5"
+description = "Unified utilities package for MiniCPM-o: includes cosyvoice + stepaudio2 and extensible utils."
+readme = "README.md"
+requires-python = ">=3.10"
+license = {text = "Apache-2.0"}
+authors = [
+  {name = "MiniCPM-o Utils Maintainers"},
+]
+keywords = ["minicpmo", "audio", "tts", "utils", "cosyvoice", "stepaudio2"]
+classifiers = [
+  "Development Status :: 4 - Beta",
+  "Intended Audience :: Developers",
+  "License :: OSI Approved :: Apache Software License",
+  "Programming Language :: Python :: 3",
+  "Programming Language :: Python :: 3.10",
+  "Programming Language :: Python :: 3.11",
+  "Programming Language :: Python :: 3.12",
+]
+
+# NOTE:
+# - 这是“一个分发包(minicpmo-utils)”同时提供多个顶层 import 包：
+#   - cosyvoice (来自 ../cosyvoice/src)
+#   - matcha    (来自 ../cosyvoice/src)
+#   - stepaudio2(来自 ../stepaudio2/src)
+#   - s3tokenizer (来自 S3Tokenizer-main)
+#   - minicpmo (本项目扩展 utils 的统一入口：from minicpmo.utils import ...)
+dependencies = [
+  "numpy",
+  "pillow==10.4.0",
+  "librosa==0.9.0",
+  "decord==0.6.0",
+  "moviepy==2.1.2",
+  "numba==0.61.2",
+]
+
+[project.optional-dependencies]
+# cosyvoice TTS 相关依赖
+tts = [
+  "torch>=2.3.0",
+  "torchaudio>=2.3.0",
+  "transformers>=4.51.0,<4.53.0",  # 4.52+ 有兼容性问题
+  "onnxruntime>=1.18.0,<=1.21.0",
+  "onnx",
+  "hyperpyyaml",
+  "openai-whisper==20231117",
+  "tqdm",
+  "tiktoken",
+  "inflect",
+  "omegaconf>=2.0.6",
+  "conformer==0.3.2",
+  "einops==0.8.1",
+  "hydra-core",
+  "lightning==2.2.4",
+  "rich",
+  "gdown==5.2.0",
+  "matplotlib",
+  "wget",
+  "pyarrow",
+  "pyworld",
+  # 新增依赖
+  "scipy",
+  "pyyaml",
+  "regex",
+  "soundfile==0.12.1",
+  "diffusers==0.29.0"
+]
+
+# stepaudio2 基础依赖（token2wav 等）
+streaming = [
+  "minicpmo-utils[tts]",  # streaming 依赖 tts
+]
+
+# stepaudio2 Flash 推理引擎依赖（flashcosyvoice.engine 模块需要）
+streaming-flash = [
+  "minicpmo-utils[streaming]",
+  "flash-attn>=2.6.0; sys_platform == 'linux'",
+  "triton>=2.3.0; sys_platform == 'linux'",
+  "safetensors",
+  "pynvml",
+  "xxhash",
+]
+
+# Linux GPU onnxruntime 可以很重，且与环境强相关，保留为可选 extra
+gpu = [
+  "onnxruntime-gpu>=1.18.0,<=1.23.2; sys_platform == 'linux'",
+]
+
+all = [
+  "minicpmo-utils[tts,streaming,gpu]",
+]
+
+[tool.setuptools]
+include-package-data = true
+
+[tool.setuptools.packages.find]
+# 现在所有代码都在本项目的 src/ 下
+where = ["src"]
+
+[tool.setuptools.package-data]
+"cosyvoice.tokenizer.assets" = ["*.tiktoken"]
+"s3tokenizer.assets" = ["*.wav", "*.npz"]
+

minicpmo_utils-0.0.5/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

minicpmo_utils-0.0.5/src/cosyvoice/__init__.py

@@ -0,0 +1,17 @@
+"""
+CosyVoice: Text-to-Speech with Large Language Model
+"""
+
+__version__ = "0.1.0"
+
+# Lazy import to avoid requiring all dependencies at package import time
+def __getattr__(name):
+    if name in ('CosyVoice', 'CosyVoice2'):
+        from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
+        if name == 'CosyVoice':
+            return CosyVoice
+        elif name == 'CosyVoice2':
+            return CosyVoice2
+    raise AttributeError(f"module '{__name__}' has no attribute '{name}'")
+
+__all__ = ['CosyVoice', 'CosyVoice2']

