- 修复路径错误

Pass the selected LoRA checkpoint to load_model() on first load so the
model initializes with the correct rank from lora_config.json instead of
always defaulting to r=32.

On subsequent LoRA hot-swaps, detect rank incompatibility and
automatically reload the model with the new checkpoint's config,
preventing tensor shape mismatch errors (fixes #283).

Made-with: Cursor

.gitignore

@@ -2,3 +2,6 @@ launch.json
 __pycache__
 voxcpm.egg-info
 .DS_Store
+./pretrained_models/
+app_local.py
+models/

README.md

@@ -46,7 +46,7 @@ VoxCPM is a **tokenizer-free** Text-to-Speech system that directly generates con
 - 🎙️ **Ultimate Cloning** — Reproduce every vocal nuance: provide both reference audio and its transcript, and the model continues seamlessly from the reference, faithfully preserving every vocal detail — timbre, rhythm, emotion, and style (same as VoxCPM1.5)
 - 🔊 **48kHz High-Quality Audio** — Accepts 16kHz reference audio and directly outputs 48kHz studio-quality audio via AudioVAE V2's asymmetric encode/decode design, with built-in super-resolution — no external upsampler needed
 - 🧠 **Context-Aware Synthesis** — Automatically infers appropriate prosody and expressiveness from text content
-- ⚡ **Real-Time Streaming** — RTF as low as ~0.3 on NVIDIA RTX 4090, and ~0.13  accelerated by [Nano-VLLM](https://github.com/a710128/nanovllm-voxcpm)
+- ⚡ **Real-Time Streaming** — RTF as low as ~0.3 on NVIDIA RTX 4090, and ~0.13 accelerated by [Nano-vLLM](https://github.com/a710128/nanovllm-voxcpm) or [vLLM-Omni](https://github.com/vllm-project/vllm-omni) — official vLLM omni-modal serving for VoxCPM2 with PagedAttention and an OpenAI-compatible API
 - 📜 **Fully Open-Source & Commercial-Ready** — Weights and code released under the [Apache-2.0](LICENSE) license, free for commercial use
 
 
@@ -92,7 +92,7 @@ Chinese Dialect: 四川话, 粤语, 吴语, 东北话, 河南话, 陕西话, 山
 pip install voxcpm
 ```
 
-> **Requirements:** Python ≥ 3.10, PyTorch ≥ 2.5.0, CUDA ≥ 12.0. See [Quick Start Docs](https://voxcpm.readthedocs.io/en/latest/quickstart.html) for details.
+> **Requirements:** Python ≥ 3.10 (<3.13), PyTorch ≥ 2.5.0, CUDA ≥ 12.0. See [Quick Start Docs](https://voxcpm.readthedocs.io/en/latest/quickstart.html) for details.
 
 ### Python API
 
@@ -123,12 +123,12 @@ pip install modelscope
 ```
 
 ```python
-from modelscope.hub.snapshot_download import snapshot_download
+from modelscope import snapshot_download
+snapshot_download("OpenBMB/VoxCPM2", local_dir='./pretrained_models/VoxCPM2') # specify the local directory to save the model
+
 from voxcpm import VoxCPM
 import soundfile as sf
-
+model = VoxCPM.from_pretrained("./pretrained_models/VoxCPM2", load_denoiser=False)
-local_model_dir = snapshot_download("OpenBMB/VoxCPM2")
-model = VoxCPM.from_pretrained(local_model_dir, load_denoiser=False)
 
 wav = model.generate(
     text="VoxCPM2 is the current recommended release for realistic multilingual speech synthesis.",
@@ -239,7 +239,7 @@ voxcpm --help
 ### Web Demo
 
 ```bash
-python app.py   # then open http://localhost:7860
+python app.py --port 8808  # then open in browser: http://localhost:8808
 ```
 
 ### 🚢 Production Deployment (Nano-vLLM)
@@ -262,6 +262,32 @@ server.stop()
 
 > **RTF as low as ~0.13 on NVIDIA RTX 4090** (vs ~0.3 with the standard PyTorch implementation), with support for batched concurrent requests and a FastAPI HTTP server. See the [Nano-vLLM-VoxCPM repo](https://github.com/a710128/nanovllm-voxcpm) for deployment details.
 
+### 🏭 Production Serving (vLLM-Omni)
+
+For production multi-tenant deployments, use [**vLLM-Omni**](https://github.com/vllm-project/vllm-omni) — the official vLLM project's omni-modal extension with native **VoxCPM2** support. PagedAttention KV cache, continuous batching, and a drop-in **OpenAI-compatible** `/v1/audio/speech` endpoint.
+
+```bash
+# Install from source (latest main — vllm-omni is rapidly evolving)
+uv pip install vllm==0.19.0 --torch-backend=auto
+git clone https://github.com/vllm-project/vllm-omni.git && cd vllm-omni
+uv pip install -e .
+```
+
+See the [vLLM-Omni installation guide](https://vllm-omni.readthedocs.io/en/latest/getting_started/installation/) for other platforms (ROCm, XPU, MUSA, NPU) and Docker images.
+
+```bash
+# Launch an OpenAI-compatible TTS server (--omni enables omni-modal serving)
+vllm serve openbmb/VoxCPM2 --omni --port 8000
+
+# Call it from any OpenAI client
+curl http://localhost:8000/v1/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{"model":"openbmb/VoxCPM2","input":"Hello from VoxCPM2 on vLLM-Omni!","voice":"default"}' \
+  --output out.wav
+```
+
+> Built on the upstream vLLM scheduler, with batched concurrent requests, streaming chunk delivery, and multi-GPU deployment out of the box. See the [VoxCPM2 example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/voxcpm2) for full deployment recipes.
+
 > **Full parameter reference, multi-scenario examples, and voice cloning tips →** [Quick Start Guide](https://voxcpm.readthedocs.io/en/latest/quickstart.html) | [Usage Guide](https://voxcpm.readthedocs.io/en/latest/usage_guide.html) | [Cookbook](https://voxcpm.readthedocs.io/en/latest/cookbook.html)
 
 ---
@@ -415,10 +441,54 @@ VoxCPM2 achieves state-of-the-art or comparable results on public zero-shot and
 
 </details>
 
+
+### Internal 30-Language ASR Benchmark
+
+We additionally run an internal multilingual intelligibility benchmark with **30 languages × 500 samples**. ASR transcription is evaluated via **Gemini 3.1 Flash Lite API**.
+
+<details>
+<summary><b>Internal 30-Language ASR Benchmark (click to expand)</b></summary>
+
+| Language | Metric | VoxCPM2 | Fish S2-Pro |
+|---|---:|---:|---:|
+| ar (Arabic) | CER | 1.23% | 0.30% |
+| da (Danish) | WER | 2.70% | 3.52% |
+| de (German) | WER | 0.96% | 0.64% |
+| el (Greek) | WER | 3.17% | 4.61% |
+| en (English) | WER | 0.42% | 1.03% |
+| es (Spanish) | WER | 1.33% | 0.64% |
+| fi (Finnish) | WER | 2.24% | 2.80% |
+| fr (French) | WER | 2.16% | 2.34% |
+| he (Hebrew) | CER | 2.98% | 15.27% |
+| hi (Hindi) | CER | 0.79% | 0.91% |
+| id (Indonesian) | WER | 1.36% | 1.68% |
+| it (Italian) | WER | 1.65% | 1.08% |
+| ja (Japanese) | CER | 2.40% | 1.82% |
+| km (Khmer) | CER | 2.05% | 75.15% |
+| ko (Korean) | CER | 0.95% | 0.29% |
+| lo (Lao) | CER | 1.90% | 87.40% |
+| ms (Malay) | WER | 1.75% | 1.41% |
+| my (Burmese) | CER | 1.42% | 85.27% |
+| nl (Dutch) | WER | 1.25% | 1.68% |
+| no (Norwegian) | WER | 2.49% | 3.76% |
+| pl (Polish) | WER | 1.90% | 1.65% |
+| pt (Portuguese) | WER | 1.48% | 1.49% |
+| ru (Russian) | WER | 0.90% | 0.86% |
+| sv (Swedish) | WER | 2.22% | 2.63% |
+| sw (Swahili) | CER | 1.07% | 2.02% |
+| th (Thai) | CER | 0.94% | 1.92% |
+| tl (Tagalog) | WER | 2.63% | 4.00% |
+| tr (Turkish) | WER | 1.65% | 1.65% |
+| vi (Vietnamese) | WER | 1.56% | 5.56% |
+| zh (Chinese) | CER | 0.92% | 1.02% |
+| Average (30 languages) |  | **1.68%** | - |
+
+</details>
+
 ### InstructTTSEval
 
 <details>
-<summary><b>Instruction-Guided Voice Design Results</b></summary>
+<summary><b>Instruction-Guided Voice Design Results (click to expand)</b></summary>
 
 | Model | InstructTTSEval-ZH | | | InstructTTSEval-EN | | |
 |-------|:---:|:----:|:----:|:----:|:----:|:----:|
@@ -484,11 +554,13 @@ Full documentation: **[voxcpm.readthedocs.io](https://voxcpm.readthedocs.io/en/l
 | Project | Description |
 |---|---|
 | [**Nano-vLLM**](https://github.com/a710128/nanovllm-voxcpm) | High-throughput and Fast GPU serving |
+| [**vLLM-Omni**](https://github.com/vllm-project/vllm-omni) | Official vLLM omni-modal serving for VoxCPM2 — PagedAttention, OpenAI-compatible API |
 | [**VoxCPM.cpp**](https://github.com/bluryar/VoxCPM.cpp) | GGML/GGUF: CPU, CUDA, Vulkan inference |
 | [**VoxCPM-ONNX**](https://github.com/bluryar/VoxCPM-ONNX) | ONNX export for CPU inference |
 | [**VoxCPMANE**](https://github.com/0seba/VoxCPMANE) | Apple Neural Engine backend |
 | [**voxcpm_rs**](https://github.com/madushan1000/voxcpm_rs) | Rust re-implementation |
 | [**ComfyUI-VoxCPM**](https://github.com/wildminder/ComfyUI-VoxCPM) | ComfyUI node-based workflows |
+| [**ComfyUI_RH_VoxCPM**](https://github.com/HM-RunningHub/ComfyUI_RH_VoxCPM) | Feature-complete ComfyUI workflow for VoxCPM 2 with multi-speaker generation, LoRA, and auto-ASR |
 | [**ComfyUI-VoxCPMTTS**](https://github.com/1038lab/ComfyUI-VoxCPMTTS) | ComfyUI TTS extension |
 | [**TTS WebUI**](https://github.com/rsxdalv/tts_webui_extension.vox_cpm) | Browser-based TTS extension |
 

README_zh.md

@@ -46,7 +46,7 @@ VoxCPM 是一个**无离散音频分词器**（Tokenizer-Free）的语音合成
 - 🎙️ **极致克隆** — 提供参考音频及其文本内容，模型接着参考音频进行无缝续写，从而精准还原声音细节特征（与 VoxCPM1.5 一致）
 - 🔊 **48kHz 高质量音频** — 输入 16kHz 参考音频，通过 AudioVAE V2 的非对称编解码设计直接输出 48kHz 高质量音频，内置超分能力
 - 🧠 **语境感知合成** — 根据文本内容自动推断合适的韵律和表现力
-- ⚡ **实时流式合成** — 在 NVIDIA RTX 4090 上 RTF 低至 ~0.3，通过 [Nano-VLLM](https://github.com/a710128/nanovllm-voxcpm) 加速后可达 ~0.13
+- ⚡ **实时流式合成** — 在 NVIDIA RTX 4090 上 RTF 低至 ~0.3，通过 [Nano-vLLM](https://github.com/a710128/nanovllm-voxcpm) 或 [vLLM-Omni](https://github.com/vllm-project/vllm-omni)（官方 vLLM 全模态服务，原生支持 VoxCPM2，提供 PagedAttention 与 OpenAI 兼容 API）加速后可达 ~0.13
 - 📜 **完全开源，商用就绪** — 权重和代码基于 [Apache-2.0](LICENSE) 协议发布，免费商用
 
 <summary><b>🌍 支持的语言（30种）</b></summary>
@@ -91,7 +91,7 @@ VoxCPM 是一个**无离散音频分词器**（Tokenizer-Free）的语音合成
 pip install voxcpm
 ```
 
-> **环境要求：** Python ≥ 3.10，PyTorch ≥ 2.5.0，CUDA ≥ 12.0。详见 [快速开始文档](https://voxcpm.readthedocs.io/zh-cn/latest/quickstart.html)。
+> **环境要求：** Python ≥ 3.10 (<3.13)，PyTorch ≥ 2.5.0，CUDA ≥ 12.0。详见 [快速开始文档](https://voxcpm.readthedocs.io/zh-cn/latest/quickstart.html)。
 
 ### Python API
 
@@ -122,12 +122,12 @@ pip install modelscope
 ```
 
 ```python
-from modelscope.hub.snapshot_download import snapshot_download
+from modelscope import snapshot_download
+snapshot_download("OpenBMB/VoxCPM2", local_dir='./pretrained_models/VoxCPM2') # 指定模型保存的本地路径
+
 from voxcpm import VoxCPM
 import soundfile as sf
-
+model = VoxCPM.from_pretrained('./pretrained_models/VoxCPM2', load_denoiser=False)
-local_model_dir = snapshot_download("OpenBMB/VoxCPM2")
-model = VoxCPM.from_pretrained(local_model_dir, load_denoiser=False)
 
 wav = model.generate(
     text="VoxCPM2 是目前推荐使用的多语言语音合成版本。",
@@ -238,7 +238,7 @@ voxcpm --help
 ### Web Demo
 
 ```bash
-python app.py   # 然后打开 http://localhost:7860
+python app.py --port 8808  # 然后在浏览器打开 http://localhost:8808
 ```
 
 ### 🚢 生产部署（Nano-vLLM）
@@ -261,6 +261,32 @@ server.stop()
 
 > **在 NVIDIA RTX 4090 上 RTF 低至 ~0.13**（标准 PyTorch 实现约 ~0.3），支持批量并发请求和 FastAPI HTTP 服务。详见 [Nano-vLLM-VoxCPM 仓库](https://github.com/a710128/nanovllm-voxcpm)。
 
+### 🏭 生产环境部署（vLLM-Omni）
+
+如需生产级多租户部署，使用 [**vLLM-Omni**](https://github.com/vllm-project/vllm-omni) — 官方 vLLM 项目的全模态扩展，原生支持 **VoxCPM2**。具备 PagedAttention KV 缓存、连续批处理，以及与 OpenAI 完全兼容的 `/v1/audio/speech` 接口。
+
+```bash
+# 从源码安装（最新 main 分支 —— vllm-omni 正在快速迭代）
+uv pip install vllm==0.19.0 --torch-backend=auto
+git clone https://github.com/vllm-project/vllm-omni.git && cd vllm-omni
+uv pip install -e .
+```
+
+其他平台（ROCm、XPU、MUSA、NPU）与 Docker 镜像请参考 [vLLM-Omni 安装文档](https://vllm-omni.readthedocs.io/en/latest/getting_started/installation/)。
+
+```bash
+# 启动 OpenAI 兼容的 TTS 服务（--omni 启用全模态服务）
+vllm serve openbmb/VoxCPM2 --omni --port 8000
+
+# 任意 OpenAI 客户端均可调用
+curl http://localhost:8000/v1/audio/speech \
+  -H "Content-Type: application/json" \
+  -d '{"model":"openbmb/VoxCPM2","input":"你好，欢迎使用 VoxCPM2 on vLLM-Omni！","voice":"default"}' \
+  --output out.wav
+```
+
+> 基于上游 vLLM 调度器构建，开箱即用支持批量并发、流式分块输出和多 GPU 部署。完整示例见 [VoxCPM2 部署样例](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/voxcpm2)。
+
 > **完整参数说明、多场景示例与声音克隆技巧 →** [快速开始指南](https://voxcpm.readthedocs.io/zh-cn/latest/quickstart.html) | [使用指南](https://voxcpm.readthedocs.io/zh-cn/latest/usage_guide.html) | [Cookbook](https://voxcpm.readthedocs.io/zh-cn/latest/cookbook.html)
 
 ---
@@ -414,10 +440,54 @@ VoxCPM2 在公开的零样本和可控 TTS 基准测试中取得了 SOTA 或可
 
 </details>
 
+### Internal 30-Language ASR Benchmark
+
+我们额外进行了内部多语言可懂度评测：**30 语种 × 500 样本**，ASR 转写评估使用 **Gemini 3.1 Flash Lite API**。
+
+<details>
+<summary><b>内部30语种评测集ASR结果（点击展开）</b></summary>
+
+| 语言 | 指标 | VoxCPM2 | Fish S2-Pro |
+|---|---:|---:|---:|
+| ar (阿拉伯语) | CER | 1.23% | 0.30% |
+| da (丹麦语) | WER | 2.70% | 3.52% |
+| de (德语) | WER | 0.96% | 0.64% |
+| el (希腊语) | WER | 3.17% | 4.61% |
+| en (英语) | WER | 0.42% | 1.03% |
+| es (西班牙语) | WER | 1.33% | 0.64% |
+| fi (芬兰语) | WER | 2.24% | 2.80% |
+| fr (法语) | WER | 2.16% | 2.34% |
+| he (希伯来语) | CER | 2.98% | 15.27% |
+| hi (印地语) | CER | 0.79% | 0.91% |
+| id (印尼语) | WER | 1.36% | 1.68% |
+| it (意大利语) | WER | 1.65% | 1.08% |
+| ja (日语) | CER | 2.40% | 1.82% |
+| km (高棉语) | CER | 2.05% | 75.15% |
+| ko (韩语) | CER | 0.95% | 0.29% |
+| lo (老挝语) | CER | 1.90% | 87.40% |
+| ms (马来语) | WER | 1.75% | 1.41% |
+| my (缅甸语) | CER | 1.42% | 85.27% |
+| nl (荷兰语) | WER | 1.25% | 1.68% |
+| no (挪威语) | WER | 2.49% | 3.76% |
+| pl (波兰语) | WER | 1.90% | 1.65% |
+| pt (葡萄牙语) | WER | 1.48% | 1.49% |
+| ru (俄语) | WER | 0.90% | 0.86% |
+| sv (瑞典语) | WER | 2.22% | 2.63% |
+| sw (斯瓦希里语) | CER | 1.07% | 2.02% |
+| th (泰语) | CER | 0.94% | 1.92% |
+| tl (菲律宾语) | WER | 2.63% | 4.00% |
+| tr (土耳其语) | WER | 1.65% | 1.65% |
+| vi (越南语) | WER | 1.56% | 5.56% |
+| zh (中文) | CER | 0.92% | 1.02% |
+| 平均（30 语种） |  | **1.68%** | - |
+
+</details>
+
+
 ### InstructTTSEval
 
 <details>
-<summary><b>指令驱动音色设计结果</b></summary>
+<summary><b>指令驱动音色设计结果 (点击展开)</b></summary>
 
 | Model | InstructTTSEval-ZH | | | InstructTTSEval-EN | | |
 |-------|:---:|:----:|:----:|:----:|:----:|:----:|
@@ -477,11 +547,13 @@ python lora_ft_webui.py   # 然后打开 http://localhost:7860
 | 项目 | 说明 |
 |---|---|
 | [**Nano-vLLM**](https://github.com/a710128/nanovllm-voxcpm) | 高吞吐快速 GPU 推理引擎 |
+| [**vLLM-Omni**](https://github.com/vllm-project/vllm-omni) | 官方 vLLM 全模态服务（原生支持 VoxCPM2）— PagedAttention、OpenAI 兼容 API |
 | [**VoxCPM.cpp**](https://github.com/bluryar/VoxCPM.cpp) | GGML/GGUF：CPU、CUDA、Vulkan 推理 |
 | [**VoxCPM-ONNX**](https://github.com/bluryar/VoxCPM-ONNX) | ONNX 导出，支持 CPU 推理 |
 | [**VoxCPMANE**](https://github.com/0seba/VoxCPMANE) | Apple Neural Engine 后端 |
 | [**voxcpm_rs**](https://github.com/madushan1000/voxcpm_rs) | Rust 重新实现 |
 | [**ComfyUI-VoxCPM**](https://github.com/wildminder/ComfyUI-VoxCPM) | ComfyUI 节点工作流 |
+| [**ComfyUI_RH_VoxCPM**](https://github.com/HM-RunningHub/ComfyUI_RH_VoxCPM) | 面向 VoxCPM 2 的功能更完整的 ComfyUI 工作流，支持多说话人、LoRA 和自动 ASR |
 | [**ComfyUI-VoxCPMTTS**](https://github.com/1038lab/ComfyUI-VoxCPMTTS) | ComfyUI TTS 扩展 |
 | [**TTS WebUI**](https://github.com/rsxdalv/tts_webui_extension.vox_cpm) | 浏览器端 TTS 扩展 |
 

app.py

@@ -1,4 +1,5 @@
 import os
+import re
 import sys
 import logging
 import numpy as np
@@ -9,8 +10,6 @@ from funasr import AutoModel
 from pathlib import Path
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
-if os.environ.get("HF_REPO_ID", "").strip() == "":
-    os.environ["HF_REPO_ID"] = "openbmb/VoxCPM2"
 
 import voxcpm
 
@@ -55,7 +54,7 @@ _EXAMPLES_FOOTER_EN = (
 )
 
 _USAGE_INSTRUCTIONS_ZH = (
-    "**VoxCPM2 — 三种语音生成方式：**\n\n"
+    "**三种语音生成方式：**\n\n"
     "🎨 **声音设计（Voice Design）**  \n"
     "无需参考音频。在 **Control Instruction** 中描述目标音色特征"
     "（性别、年龄、语气、情绪、语速等），VoxCPM2 即可为你从零创造独一无二的声音。\n\n"
@@ -66,6 +65,8 @@ _USAGE_INSTRUCTIONS_ZH = (
     "开启 **极致克隆模式** 并提供参考音频的文字内容（可自动识别）。"
     "模型会将参考音频视为已说出的前文，以**音频续写**的方式完整还原参考音频中的所有声音细节。"
     "注意：该模式与可控克隆模式互斥，将禁用Control Instruction。\n\n"
+    "目前支持的方言包括：\n"
+    "「四川话、粤语、吴语、东北话、河南话、陕西话、山东话、天津话、闽南话」"
 )
 