minicpmo_utils-0.0.5/src/cosyvoice/bin/average_model.py

@@ -0,0 +1,93 @@
+# Copyright (c) 2020 Mobvoi Inc (Di Wu)
+# 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 argparse
+import glob
+
+import yaml
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser(description='average model')
+    parser.add_argument('--dst_model', required=True, help='averaged model')
+    parser.add_argument('--src_path',
+                        required=True,
+                        help='src model path for average')
+    parser.add_argument('--val_best',
+                        action="store_true",
+                        help='averaged model')
+    parser.add_argument('--num',
+                        default=5,
+                        type=int,
+                        help='nums for averaged model')
+
+    args = parser.parse_args()
+    print(args)
+    return args
+
+
+def main():
+    args = get_args()
+    val_scores = []
+    if args.val_best:
+        yamls = glob.glob('{}/*.yaml'.format(args.src_path))
+        yamls = [
+            f for f in yamls
+            if not (os.path.basename(f).startswith('train')
+                    or os.path.basename(f).startswith('init'))
+        ]
+        for y in yamls:
+            with open(y, 'r') as f:
+                dic_yaml = yaml.load(f, Loader=yaml.BaseLoader)
+                loss = float(dic_yaml['loss_dict']['loss'])
+                epoch = int(dic_yaml['epoch'])
+                step = int(dic_yaml['step'])
+                tag = dic_yaml['tag']
+                val_scores += [[epoch, step, loss, tag]]
+        sorted_val_scores = sorted(val_scores,
+                                   key=lambda x: x[2],
+                                   reverse=False)
+        print("best val (epoch, step, loss, tag) = " +
+              str(sorted_val_scores[:args.num]))
+        path_list = [
+            args.src_path + '/epoch_{}_whole.pt'.format(score[0])
+            for score in sorted_val_scores[:args.num]
+        ]
+    print(path_list)
+    avg = {}
+    num = args.num
+    assert num == len(path_list)
+    for path in path_list:
+        print('Processing {}'.format(path))
+        states = torch.load(path, map_location=torch.device('cpu'))
+        for k in states.keys():
+            if k not in ['step', 'epoch']:
+                if k not in avg.keys():
+                    avg[k] = states[k].clone()
+                else:
+                    avg[k] += states[k]
+    # average
+    for k in avg.keys():
+        if avg[k] is not None:
+            # pytorch 1.6 use true_divide instead of /=
+            avg[k] = torch.true_divide(avg[k], num)
+    print('Saving to {}'.format(args.dst_model))
+    torch.save(avg, args.dst_model)
+
+
+if __name__ == '__main__':
+    main()

minicpmo_utils-0.0.5/src/cosyvoice/bin/export_jit.py

@@ -0,0 +1,103 @@
+# 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.
+
+from __future__ import print_function
+
+import argparse
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+import os
+import sys
+import torch
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../..'.format(ROOT_DIR))
+sys.path.append('{}/../../third_party/Matcha-TTS'.format(ROOT_DIR))
+from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
+from cosyvoice.utils.file_utils import logging
+
+
+def get_args():
+    parser = argparse.ArgumentParser(description='export your model for deployment')
+    parser.add_argument('--model_dir',
+                        type=str,
+                        default='pretrained_models/CosyVoice-300M',
+                        help='local path')
+    args = parser.parse_args()
+    print(args)
+    return args
+
+
+def get_optimized_script(model, preserved_attrs=[]):
+    script = torch.jit.script(model)
+    if preserved_attrs != []:
+        script = torch.jit.freeze(script, preserved_attrs=preserved_attrs)
+    else:
+        script = torch.jit.freeze(script)
+    script = torch.jit.optimize_for_inference(script)
+    return script
+
+
+def main():
+    args = get_args()
+    logging.basicConfig(level=logging.DEBUG,
+                        format='%(asctime)s %(levelname)s %(message)s')
+
+    torch._C._jit_set_fusion_strategy([('STATIC', 1)])
+    torch._C._jit_set_profiling_mode(False)
+    torch._C._jit_set_profiling_executor(False)
+
+    try:
+        model = CosyVoice(args.model_dir)
+    except Exception:
+        try:
+            model = CosyVoice2(args.model_dir)
+        except Exception:
+            raise TypeError('no valid model_type!')
+
+    if not isinstance(model, CosyVoice2):
+        # 1. export llm text_encoder
+        llm_text_encoder = model.model.llm.text_encoder
+        script = get_optimized_script(llm_text_encoder)
+        script.save('{}/llm.text_encoder.fp32.zip'.format(args.model_dir))
+        script = get_optimized_script(llm_text_encoder.half())
+        script.save('{}/llm.text_encoder.fp16.zip'.format(args.model_dir))
+        logging.info('successfully export llm_text_encoder')
+
+        # 2. export llm llm
+        llm_llm = model.model.llm.llm
+        script = get_optimized_script(llm_llm, ['forward_chunk'])
+        script.save('{}/llm.llm.fp32.zip'.format(args.model_dir))
+        script = get_optimized_script(llm_llm.half(), ['forward_chunk'])
+        script.save('{}/llm.llm.fp16.zip'.format(args.model_dir))
+        logging.info('successfully export llm_llm')
+
+        # 3. export flow encoder
+        flow_encoder = model.model.flow.encoder
+        script = get_optimized_script(flow_encoder)
+        script.save('{}/flow.encoder.fp32.zip'.format(args.model_dir))
+        script = get_optimized_script(flow_encoder.half())
+        script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
+        logging.info('successfully export flow_encoder')
+    else:
+        # 3. export flow encoder
+        flow_encoder = model.model.flow.encoder
+        script = get_optimized_script(flow_encoder)
+        script.save('{}/flow.encoder.fp32.zip'.format(args.model_dir))
+        script = get_optimized_script(flow_encoder.half())
+        script.save('{}/flow.encoder.fp16.zip'.format(args.model_dir))
+        logging.info('successfully export flow_encoder')
+
+
+if __name__ == '__main__':
+    main()