 _EXAMPLES_FOOTER_ZH = (
@@ -221,11 +222,11 @@ _APP_THEME = gr.themes.Soft(
 # ---------- Model ----------
 
 class VoxCPMDemo:
-    def __init__(self, model_dir: Optional[str] = None) -> None:
+    def __init__(self, model_id: str = "openbmb/VoxCPM2") -> None:
         self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        logger.info(f"Running on device: {self.device}")
+        logger.info(f"运行在设备上: {self.device}")
 
-        self.asr_model_id = "iic/SenseVoiceSmall"
+        self.asr_model_id = "./models/iic/SenseVoiceSmall"
         self.asr_model: Optional[AutoModel] = AutoModel(
             model=self.asr_model_id,
             disable_update=True,
@@ -234,36 +235,13 @@ class VoxCPMDemo:
         )
 
         self.voxcpm_model: Optional[voxcpm.VoxCPM] = None
-        self.explicit_model_dir = model_dir
+        self._model_id = model_id
-
-    def _resolve_model_dir(self) -> str:
-        if self.explicit_model_dir and os.path.isdir(self.explicit_model_dir):
-            return self.explicit_model_dir
-        env_model_dir = os.environ.get("VOXCPM_MODEL_DIR", "").strip()
-        if env_model_dir and os.path.isdir(env_model_dir):
-            return env_model_dir
-        repo_id = os.environ.get("HF_REPO_ID", "").strip()
-        if len(repo_id) > 0:
-            target_dir = os.path.join("models", repo_id.replace("/", "__"))
-            if not os.path.isdir(target_dir):
-                try:
-                    from huggingface_hub import snapshot_download
-                    os.makedirs(target_dir, exist_ok=True)
-                    logger.info(f"Downloading model from HF repo '{repo_id}' to '{target_dir}' ...")
-                    snapshot_download(repo_id=repo_id, local_dir=target_dir, local_dir_use_symlinks=False)
-                except Exception as e:
-                    logger.warning(f"HF download failed: {e}. Falling back to 'models'.")
-                    return "models"
-            return target_dir
-        return "models"
 
     def get_or_load_voxcpm(self) -> voxcpm.VoxCPM:
         if self.voxcpm_model is not None:
             return self.voxcpm_model
-        logger.info("Model not loaded, initializing...")
+        logger.info(f"Loading model: {self._model_id}")
-        model_dir = self._resolve_model_dir()
+        self.voxcpm_model = voxcpm.VoxCPM.from_pretrained(self._model_id, optimize=True,zipenhancer_model_id="./models/iic/speech_zipenhancer_ans_multiloss_16k_base")
-        logger.info(f"Using model dir: {model_dir}")
-        self.voxcpm_model = voxcpm.VoxCPM(voxcpm_model_path=model_dir, optimize=True)
         logger.info("Model loaded successfully.")
         return self.voxcpm_model
 
@@ -315,6 +293,9 @@ class VoxCPMDemo:
             raise ValueError("Please input text to synthesize.")
 
         control = (control_instruction or "").strip()
+        # Strip any parentheses (half-width/full-width) from control text to avoid
+        # breaking the "(control)text" prompt format expected by the model.
+        control = re.sub(r"[()（）]", "", control).strip()
         final_text = f"({control}){text}" if control else text
 
         audio_path = reference_wav_path_input if reference_wav_path_input else None
@@ -507,9 +488,9 @@ def run_demo(
     server_name: str = "0.0.0.0",
     server_port: int = 8808,
     show_error: bool = True,
-    model_dir: Optional[str] = None,
+    model_id: str = "./models/OpenBMB/VoxCPM2",
 ):
-    demo = VoxCPMDemo(model_dir=model_dir)
+    demo = VoxCPMDemo(model_id=model_id)
     interface = create_demo_interface(demo)
     interface.queue(max_size=10, default_concurrency_limit=1).launch(
         server_name=server_name,
@@ -524,7 +505,10 @@ def run_demo(
 if __name__ == "__main__":
     import argparse
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model-dir", type=str, default=None, help="Path to VoxCPM2 checkpoint directory")
+    parser.add_argument(
-    parser.add_argument("--port", type=int, default=8808, help="Server port")
+        "--model-id", type=str, default="./models/OpenBMB/VoxCPM2",
+        help="本地路径或HuggingFace仓库ID（默认：./models/openbmb/VoxCPM2）",
+    )
+    parser.add_argument("--port", type=int, default=8808, help="服务端口")
     args = parser.parse_args()
-    run_demo(model_dir=args.model_dir, server_port=args.port)
+    run_demo(model_id=args.model_id, server_port=args.port)

app_old.py

@@ -1,280 +0,0 @@
-import os
-import sys
-import numpy as np
-import torch
-import gradio as gr
-from typing import Optional, Tuple
-from funasr import AutoModel
-from pathlib import Path
-os.environ["TOKENIZERS_PARALLELISM"] = "false"
-if os.environ.get("HF_REPO_ID", "").strip() == "":
-    os.environ["HF_REPO_ID"] = "openbmb/VoxCPM1.5"
-
-import voxcpm
-
-
-class VoxCPMDemo:
-    def __init__(self) -> None:
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        print(f"🚀 Running on device: {self.device}", file=sys.stderr)
-
-        # ASR model for prompt text recognition
-        self.asr_model_id = "iic/SenseVoiceSmall"
-        self.asr_model: Optional[AutoModel] = AutoModel(
-            model=self.asr_model_id,
-            disable_update=True,
-            log_level='DEBUG',
-            device="cuda:0" if self.device == "cuda" else "cpu",
-        )
-
-        # TTS model (lazy init)
-        self.voxcpm_model: Optional[voxcpm.VoxCPM] = None
-        self.default_local_model_dir = "./models/VoxCPM1.5"
-
-    # ---------- Model helpers ----------
-    def _resolve_model_dir(self) -> str:
-        """
-        Resolve model directory:
-        1) Use local checkpoint directory if exists
-        2) If HF_REPO_ID env is set, download into models/{repo}
-        3) Fallback to 'models'
-        """
-        if os.path.isdir(self.default_local_model_dir):
-            return self.default_local_model_dir
-
-        repo_id = os.environ.get("HF_REPO_ID", "").strip()
-        if len(repo_id) > 0:
-            target_dir = os.path.join("models", repo_id.replace("/", "__"))
-            if not os.path.isdir(target_dir):
-                try:
-                    from huggingface_hub import snapshot_download  # type: ignore
-                    os.makedirs(target_dir, exist_ok=True)
-                    print(f"Downloading model from HF repo '{repo_id}' to '{target_dir}' ...", file=sys.stderr)
-                    snapshot_download(repo_id=repo_id, local_dir=target_dir, local_dir_use_symlinks=False)
-                except Exception as e:
-                    print(f"Warning: HF download failed: {e}. Falling back to 'data'.", file=sys.stderr)
-                    return "models"
-            return target_dir
-        return "models"
-
-    def get_or_load_voxcpm(self) -> voxcpm.VoxCPM:
-        if self.voxcpm_model is not None:
-            return self.voxcpm_model
-        print("Model not loaded, initializing...", file=sys.stderr)
-        model_dir = self._resolve_model_dir()
-        print(f"Using model dir: {model_dir}", file=sys.stderr)
-        self.voxcpm_model = voxcpm.VoxCPM(voxcpm_model_path=model_dir)
-        print("Model loaded successfully.", file=sys.stderr)
-        return self.voxcpm_model
-
-    # ---------- Functional endpoints ----------
-    def prompt_wav_recognition(self, prompt_wav: Optional[str]) -> str:
-        if prompt_wav is None:
-            return ""
-        res = self.asr_model.generate(input=prompt_wav, language="auto", use_itn=True)
-        text = res[0]["text"].split('|>')[-1]
-        return text
-
-    def generate_tts_audio(
-        self,
-        text_input: str,
-        prompt_wav_path_input: Optional[str] = None,
-        prompt_text_input: Optional[str] = None,
-        cfg_value_input: float = 2.0,
-        inference_timesteps_input: int = 10,
-        do_normalize: bool = True,
-        denoise: bool = True,
-    ) -> Tuple[int, np.ndarray]:
-        """
-        Generate speech from text using VoxCPM; optional reference audio for voice style guidance.
-        Returns (sample_rate, waveform_numpy)
-        """
-        current_model = self.get_or_load_voxcpm()
-
-        text = (text_input or "").strip()
-        if len(text) == 0:
-            raise ValueError("Please input text to synthesize.")
-
-        prompt_wav_path = prompt_wav_path_input if prompt_wav_path_input else None
-        prompt_text = prompt_text_input if prompt_text_input else None
-
-        print(f"Generating audio for text: '{text[:60]}...'", file=sys.stderr)
-        wav = current_model.generate(
-            text=text,
-            prompt_text=prompt_text,
-            prompt_wav_path=prompt_wav_path,
-            cfg_value=float(cfg_value_input),
-            inference_timesteps=int(inference_timesteps_input),
-            normalize=do_normalize,
-            denoise=denoise,
-        )
-        return (current_model.tts_model.sample_rate, wav)
-
-
-# ---------- UI Builders ----------
-
-_APP_THEME = gr.themes.Soft(
-    primary_hue="blue",
-    secondary_hue="gray",
-    neutral_hue="slate",
-    font=[gr.themes.GoogleFont("Inter"), "Arial", "sans-serif"],
-)
-
-_CUSTOM_CSS = """
-.logo-container {
-    text-align: center;
-    margin: 0.5rem 0 1rem 0;
-}
-.logo-container img {
-    height: 80px;
-    width: auto;
-    max-width: 200px;
-    display: inline-block;
-}
-/* Bold accordion labels */
-#acc_quick details > summary,
-#acc_tips details > summary {
-    font-weight: 600 !important;
-    font-size: 1.1em !important;
-}
-/* Bold labels for specific checkboxes */
-#chk_denoise label,
-#chk_denoise span,
-#chk_normalize label,
-#chk_normalize span {
-    font-weight: 600;
-}
-"""
-
-
-def create_demo_interface(demo: VoxCPMDemo):
-    """Build the Gradio UI for VoxCPM demo."""
-    gr.set_static_paths(paths=[Path.cwd().absolute()/"assets"])
-
-    with gr.Blocks() as interface:
-        # Header logo
-        gr.HTML('<div class="logo-container"><img src="/gradio_api/file=assets/voxcpm_logo.png" alt="VoxCPM Logo"></div>')
-
-        # Quick Start
-        with gr.Accordion("📋 Quick Start Guide ｜快速入门", open=False, elem_id="acc_quick"):
-            gr.Markdown("""
-            ### How to Use ｜使用说明
-            1. **(Optional) Provide a Voice Prompt** - Upload or record an audio clip to provide the desired voice characteristics for synthesis.  
-               **（可选）提供参考声音** - 上传或录制一段音频，为声音合成提供音色、语调和情感等个性化特征
-            2. **(Optional) Enter prompt text** - If you provided a voice prompt, enter the corresponding transcript here (auto-recognition available).  
-               **（可选项）输入参考文本** - 如果提供了参考语音，请输入其对应的文本内容（支持自动识别）。
-            3. **Enter target text** - Type the text you want the model to speak.  
-               **输入目标文本** - 输入您希望模型朗读的文字内容。
-            4. **Generate Speech** - Click the "Generate" button to create your audio.  
-               **生成语音** - 点击"生成"按钮，即可为您创造出音频。
-            """)
-
-        # Pro Tips
-        with gr.Accordion("💡 Pro Tips ｜使用建议", open=False, elem_id="acc_tips"):
-            gr.Markdown("""
-            ### Prompt Speech Enhancement｜参考语音降噪
-            - **Enable** to remove background noise for a clean voice, with an external ZipEnhancer component. However, this will limit the audio sampling rate to 16kHz, restricting the cloning quality ceiling.  
-              **启用**：通过 ZipEnhancer 组件消除背景噪音，但会将音频采样率限制在16kHz，限制克隆上限。
-            - **Disable** to preserve the original audio's all information, including background atmosphere, and support audio cloning up to 44.1kHz sampling rate.  
-              **禁用**：保留原始音频的全部信息，包括背景环境声，最高支持44.1kHz的音频复刻。
-
-            ### Text Normalization｜文本正则化
-            - **Enable** to process general text with an external WeTextProcessing component.  
-              **启用**：使用 WeTextProcessing 组件，可支持常见文本的正则化处理。
-            - **Disable** to use VoxCPM's native text understanding ability. For example, it supports phonemes input (For Chinese, phonemes are converted using pinyin, {ni3}{hao3}; For English, phonemes are converted using CMUDict, {HH AH0 L OW1}), try it!  
-              **禁用**：将使用 VoxCPM 内置的文本理解能力。如，支持音素输入（如中文转拼音：{ni3}{hao3}；英文转CMUDict：{HH AH0 L OW1}）和公式符号合成，尝试一下！
-
-            ### CFG Value｜CFG 值
-            - **Lower CFG** if the voice prompt sounds strained or expressive, or instability occurs with long text input.  
-              **调低**：如果提示语音听起来不自然或过于夸张，或者长文本输入出现稳定性问题。
-            - **Higher CFG** for better adherence to the prompt speech style or input text, or instability occurs with too short text input.
-              **调高**：为更好地贴合提示音频的风格或输入文本， 或者极短文本输入出现稳定性问题。
-
-            ### Inference Timesteps｜推理时间步
-            - **Lower** for faster synthesis speed.  
-              **调低**：合成速度更快。
-            - **Higher** for better synthesis quality.  
-              **调高**：合成质量更佳。
-            """)
-
-        # Main controls
-        with gr.Row():
-            with gr.Column():
-                prompt_wav = gr.Audio(
-                    sources=["upload", 'microphone'],
-                    type="filepath",
-                    label="Prompt Speech (Optional, or let VoxCPM improvise)",
-                    value="./examples/example.wav",
-                )
-                DoDenoisePromptAudio = gr.Checkbox(
-                    value=False,
-                    label="Prompt Speech Enhancement",
-                    elem_id="chk_denoise",
-                    info="We use ZipEnhancer model to denoise the prompt audio."
-                )
-                with gr.Row():
-                    prompt_text = gr.Textbox(
-                        value="Just by listening a few minutes a day, you'll be able to eliminate negative thoughts by conditioning your mind to be more positive.",
-                        label="Prompt Text",
-                        placeholder="Please enter the prompt text. Automatic recognition is supported, and you can correct the results yourself..."
-                    )
-                run_btn = gr.Button("Generate Speech", variant="primary")
-
-            with gr.Column():
-                cfg_value = gr.Slider(
-                    minimum=1.0,
-                    maximum=3.0,
-                    value=2.0,
-                    step=0.1,
-                    label="CFG Value (Guidance Scale)",
-                    info="Higher values increase adherence to prompt, lower values allow more creativity"
-                )
-                inference_timesteps = gr.Slider(
-                    minimum=4,
-                    maximum=30,
-                    value=10,
-                    step=1,
-                    label="Inference Timesteps",
-                    info="Number of inference timesteps for generation (higher values may improve quality but slower)"
-                )
-                with gr.Row():
-                    text = gr.Textbox(
-                        value="VoxCPM is an innovative end-to-end TTS model from ModelBest, designed to generate highly realistic speech.",
-                        label="Target Text",
-                    )
-                with gr.Row():
-                    DoNormalizeText = gr.Checkbox(
-                        value=False,
-                        label="Text Normalization",
-                        elem_id="chk_normalize",
-                        info="We use wetext library to normalize the input text."
-                    )
-                audio_output = gr.Audio(label="Output Audio")
-
-        # Wiring
-        run_btn.click(
-            fn=demo.generate_tts_audio,
-            inputs=[text, prompt_wav, prompt_text, cfg_value, inference_timesteps, DoNormalizeText, DoDenoisePromptAudio],
-            outputs=[audio_output],
-            show_progress=True,
-            api_name="generate",
-        )
-        prompt_wav.change(fn=demo.prompt_wav_recognition, inputs=[prompt_wav], outputs=[prompt_text])
-
-    return interface
-
-
-def run_demo(server_name: str = "localhost", server_port: int = 7860, show_error: bool = True):
-    demo = VoxCPMDemo()
-    interface = create_demo_interface(demo)
-    interface.queue(max_size=10, default_concurrency_limit=1).launch(
-        server_name=server_name,
-        server_port=server_port,
-        show_error=show_error,
-        theme=_APP_THEME,
-        css=_CUSTOM_CSS,
-    )
-
-
-if __name__ == "__main__":
-    run_demo()

conf/voxcpm_v2/voxcpm_finetune_all.yaml

@@ -2,7 +2,7 @@ pretrained_path: /path/to/VoxCPM2/
 train_manifest: /path/to/train.jsonl
 val_manifest: null
 sample_rate: 16000        # AudioVAE encoder input rate; must match audio_vae_config.sample_rate
-out_sample_rate: 48000    # AudioVAE decoder output rate; only used at inference, not during training
+out_sample_rate: 48000    # AudioVAE decoder output rate; used for TensorBoard audio logging
 batch_size: 2
 grad_accum_steps: 8  # effective batch size = batch_size × grad_accum_steps = 16
 num_workers: 8
@@ -15,6 +15,7 @@ weight_decay: 0.01
 warmup_steps: 100
 max_steps: 1000
 max_batch_tokens: 8192
+max_grad_norm: 1.0        # gradient clipping max norm; 0 = disabled
 save_path: /path/to/checkpoints/finetune_all
 tensorboard: /path/to/logs/finetune_all
 lambdas:

conf/voxcpm_v2/voxcpm_finetune_lora.yaml

@@ -2,7 +2,7 @@ pretrained_path: /path/to/VoxCPM2/
 train_manifest: /path/to/train.jsonl
 val_manifest: null
 sample_rate: 16000        # AudioVAE encoder input rate; must match audio_vae_config.sample_rate
-out_sample_rate: 48000    # AudioVAE decoder output rate; only used at inference, not during training
+out_sample_rate: 48000    # AudioVAE decoder output rate; used for TensorBoard audio logging
 batch_size: 2
 grad_accum_steps: 8  # effective batch size = batch_size × grad_accum_steps = 16
 num_workers: 8
@@ -15,6 +15,7 @@ weight_decay: 0.01
 warmup_steps: 100
 max_steps: 1000
 max_batch_tokens: 8192
+max_grad_norm: 1.0        # gradient clipping max norm; 0 = disabled
 save_path: /path/to/checkpoints/finetune_lora
 tensorboard: /path/to/logs/finetune_lora
 lambdas:

lora_ft_webui.py

@@ -14,8 +14,10 @@ from typing import Optional
 project_root = Path(__file__).parent
 sys.path.insert(0, str(project_root / "src"))
 
-# Default pretrained model path relative to this repo
+# Default pretrained model path: prefer VoxCPM2 if it exists, fallback to VoxCPM1.5
-default_pretrained_path = str(project_root / "models" / "openbmb__VoxCPM1.5")
+_v2_path = project_root / "models" / "openbmb__VoxCPM2"
+_v15_path = project_root / "models" / "openbmb__VoxCPM1.5"
+default_pretrained_path = str(_v2_path if _v2_path.exists() else _v15_path)
 
 from voxcpm.core import VoxCPM
 from voxcpm.model.voxcpm import LoRAConfig
@@ -279,27 +281,48 @@ def run_inference(text, prompt_wav, prompt_text, lora_selection, cfg_scale, step
                     print(f"Warning: Failed to read base_model from LoRA config: {e}", file=sys.stderr)
 
         # 加载模型
+        lora_to_load = lora_selection if lora_selection and lora_selection != "None" else None
         try:
             print(f"Loading base model: {base_model_path}", file=sys.stderr)
-            load_model(base_model_path)
+            load_model(base_model_path, lora_to_load)
-            if lora_selection and lora_selection != "None":
+            if lora_to_load:
-                print(f"Model loaded for LoRA: {lora_selection}", file=sys.stderr)
+                print(f"Model loaded with LoRA: {lora_selection}", file=sys.stderr)
         except Exception as e:
             error_msg = f"Failed to load model from {base_model_path}: {str(e)}"
             print(error_msg, file=sys.stderr)
             return None, error_msg
+        lora_just_loaded = lora_to_load
+    else:
+        lora_just_loaded = None
 
     # Handle LoRA hot-swapping
     assert current_model is not None, "Model must be loaded before inference"
     if lora_selection and lora_selection != "None":
         full_lora_path = os.path.join("lora", lora_selection)
+
+        if lora_just_loaded != lora_selection:
+            new_lora_config, new_base_model = load_lora_config_from_checkpoint(full_lora_path)
+            current_r = current_model.tts_model.lora_config.r if current_model.tts_model.lora_config else None
+            new_r = new_lora_config.r if new_lora_config else None
+
+            if new_r is not None and current_r is not None and new_r != current_r:
+                print(f"LoRA rank mismatch (model r={current_r}, checkpoint r={new_r}), reloading...", file=sys.stderr)
+                reload_base = (
+                    new_base_model if new_base_model and os.path.exists(new_base_model)
+                    else (pretrained_path if pretrained_path and pretrained_path.strip() else default_pretrained_path)
+                )
+                try:
+                    load_model(reload_base, lora_selection)
+                except Exception as e:
+                    return None, f"Failed to reload model for LoRA rank change: {e}"
+            else:
                 print(f"Hot-loading LoRA: {full_lora_path}", file=sys.stderr)
                 try:
                     current_model.load_lora(full_lora_path)
-            current_model.set_lora_enabled(True)
                 except Exception as e:
                     print(f"Error loading LoRA: {e}", file=sys.stderr)
                     return None, f"Error loading LoRA: {e}"
+        current_model.set_lora_enabled(True)
     else:
         print("Disabling LoRA", file=sys.stderr)
         current_model.set_lora_enabled(False)
@@ -368,6 +391,7 @@ def start_training(
     warmup_steps=100,
     max_steps=None,
     sample_rate=44100,
+    max_grad_norm=1.0,
     # LoRA advanced
     enable_lm=True,
     enable_dit=True,
@@ -409,11 +433,25 @@ def start_training(
     # Resolve max_steps default
     resolved_max_steps = int(max_steps) if max_steps not in (None, "", 0) else int(num_iters)
 