minicpmo_utils-0.0.5/src/cosyvoice/bin/export_onnx.py

@@ -0,0 +1,120 @@
+# Copyright (c) 2024 Antgroup Inc (authors: Zhoubofan, hexisyztem@icloud.com)
+# 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.
+
+from __future__ import print_function
+
+import argparse
+import logging
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+import os
+import sys
+import onnxruntime
+import random
+import torch
+from tqdm import tqdm
+ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
+sys.path.append('{}/../..'.format(ROOT_DIR))
+sys.path.append('{}/../../third_party/Matcha-TTS'.format(ROOT_DIR))
+from cosyvoice.cli.cosyvoice import CosyVoice, CosyVoice2
+from cosyvoice.utils.file_utils import logging
+
+
+def get_dummy_input(batch_size, seq_len, out_channels, device):
+    x = torch.rand((batch_size, out_channels, seq_len), dtype=torch.float32, device=device)
+    mask = torch.ones((batch_size, 1, seq_len), dtype=torch.float32, device=device)
+    mu = torch.rand((batch_size, out_channels, seq_len), dtype=torch.float32, device=device)
+    t = torch.rand((batch_size), dtype=torch.float32, device=device)
+    spks = torch.rand((batch_size, out_channels), dtype=torch.float32, device=device)
+    cond = torch.rand((batch_size, out_channels, seq_len), dtype=torch.float32, device=device)
+    return x, mask, mu, t, spks, cond
+
+
+def get_args():
+    parser = argparse.ArgumentParser(description='export your model for deployment')
+    parser.add_argument('--model_dir',
+                        type=str,
+                        default='pretrained_models/CosyVoice-300M',
+                        help='local path')
+    args = parser.parse_args()
+    print(args)
+    return args
+
+
+@torch.no_grad()
+def main():
+    args = get_args()
+    logging.basicConfig(level=logging.DEBUG,
+                        format='%(asctime)s %(levelname)s %(message)s')
+
+    try:
+        model = CosyVoice(args.model_dir)
+    except Exception:
+        try:
+            model = CosyVoice2(args.model_dir)
+        except Exception:
+            raise TypeError('no valid model_type!')
+
+    # 1. export flow decoder estimator
+    estimator = model.model.flow.decoder.estimator
+    estimator.eval()
+
+    device = model.model.device
+    batch_size, seq_len = 2, 256
+    out_channels = model.model.flow.decoder.estimator.out_channels
+    x, mask, mu, t, spks, cond = get_dummy_input(batch_size, seq_len, out_channels, device)
+    torch.onnx.export(
+        estimator,
+        (x, mask, mu, t, spks, cond),
+        '{}/flow.decoder.estimator.fp32.onnx'.format(args.model_dir),
+        export_params=True,
+        opset_version=18,
+        do_constant_folding=True,
+        input_names=['x', 'mask', 'mu', 't', 'spks', 'cond'],
+        output_names=['estimator_out'],
+        dynamic_axes={
+            'x': {2: 'seq_len'},
+            'mask': {2: 'seq_len'},
+            'mu': {2: 'seq_len'},
+            'cond': {2: 'seq_len'},
+            'estimator_out': {2: 'seq_len'},
+        }
+    )
+
+    # 2. test computation consistency
+    option = onnxruntime.SessionOptions()
+    option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
+    option.intra_op_num_threads = 1
+    providers = ['CUDAExecutionProvider' if torch.cuda.is_available() else 'CPUExecutionProvider']
+    estimator_onnx = onnxruntime.InferenceSession('{}/flow.decoder.estimator.fp32.onnx'.format(args.model_dir),
+                                                  sess_options=option, providers=providers)
+
+    for _ in tqdm(range(10)):
+        x, mask, mu, t, spks, cond = get_dummy_input(batch_size, random.randint(16, 512), out_channels, device)
+        output_pytorch = estimator(x, mask, mu, t, spks, cond)
+        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_onnx = estimator_onnx.run(None, ort_inputs)[0]
+        torch.testing.assert_allclose(output_pytorch, torch.from_numpy(output_onnx).to(device), rtol=1e-2, atol=1e-4)
+    logging.info('successfully export estimator')
+
+
+if __name__ == "__main__":
+    main()