+    # Auto-detect out_sample_rate from model config
+    out_sample_rate = 0
+    config_file = os.path.join(pretrained_path, "config.json")
+    if os.path.isfile(config_file):
+        try:
+            with open(config_file, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+            out_sr = cfg.get("audio_vae_config", {}).get("out_sample_rate")
+            if out_sr:
+                out_sample_rate = int(out_sr)
+        except Exception:
+            pass
+
     config = {
         "pretrained_path": pretrained_path,
         "train_manifest": train_manifest,
         "val_manifest": val_manifest,
         "sample_rate": int(sample_rate),
+        "out_sample_rate": out_sample_rate,
         "batch_size": int(batch_size),
         "grad_accum_steps": int(grad_accum_steps),
         "num_workers": int(num_workers),
@@ -425,6 +463,7 @@ def start_training(
         "weight_decay": float(weight_decay),
         "warmup_steps": int(warmup_steps),
         "max_steps": resolved_max_steps,
+        "max_grad_norm": float(max_grad_norm),
         "save_path": checkpoints_dir,
         "tensorboard": tensorboard_path if tensorboard_path else logs_dir,
         "lambdas": {"loss/diff": 1.0, "loss/stop": 1.0},
@@ -932,17 +971,19 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
                         with gr.Row():
                             max_steps = gr.Number(label="最大步数 (max_steps, 0→默认num_iters)", value=0, precision=0)
                             sample_rate = gr.Number(label="采样率 (sample_rate)", value=44100, precision=0)
-                            tensorboard_path = gr.Textbox(label="Tensorboard 路径 (可选)", value="")
+                            max_grad_norm = gr.Number(label="梯度裁剪 (max_grad_norm, 0=关闭)", value=1.0)
                         with gr.Row():
+                            tensorboard_path = gr.Textbox(label="Tensorboard 路径 (可选)", value="")
                             enable_lm = gr.Checkbox(label="启用 LoRA LM (enable_lm)", value=True)
                             enable_dit = gr.Checkbox(label="启用 LoRA DIT (enable_dit)", value=True)
+                        with gr.Row():
                             enable_proj = gr.Checkbox(label="启用投影 (enable_proj)", value=False)
                             dropout = gr.Number(label="LoRA Dropout", value=0.0)
 
                         gr.Markdown("#### 分发选项 (Distribution)")
                         with gr.Row():
                             hf_model_id = gr.Textbox(
-                                label="HuggingFace Model ID (e.g., openbmb/VoxCPM1.5)", value="openbmb/VoxCPM1.5"
+                                label="HuggingFace Model ID (e.g., openbmb/VoxCPM2)", value=""
                             )
                             distribute = gr.Checkbox(label="分发模式 (distribute)", value=False)
 
@@ -992,6 +1033,7 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
                     warmup_steps,
                     max_steps,
                     sample_rate,
+                    max_grad_norm,
                     enable_lm,
                     enable_dit,
                     enable_proj,
@@ -1150,12 +1192,13 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
                 "warmup_steps": "warmup_steps",
                 "max_steps": "最大步数 (max_steps)",
                 "sample_rate": "采样率 (sample_rate)",
+                "max_grad_norm": "梯度裁剪 (max_grad_norm, 0=关闭)",
                 "enable_lm": "启用 LoRA LM (enable_lm)",
                 "enable_dit": "启用 LoRA DIT (enable_dit)",
                 "enable_proj": "启用投影 (enable_proj)",
                 "dropout": "LoRA Dropout",
                 "tensorboard_path": "Tensorboard 路径 (可选)",
-                "hf_model_id": "HuggingFace Model ID (e.g., openbmb/VoxCPM1.5)",
+                "hf_model_id": "HuggingFace Model ID (e.g., openbmb/VoxCPM2)",
                 "distribute": "分发模式 (distribute)",
             }
         else:
@@ -1168,12 +1211,13 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
                 "warmup_steps": "Warmup Steps",
                 "max_steps": "Max Steps",
                 "sample_rate": "Sample Rate",
+                "max_grad_norm": "Max Grad Norm (0=disabled)",
                 "enable_lm": "Enable LoRA LM",
                 "enable_dit": "Enable LoRA DIT",
                 "enable_proj": "Enable Projection",
                 "dropout": "LoRA Dropout",
                 "tensorboard_path": "Tensorboard Path (Optional)",
-                "hf_model_id": "HuggingFace Model ID (e.g., openbmb/VoxCPM1.5)",
+                "hf_model_id": "HuggingFace Model ID (e.g., openbmb/VoxCPM2)",
                 "distribute": "Distribute Mode",
             }
 
@@ -1203,11 +1247,12 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
             gr.update(label=adv["warmup_steps"]),
             gr.update(label=adv["max_steps"]),
             gr.update(label=adv["sample_rate"]),
+            gr.update(label=adv["max_grad_norm"]),
+            gr.update(label=adv["tensorboard_path"]),
             gr.update(label=adv["enable_lm"]),
             gr.update(label=adv["enable_dit"]),
             gr.update(label=adv["enable_proj"]),
             gr.update(label=adv["dropout"]),
-            gr.update(label=adv["tensorboard_path"]),
             # Distribution options
             gr.update(label=adv["hf_model_id"]),
             gr.update(label=adv["distribute"]),
@@ -1254,11 +1299,12 @@ with gr.Blocks(title="VoxCPM LoRA WebUI", theme=gr.themes.Soft(), css=custom_css
             warmup_steps,
             max_steps,
             sample_rate,
+            max_grad_norm,
+            tensorboard_path,
             enable_lm,
             enable_dit,
             enable_proj,
             dropout,
-            tensorboard_path,
             # distribution outputs
             hf_model_id,
             distribute,

modeldown.py

@@ -0,0 +1,34 @@
+"""
+模型下载脚本
+
+"""
+from modelscope import snapshot_download
+
+
+def download(repo_id:str, local_dir:str):
+    """
+   下载模型仓库或单个文件
+
+
+    Args:
+        repo_id (str): 用户名/仓库名，例如 'stabilityai/sdxl-turbo'
+        local_dir (str or Path): 下载文件放置的本地目录路径
+
+    Returns:
+        str: 下载文件的本地路径
+
+    Raises:
+        ValueError: 当 repo_id 格式不正确时
+    """
+    model_dir = snapshot_download(
+        repo_id,
+        repo_type='model',
+        local_dir=f"{local_dir}/{repo_id}",
+    )
+    
+
+
+if __name__ == "__main__":
+    download("OpenBMB/VoxCPM2", "./models")
+    download("iic/SenseVoiceSmall", "./models")
+    download('iic/speech_zipenhancer_ans_multiloss_16k_base',"./models")

pyproject.toml

@@ -47,8 +47,7 @@ dependencies = [
     "funasr",
     "spaces",
     "argbind",
-    "safetensors"
+    "safetensors",
-
 ]
 
 [project.optional-dependencies]

scripts/test_pick_runtime_dtype.py

@@ -0,0 +1,108 @@
+"""Unit checks for pick_runtime_dtype / get_dtype consistency.
+
+Loads src/voxcpm/model/utils.py directly to avoid the heavy voxcpm package
+init. Run with: `python scripts/test_pick_runtime_dtype.py`.
+"""
+import importlib.util
+import os
+import pathlib
+import sys
+
+REPO_ROOT = pathlib.Path(__file__).resolve().parent.parent
+UTILS = str(REPO_ROOT / "src" / "voxcpm" / "model" / "utils.py")
+spec = importlib.util.spec_from_file_location("voxcpm_utils", UTILS)
+utils = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(utils)
+
+_LOW_PRECISION_DTYPES = utils._LOW_PRECISION_DTYPES
+_VALID_DTYPE_OVERRIDES = utils._VALID_DTYPE_OVERRIDES
+get_dtype = utils.get_dtype
+pick_runtime_dtype = utils.pick_runtime_dtype
+
+
+def expect(actual, expected, label):
+    ok = actual == expected
+    mark = "OK " if ok else "FAIL"
+    print(f"[{mark}] {label}: got={actual!r} expected={expected!r}")
+    return ok
+
+
+def expect_raises(fn, exc_type, label):
+    try:
+        fn()
+    except exc_type as e:
+        print(f"[OK ] {label}: raised {exc_type.__name__}: {e}")
+        return True
+    except Exception as e:
+        print(f"[FAIL] {label}: raised {type(e).__name__} not {exc_type.__name__}: {e}")
+        return False
+    print(f"[FAIL] {label}: no exception raised")
+    return False
+
+
+results = []
+
+print("=== override set sanity ===")
+results.append(expect("half" not in _VALID_DTYPE_OVERRIDES, True, "half removed from _VALID_DTYPE_OVERRIDES"))
+results.append(expect("half" not in _LOW_PRECISION_DTYPES, True, "half removed from _LOW_PRECISION_DTYPES"))
+
+print("\n=== every accepted override parses through get_dtype ===")
+for dt in sorted(_VALID_DTYPE_OVERRIDES):
+    try:
+        torch_dtype = get_dtype(dt)
+        print(f"[OK ] get_dtype({dt!r}) -> {torch_dtype}")
+        results.append(True)
+    except Exception as e:
+        print(f"[FAIL] get_dtype({dt!r}) raised: {e}")
+        results.append(False)
+
+print("\n=== pick_runtime_dtype: non-mps is a no-op ===")
+results.append(expect(pick_runtime_dtype("cuda", "bfloat16"), "bfloat16", "cuda/bf16 untouched"))
+results.append(expect(pick_runtime_dtype("cpu", "float16"), "float16", "cpu/fp16 untouched"))
+results.append(expect(pick_runtime_dtype("cuda", "float32"), "float32", "cuda/fp32 untouched"))
+
+print("\n=== pick_runtime_dtype: mps forces fp32 for low-precision ===")
+os.environ.pop("VOXCPM_MPS_DTYPE", None)
+results.append(expect(pick_runtime_dtype("mps", "bfloat16"), "float32", "mps/bf16 -> fp32"))
+results.append(expect(pick_runtime_dtype("mps", "bf16"), "float32", "mps/bf16-alias -> fp32"))
+results.append(expect(pick_runtime_dtype("mps", "float16"), "float32", "mps/fp16 -> fp32"))
+results.append(expect(pick_runtime_dtype("mps", "fp16"), "float32", "mps/fp16-alias -> fp32"))
+results.append(expect(pick_runtime_dtype("mps", "float32"), "float32", "mps/fp32 stays"))
+results.append(expect(pick_runtime_dtype("mps", "fp32"), "fp32", "mps/fp32-alias stays"))
+
+print("\n=== pick_runtime_dtype: VOXCPM_MPS_DTYPE override ===")
+os.environ["VOXCPM_MPS_DTYPE"] = "bfloat16"
+results.append(expect(pick_runtime_dtype("mps", "bfloat16"), "bfloat16", "override bf16 honored"))
+
+os.environ["VOXCPM_MPS_DTYPE"] = "FP16"
+results.append(expect(pick_runtime_dtype("mps", "bfloat16"), "fp16", "override is case-insensitive"))
+
+os.environ["VOXCPM_MPS_DTYPE"] = "  float32  "
+results.append(expect(pick_runtime_dtype("mps", "bfloat16"), "float32", "override is whitespace-trimmed"))
+
+print("\n=== pick_runtime_dtype: 'half' is no longer a valid override ===")
+os.environ["VOXCPM_MPS_DTYPE"] = "half"
+results.append(
+    expect_raises(
+        lambda: pick_runtime_dtype("mps", "bfloat16"),
+        ValueError,
+        "override=half now rejected (was the bug)",
+    )
+)
+
+os.environ["VOXCPM_MPS_DTYPE"] = "garbage"
+results.append(
+    expect_raises(
+        lambda: pick_runtime_dtype("mps", "bfloat16"),
+        ValueError,
+        "override=garbage still rejected",
+    )
+)
+
+os.environ.pop("VOXCPM_MPS_DTYPE", None)
+
+print("\n=== summary ===")
+passed = sum(results)
+total = len(results)
+print(f"{passed}/{total} passed")
+sys.exit(0 if passed == total else 1)

scripts/train_voxcpm_finetune.py

@@ -30,7 +30,8 @@ except ImportError:
 import json
 
 from voxcpm.model import VoxCPMModel, VoxCPM2Model
-from voxcpm.model.voxcpm import LoRAConfig
+from voxcpm.model.voxcpm import LoRAConfig as LoRAConfigV1
+from voxcpm.model.voxcpm2 import LoRAConfig as LoRAConfigV2
 from voxcpm.training import (
     Accelerator,
     BatchProcessor,
@@ -46,7 +47,7 @@ def train(
     train_manifest: str,
     val_manifest: str = "",
     sample_rate: int = 16_000,
-    out_sample_rate: int = 0,  # accepted from YAML for documentation; not used in training
+    out_sample_rate: int = 0,  # AudioVAE decoder output rate; used for TensorBoard audio logging
     batch_size: int = 1,
     grad_accum_steps: int = 1,
     num_workers: int = 2,
@@ -64,12 +65,12 @@ def train(
     lambdas: Dict[str, float] = {"loss/diff": 1.0, "loss/stop": 1.0},
     lora: dict = None,
     config_path: str = "",
+    max_grad_norm: float = 0.0,  # gradient clipping; 0 = disabled (backward compat)
     # Distribution options (for LoRA checkpoints)
     hf_model_id: str = "",  # HuggingFace model ID (e.g., "openbmb/VoxCPM1.5")
     distribute: bool = False,  # If True, save hf_model_id as base_model; otherwise save pretrained_path
 ):
     _ = config_path
-    _ = out_sample_rate
 
     # Validate distribution options
     if lora is not None and distribute and not hf_model_id:
@@ -93,6 +94,7 @@ def train(
     with open(os.path.join(pretrained_path, "config.json"), "r", encoding="utf-8") as _f:
         _arch = json.load(_f).get("architecture", "voxcpm").lower()
     _model_cls = VoxCPM2Model if _arch == "voxcpm2" else VoxCPMModel
+    LoRAConfig = LoRAConfigV2 if _arch == "voxcpm2" else LoRAConfigV1
     if accelerator.rank == 0:
         print(f"Detected architecture: {_arch} -> {_model_cls.__name__}", file=sys.stderr)
     base_model = _model_cls.from_local(
@@ -178,8 +180,12 @@ def train(
         dataset_cnt=dataset_cnt,
         device=accelerator.device,
     )
-    # Save audio_vae for audio generation
+    # Save audio_vae and output sample rate for audio generation.
+    # Prefer model's actual output rate; fall back to YAML out_sample_rate or encode rate.
     audio_vae_for_gen = base_model.audio_vae
+    out_sr = base_model.sample_rate  # decoder output rate (e.g. 48000 for V2)
+    if out_sr == 0 and out_sample_rate > 0:
+        out_sr = out_sample_rate
     del base_model.audio_vae
     model = accelerator.prepare_model(base_model)
     unwrapped_model = accelerator.unwrap(model)
@@ -312,8 +318,8 @@ def train(
             scaler = getattr(accelerator, "scaler", None)
             if scaler is not None:
                 scaler.unscale_(optimizer)
-            # Use large max_norm to only compute grad_norm without actual clipping
+            effective_max_norm = max_grad_norm if max_grad_norm > 0 else 1e9
-            grad_norm = torch.nn.utils.clip_grad_norm_(unwrapped_model.parameters(), max_norm=1e9)
+            grad_norm = torch.nn.utils.clip_grad_norm_(unwrapped_model.parameters(), max_norm=effective_max_norm)
 
             accelerator.step(optimizer)
             accelerator.update()
@@ -341,6 +347,7 @@ def train(
                     val_ds=val_ds,
                     audio_vae=audio_vae_for_gen,
                     sample_rate=sample_rate,
+                    out_sample_rate=out_sr,
                     val_texts=val_texts,
                     tokenizer=tokenizer,
                     valid_interval=valid_interval,
@@ -367,6 +374,7 @@ def validate(
     val_ds=None,
     audio_vae=None,
     sample_rate=22050,
+    out_sample_rate=0,
     val_texts=None,
     tokenizer=None,
     valid_interval=1000,
@@ -432,6 +440,7 @@ def validate(
                 step,
                 accelerator,
                 sample_rate,
+                out_sample_rate=out_sample_rate,
                 val_texts=val_texts,
                 tokenizer=tokenizer,
                 valid_interval=valid_interval,
@@ -534,6 +543,7 @@ def generate_sample_audio(
     step,
     accelerator,
     sample_rate=22050,
+    out_sample_rate=0,
     val_texts=None,
     tokenizer=None,
     pretrained_path=None,
@@ -548,6 +558,10 @@ def generate_sample_audio(
     log(f"[Audio] Starting audio generation for {num_samples} samples at step {step}")
 
     unwrapped_model = accelerator.unwrap(model)
+    # Determine the correct output sample rate for generated audio.
+    # out_sample_rate is the decoder output rate (e.g. 48kHz for V2);
+    # sample_rate is the encoder input rate (e.g. 16kHz for V2).
+    gen_sr = out_sample_rate if out_sample_rate > 0 else sample_rate
 
     for i in range(num_samples):
         sample = val_ds[i]
@@ -604,10 +618,10 @@ def generate_sample_audio(
             gen_audio_np = normalize_audio(gen_audio_np)
 
             tag = f"val_sample_{i}"
-            writer.add_audio(f"{tag}/generated_audio", gen_audio_np, global_step=step, sample_rate=sample_rate)
+            writer.add_audio(f"{tag}/generated_audio", gen_audio_np, global_step=step, sample_rate=gen_sr)
-            log(f"[Audio] Generated audio for sample {i}: duration={len(gen_audio_np)/sample_rate:.2f}s")
+            log(f"[Audio] Generated audio for sample {i}: duration={len(gen_audio_np)/gen_sr:.2f}s")
 
-            # Log reference audio
+            # Log reference audio (at encoder input rate, which is what val_ds provides)
             if ref_audio_np is not None:
                 writer.add_audio(
                     f"{tag}/reference_audio", normalize_audio(ref_audio_np), global_step=step, sample_rate=sample_rate
@@ -615,9 +629,9 @@ def generate_sample_audio(
 
             # Generate mel spectrogram figure
             try:
-                mel_gen = compute_mel_spectrogram(gen_audio_np, sample_rate)
+                mel_gen = compute_mel_spectrogram(gen_audio_np, gen_sr)
                 mel_ref = compute_mel_spectrogram(ref_audio_np, sample_rate) if ref_audio_np is not None else None
-                fig = create_mel_figure(gen_audio_np, mel_gen, sample_rate, step, ref_audio_np, mel_ref)
+                fig = create_mel_figure(gen_audio_np, mel_gen, gen_sr, step, ref_audio_np, mel_ref)
                 writer.add_figure(f"{tag}/mel_spectrogram", fig, global_step=step)
                 log(f"[Audio] Created mel spectrogram figure for sample {i}")
             except Exception as e:

src/voxcpm/cli.py

@@ -11,11 +11,6 @@ import os
 import sys
 from pathlib import Path
 
-import soundfile as sf
-
-from voxcpm.core import VoxCPM
-
-
 DEFAULT_HF_MODEL_ID = "openbmb/VoxCPM2"
 
 # -----------------------------
@@ -173,7 +168,9 @@ def validate_batch_args(args, parser):
 # -----------------------------
 
 
-def load_model(args) -> VoxCPM:
+def load_model(args):
+    from voxcpm.core import VoxCPM
+
     print("Loading VoxCPM model...", file=sys.stderr)
 
     zipenhancer_path = getattr(args, "zipenhancer_path", None) or os.environ.get(
@@ -209,6 +206,7 @@ def load_model(args) -> VoxCPM:
                 zipenhancer_model_path=zipenhancer_path,
                 enable_denoiser=not args.no_denoiser,
                 optimize=not args.no_optimize,
+                device=args.device,
                 lora_config=lora_config,
                 lora_weights_path=lora_weights_path,
             )
@@ -227,6 +225,7 @@ def load_model(args) -> VoxCPM:
             cache_dir=args.cache_dir,
             local_files_only=args.local_files_only,
             optimize=not args.no_optimize,
+            device=args.device,
             lora_config=lora_config,
             lora_weights_path=lora_weights_path,
         )
@@ -264,6 +263,8 @@ def _run_single(args, parser, *, text: str, output: str, prompt_text: str | None
         and (args.prompt_audio is not None or args.reference_audio is not None),
     )
 
+    import soundfile as sf
+
     sf.write(str(output_path), audio_array, model.tts_model.sample_rate)
 
     duration = len(audio_array) / model.tts_model.sample_rate
@@ -286,7 +287,27 @@ def cmd_clone(args, parser):
     )
 
 
+def cmd_validate(args, parser):
+    from voxcpm.training.validate import (
+        print_validation_report,
+        validate_manifest,
+    )
+
+    manifest = str(require_file_exists(args.manifest, parser, "manifest file"))
+    result = validate_manifest(
+        manifest_path=manifest,
+        sample_rate=args.sample_rate,
+        max_samples=args.max_samples,
+        verbose=args.verbose,
+    )
+    print_validation_report(result, manifest)
+    if not result.is_valid:
+        sys.exit(1)
+
+
 def cmd_batch(args, parser):
+    import soundfile as sf
+
     input_file = require_file_exists(args.input, parser, "input file")
     output_dir = Path(args.output_dir)
     output_dir.mkdir(parents=True, exist_ok=True)
@@ -403,6 +424,12 @@ def _add_model_args(parser):
         default=DEFAULT_HF_MODEL_ID,
         help=f"Hugging Face repo id (default: {DEFAULT_HF_MODEL_ID})",
     )
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="auto",
+        help="Runtime device: auto, cpu, mps, cuda, or cuda:N (default: auto)",
+    )
     parser.add_argument(
         "--cache-dir", type=str, help="Cache directory for Hub downloads"
     )
@@ -524,6 +551,30 @@ Examples:
     _add_model_args(batch_parser)
     _add_lora_args(batch_parser)
 
+    # Validate subcommand
+    validate_parser = subparsers.add_parser(
+        "validate",
+        help="Validate a training data manifest (JSONL) before fine-tuning",
+    )
+    validate_parser.add_argument(
+        "--manifest", "-m", required=True, help="Path to JSONL training manifest"
+    )
+    validate_parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16_000,
+        help="Expected audio sample rate in Hz (default: 16000)",
+    )
+    validate_parser.add_argument(
+        "--max-samples",
+        type=int,
+        default=0,
+        help="Maximum number of samples to validate (0 = all, default: 0)",
+    )
+    validate_parser.add_argument(
+        "--verbose", "-v", action="store_true", help="Print per-sample progress"
+    )
+
     # Legacy root arguments
     parser.add_argument("--input", "-i", help="Input text file (batch mode only)")
     parser.add_argument(
@@ -576,6 +627,9 @@ def main():
     parser = _build_parser()
     args = parser.parse_args()
 
+    if args.command == "validate":
+        return cmd_validate(args, parser)
+
     validate_ranges(args, parser)
 
     if args.command == "design":

src/voxcpm/core.py

@@ -8,6 +8,7 @@ from typing import Generator, Optional
 from huggingface_hub import snapshot_download
 from .model.voxcpm import VoxCPMModel, LoRAConfig
 from .model.voxcpm2 import VoxCPM2Model
+from .model.utils import next_and_close
 
 
 class VoxCPM:
@@ -17,6 +18,7 @@ class VoxCPM:
         zipenhancer_model_path: str | None = "iic/speech_zipenhancer_ans_multiloss_16k_base",
         enable_denoiser: bool = True,
         optimize: bool = True,
+        device: str | None = None,
         lora_config: Optional[LoRAConfig] = None,
         lora_weights_path: Optional[str] = None,
     ):
@@ -30,6 +32,9 @@ class VoxCPM:
                 id or local path. If None, denoiser will not be initialized.
             enable_denoiser: Whether to initialize the denoiser pipeline.
             optimize: Whether to optimize the model with torch.compile. True by default, but can be disabled for debugging.
+            device: Runtime device. If set to ``None`` or ``"auto"``, VoxCPM
+                will choose automatically (preferring CUDA, then MPS, then CPU).
+                If set explicitly, that device is used or a clear error is raised.
             lora_config: LoRA configuration for fine-tuning. If lora_weights_path is
                 provided without lora_config, a default config will be created.
             lora_weights_path: Path to pre-trained LoRA weights (.pth file or directory
@@ -56,10 +61,20 @@ class VoxCPM:
         arch = config.get("architecture", "voxcpm").lower()
 
         if arch == "voxcpm2":
-            self.tts_model = VoxCPM2Model.from_local(voxcpm_model_path, optimize=optimize, lora_config=lora_config)
+            self.tts_model = VoxCPM2Model.from_local(
+                voxcpm_model_path,
+                optimize=optimize,
+                device=device,
+                lora_config=lora_config,
+            )
             print("Loaded VoxCPM2Model", file=sys.stderr)
         elif arch == "voxcpm":
-            self.tts_model = VoxCPMModel.from_local(voxcpm_model_path, optimize=optimize, lora_config=lora_config)
+            self.tts_model = VoxCPMModel.from_local(
+                voxcpm_model_path,
+                optimize=optimize,
+                device=device,
+                lora_config=lora_config,
+            )
             print("Loaded VoxCPMModel", file=sys.stderr)
         else:
             raise ValueError(f"Unsupported architecture: {arch}")
@@ -94,6 +109,7 @@ class VoxCPM:
         cache_dir: str = None,
         local_files_only: bool = False,
         optimize: bool = True,
+        device: str | None = None,
         lora_config: Optional[LoRAConfig] = None,
         lora_weights_path: Optional[str] = None,
         **kwargs,
@@ -109,6 +125,9 @@ class VoxCPM:
             cache_dir: Custom cache directory for the snapshot.
             local_files_only: If True, only use local files and do not attempt
                 to download.
+            device: Runtime device. Use ``None``/``"auto"`` for automatic
+                fallback, or an explicit value such as ``"cpu"``, ``"mps"``,
+                ``"cuda"``, or ``"cuda:0"``.
             lora_config: LoRA configuration for fine-tuning. If lora_weights_path is
                 provided without lora_config, a default config will be created with
                 enable_lm=True and enable_dit=True.
@@ -130,7 +149,7 @@ class VoxCPM:
         if not repo_id:
             raise ValueError("You must provide hf_model_id")
 
-        # Load from local path if provided
+        # 从本地路径加载（如果提供）
         if os.path.isdir(repo_id):
             local_path = repo_id
         else:
@@ -146,13 +165,14 @@ class VoxCPM:
             zipenhancer_model_path=zipenhancer_model_id if load_denoiser else None,
             enable_denoiser=load_denoiser,
             optimize=optimize,
+            device=device,
             lora_config=lora_config,
             lora_weights_path=lora_weights_path,
             **kwargs,
         )
 
     def generate(self, *args, **kwargs) -> np.ndarray:
-        return next(self._generate(*args, streaming=False, **kwargs))
+        return next_and_close(self._generate(*args, streaming=False, **kwargs))
 
     def generate_streaming(self, *args, **kwargs) -> Generator[np.ndarray, None, None]:
         return self._generate(*args, streaming=True, **kwargs)
@@ -200,7 +220,7 @@ class VoxCPM:
             Yields audio chunks for each generation step if ``streaming=True``,
             otherwise yields a single array containing the final audio.
         """
-        if not text.strip() or not isinstance(text, str):
+        if not isinstance(text, str) or not text.strip():
             raise ValueError("target text must be a non-empty string")
 
         if prompt_wav_path is not None:
@@ -273,8 +293,15 @@ class VoxCPM:
                 streaming=streaming,
             )
 
+            if streaming:
+                try:
                     for wav, _, _ in generate_result:
                         yield wav.squeeze(0).cpu().numpy()
+                finally:
+                    generate_result.close()
+            else:
+                wav, _, _ = next_and_close(generate_result)
+                yield wav.squeeze(0).cpu().numpy()
 
         finally:
             for tmp_path in temp_files:

src/voxcpm/model/utils.py

@@ -1,7 +1,25 @@
-from typing import List
+import os
+from typing import List, Optional
 import torch
 from transformers import PreTrainedTokenizer
 
+_LOW_PRECISION_DTYPES = {"bfloat16", "bf16", "float16", "fp16"}
+_VALID_DTYPE_OVERRIDES = {
+    "bfloat16", "bf16",
+    "float16", "fp16",
+    "float32", "fp32",
+}
+
+
+# Ref: https://github.com/OpenBMB/VoxCPM/issues/256#issuecomment-4235252732
+# Explicitly close partially-consumed generators so inference_mode cleanup
+# does not get deferred to Python's GC/finalizer path.
+def next_and_close(gen):
+    try:
+        return next(gen)
+    finally:
+        gen.close()
+
 
 def mask_multichar_chinese_tokens(tokenizer: PreTrainedTokenizer):
     """Create a tokenizer wrapper that converts multi-character Chinese tokens to single characters.
@@ -119,3 +137,95 @@ def get_dtype(dtype: str):
         return torch.float32
     else:
         raise ValueError(f"Unsupported dtype: {dtype}")
+
+
+def _has_mps() -> bool:
+    return hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+
+
+def pick_runtime_dtype(device: str, configured_dtype: str) -> str:
+    """Pick a safe runtime dtype for the resolved device.
+
+    On Apple Silicon (MPS), bfloat16/float16 produce enough numerical drift
+    in the diffusion AR loop that the output is glitched and the model's
+    badcase detector triggers infinite retries. float32 is the only stable
+    option today. CUDA and CPU keep whatever the checkpoint was trained with.
+
+    Users can override with ``VOXCPM_MPS_DTYPE`` (e.g. ``bfloat16``) when
+    they want to test future MPS improvements.
+    """
+    if device != "mps":
+        return configured_dtype
+
+    override = os.environ.get("VOXCPM_MPS_DTYPE", "").strip().lower()
+    if override:
+        if override not in _VALID_DTYPE_OVERRIDES:
+            raise ValueError(
+                f"VOXCPM_MPS_DTYPE='{override}' is not one of "
+                f"{sorted(_VALID_DTYPE_OVERRIDES)}"
+            )
+        return override
+
+    if (configured_dtype or "").lower() in _LOW_PRECISION_DTYPES:
+        return "float32"
+    return configured_dtype
+
+
+def auto_select_device(preferred_device: Optional[str] = "cuda") -> str:
+    """
+    Choose a runtime device automatically.
+
+    Preference order:
+    - if the preferred device is available, use it
+    - otherwise fall back to CUDA -> MPS -> CPU
+    """
+    preferred = (preferred_device or "cuda").strip().lower()
+
+    if preferred.startswith("cuda") and torch.cuda.is_available():
+        return preferred
+    if preferred == "mps" and _has_mps():
+        return "mps"
+    if preferred == "cpu":
+        return "cpu"
+
+    if torch.cuda.is_available():
+        return "cuda"
+    if _has_mps():
+        return "mps"
+    return "cpu"
+
+
+def resolve_runtime_device(device: Optional[str], configured_device: str = "cuda") -> str:
+    """
+    Resolve the actual runtime device.
+
+    Semantics:
+    - ``device`` is ``None`` or ``"auto"``: use automatic fallback selection
+    - otherwise: treat it as an explicit user choice and validate availability
+    """
+    explicit = None if device is None else device.strip().lower()
+
+    if explicit is None or explicit == "auto":
+        return auto_select_device(configured_device)
+
+    if explicit.startswith("cuda"):
+        if not torch.cuda.is_available():
+            raise ValueError(
+                f"Requested device '{device}', but CUDA is not available. "
+                "Use device='auto' for automatic fallback."
+            )
+        return explicit
+    if explicit == "mps":
+        if not _has_mps():
+            raise ValueError(
+                "Requested device 'mps', but MPS is not available. "
+                "Use device='auto' for automatic fallback."
+            )
+        return "mps"
+    if explicit == "cpu":
+        return "cpu"
+
+    raise ValueError(
+        f"Unsupported device '{device}'. Supported values are 'auto', 'cpu', 'mps', "
+        "'cuda', or indexed CUDA devices like 'cuda:0'."
+    )

src/voxcpm/model/voxcpm.py

@@ -44,7 +44,13 @@ from ..modules.layers.lora import apply_lora_to_named_linear_modules
 from ..modules.locdit import CfmConfig, UnifiedCFM, VoxCPMLocDiT
 from ..modules.locenc import VoxCPMLocEnc
 from ..modules.minicpm4 import MiniCPM4Config, MiniCPMModel
-from .utils import get_dtype, mask_multichar_chinese_tokens
+from .utils import (
+    get_dtype,
+    mask_multichar_chinese_tokens,
+    next_and_close,
+    pick_runtime_dtype,
+    resolve_runtime_device,
+)
 
 
 class VoxCPMEncoderConfig(BaseModel):
@@ -109,18 +115,22 @@ class VoxCPMModel(nn.Module):
         tokenizer: LlamaTokenizerFast,
         audio_vae: AudioVAE,
         lora_config: LoRAConfig = None,
+        device: str | None = None,
     ):
         super().__init__()
         self.config = config
         self.lora_config = lora_config
         self.feat_dim = config.feat_dim
         self.patch_size = config.patch_size
-        self.device = config.device
+        self.device = resolve_runtime_device(device, config.device)
-        if not torch.cuda.is_available():
+        self.config.device = self.device
-            if torch.backends.mps.is_available():
+        resolved_dtype = pick_runtime_dtype(self.device, self.config.dtype)
-                self.device = "mps"
+        if resolved_dtype != self.config.dtype:
-            else:
+            print(
-                self.device = "cpu"
+                f"[voxcpm] adjusted dtype {self.config.dtype} -> {resolved_dtype} for device {self.device}",
+                file=sys.stderr,
+            )
+            self.config.dtype = resolved_dtype
         print(f"Running on device: {self.device}, dtype: {self.config.dtype}", file=sys.stderr)
 
         # Text-Semantic LM
@@ -227,6 +237,7 @@ class VoxCPMModel(nn.Module):
             self.residual_lm.forward_step = torch.compile(
                 self.residual_lm.forward_step, mode="reduce-overhead", fullgraph=True
             )
+            self._feat_encoder_raw = self.feat_encoder
             self.feat_encoder = torch.compile(self.feat_encoder, mode="reduce-overhead", fullgraph=True)
             self.feat_decoder.estimator = torch.compile(
                 self.feat_decoder.estimator, mode="reduce-overhead", fullgraph=True
@@ -337,7 +348,7 @@ class VoxCPMModel(nn.Module):
         return get_dtype(self.config.dtype)
 
     def generate(self, *args, **kwargs) -> torch.Tensor:
-        return next(self._generate(*args, streaming=False, **kwargs))
+        return next_and_close(self._generate(*args, streaming=False, **kwargs))
 
     def generate_streaming(self, *args, **kwargs) -> Generator[torch.Tensor, None, None]:
         return self._generate(*args, streaming=True, **kwargs)
@@ -463,7 +474,7 @@ class VoxCPMModel(nn.Module):
                     yield decode_audio
                 break
             else:
-                latent_pred, pred_audio_feat = next(inference_result)
+                latent_pred, pred_audio_feat = next_and_close(inference_result)
                 if retry_badcase:
                     if pred_audio_feat.shape[0] >= target_text_length * retry_badcase_ratio_threshold:
                         print(
@@ -571,7 +582,7 @@ class VoxCPMModel(nn.Module):
         return merged_cache
 
     def generate_with_prompt_cache(self, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        return next(self._generate_with_prompt_cache(*args, streaming=False, **kwargs))
+        return next_and_close(self._generate_with_prompt_cache(*args, streaming=False, **kwargs))
 
     def generate_with_prompt_cache_streaming(
         self, *args, **kwargs
@@ -690,7 +701,7 @@ class VoxCPMModel(nn.Module):
                     yield (decode_audio, target_text_token, pred_audio_feat)
                 break
             else:
-                latent_pred, pred_audio_feat = next(inference_result)
+                latent_pred, pred_audio_feat = next_and_close(inference_result)
                 if retry_badcase:
                     if pred_audio_feat.shape[0] >= target_text_length * retry_badcase_ratio_threshold:
                         print(
@@ -713,7 +724,7 @@ class VoxCPMModel(nn.Module):
             yield (decode_audio, target_text_token, pred_audio_feat)
 
     def inference(self, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
-        return next(self._inference(*args, streaming=False, **kwargs))
+        return next_and_close(self._inference(*args, streaming=False, **kwargs))
 
     def inference_streaming(self, *args, **kwargs) -> Generator[Tuple[torch.Tensor, List[torch.Tensor]], None, None]:
         return self._inference(*args, streaming=True, **kwargs)
@@ -755,7 +766,8 @@ class VoxCPMModel(nn.Module):
         """
         B, T, P, D = feat.shape
 
-        feat_embed = self.feat_encoder(feat)  # [b, t, h_feat]
+        prefill_encoder = getattr(self, "_feat_encoder_raw", self.feat_encoder)
+        feat_embed = prefill_encoder(feat)  # [b, t, h_feat]
         feat_embed = self.enc_to_lm_proj(feat_embed)
 
         if self.config.lm_config.use_mup:
@@ -845,8 +857,16 @@ class VoxCPMModel(nn.Module):
             yield feat_pred, pred_feat_seq.squeeze(0).cpu()
 
     @classmethod
-    def from_local(cls, path: str, optimize: bool = True, training: bool = False, lora_config: LoRAConfig = None):
+    def from_local(
-        config = VoxCPMConfig.model_validate_json(open(os.path.join(path, "config.json")).read())
+        cls,
+        path: str,
+        optimize: bool = True,
+        training: bool = False,
+        device: str | None = None,
+        lora_config: LoRAConfig = None,
+    ):
+        with open(os.path.join(path, "config.json"), "r", encoding="utf-8") as _cfg_f:
+            config = VoxCPMConfig.model_validate_json(_cfg_f.read())
         tokenizer = LlamaTokenizerFast.from_pretrained(path)
         audio_vae_config = getattr(config, "audio_vae_config", None)
         audio_vae = AudioVAE(config=audio_vae_config) if audio_vae_config else AudioVAE()
@@ -868,7 +888,7 @@ class VoxCPMModel(nn.Module):
             raise FileNotFoundError(
                 f"AudioVAE checkpoint not found. Expected either {audiovae_safetensors_path} or {audiovae_pth_path}"
             )
-        model = cls(config, tokenizer, audio_vae, lora_config)
+        model = cls(config, tokenizer, audio_vae, lora_config, device=device)
         if not training:
             lm_dtype = get_dtype(model.config.dtype)
             model = model.to(lm_dtype)
@@ -950,7 +970,7 @@ class VoxCPMModel(nn.Module):
         if safetensors_file and safetensors_file.exists() and SAFETENSORS_AVAILABLE:
             state_dict = load_file(str(safetensors_file), device=device)
         elif ckpt_file and ckpt_file.exists():
-            ckpt = torch.load(ckpt_file, map_location=device, weights_only=False)
+            ckpt = torch.load(ckpt_file, map_location=device, weights_only=True)
             state_dict = ckpt.get("state_dict", ckpt)
         else:
             raise FileNotFoundError(f"LoRA checkpoint not found. Expected either {safetensors_file} or {ckpt_file}")

src/voxcpm/model/voxcpm2.py

@@ -45,28 +45,17 @@ from ..modules.layers.lora import apply_lora_to_named_linear_modules
 from ..modules.locdit import CfmConfig, UnifiedCFM, VoxCPMLocDiTV2
 from ..modules.locenc import VoxCPMLocEnc
 from ..modules.minicpm4 import MiniCPM4Config, MiniCPMModel
-from .utils import get_dtype, mask_multichar_chinese_tokens
+from .utils import (
+    get_dtype,
+    mask_multichar_chinese_tokens,
+    next_and_close,
+    pick_runtime_dtype,
+    resolve_runtime_device,
+)
 
 
-def _trim_audio_silence_vad(
+# A simple function to trim audio silence using VAD, not used default
-    audio: torch.Tensor,
+def _trim_audio_silence_vad(audio: torch.Tensor, sample_rate: int, max_silence_ms: float = 200.0, top_db: float = 35.0) -> torch.Tensor:
-    sample_rate: int,
-    max_silence_ms: float = 200.0,
-    top_db: float = 35.0,
-) -> torch.Tensor:
-    """使用能量阈值（VAD 方式）截取首尾静音及尾部长段伪静音，首尾各最多保留 max_silence_ms 毫秒静音。
-
-    会同时截掉末尾的长段伪静音（低能量但非完全静音的段落，如长时间底噪）。
-
-    Args:
-        audio: (1, T) 的音频 tensor
-        sample_rate: 采样率
-        max_silence_ms: 首尾允许保留的最大静音长度（毫秒）
-        top_db: 低于参考电平多少 dB 视为静音
-
-    Returns:
-        截取后的 (1, T') tensor
-    """
     if audio.numel() == 0:
         return audio
     y = audio.squeeze(0).numpy()
@@ -85,7 +74,7 @@ def _trim_audio_silence_vad(
     except Exception:
         start, end = 0, n
 
-    # 用逐帧 RMS 找「最后一段有持续能量的位置」，截掉末尾长伪静音（低能量底噪等）
+    # Find the last frame with continuous energy, trim the long pseudo-silence at the end (low energy background noise, etc.)
     n_frames = max(0, (n - frame_length) // hop_length + 1)
     last_voice_frame = -1
     for j in range(n_frames):
@@ -168,18 +157,22 @@ class VoxCPM2Model(nn.Module):
         tokenizer: LlamaTokenizerFast,
         audio_vae: AudioVAEV2,
         lora_config: LoRAConfig = None,
+        device: str | None = None,
     ):
         super().__init__()
         self.config = config
         self.lora_config = lora_config
         self.feat_dim = config.feat_dim
         self.patch_size = config.patch_size
-        self.device = config.device
+        self.device = resolve_runtime_device(device, config.device)
-        if not torch.cuda.is_available():
+        self.config.device = self.device
-            if torch.backends.mps.is_available():
+        resolved_dtype = pick_runtime_dtype(self.device, self.config.dtype)
-                self.device = "mps"
+        if resolved_dtype != self.config.dtype:
-            else:
+            print(
-                self.device = "cpu"
+                f"[voxcpm2] adjusted dtype {self.config.dtype} -> {resolved_dtype} for device {self.device}",
+                file=sys.stderr,
+            )
+            self.config.dtype = resolved_dtype
         print(f"Running on device: {self.device}, dtype: {self.config.dtype}", file=sys.stderr)
 
         # Text-Semantic LM
@@ -246,6 +239,7 @@ class VoxCPM2Model(nn.Module):
         # Audio VAE
         self.audio_vae = audio_vae
         self.chunk_size = audio_vae.chunk_size
+        self._decode_chunk_size = getattr(audio_vae, "decode_chunk_size", audio_vae.chunk_size)
         self._encode_sample_rate = audio_vae.sample_rate
         self.sample_rate = getattr(audio_vae, "out_sample_rate", audio_vae.sample_rate)
 
@@ -291,6 +285,7 @@ class VoxCPM2Model(nn.Module):
             self.residual_lm.forward_step = torch.compile(
                 self.residual_lm.forward_step, mode="reduce-overhead", fullgraph=True
             )
+            self._feat_encoder_raw = self.feat_encoder
             self.feat_encoder = torch.compile(self.feat_encoder, mode="reduce-overhead", fullgraph=True)
             self.feat_decoder.estimator = torch.compile(
                 self.feat_decoder.estimator, mode="reduce-overhead", fullgraph=True
@@ -382,11 +377,7 @@ class VoxCPM2Model(nn.Module):
                 mu=dit_hidden,
                 patch_size=self.patch_size,
                 cond=feat_cond_for_sample,
-                n_timesteps=(
+                n_timesteps=10,
-                    self.config.dit_config.cfm_config.inference_cfg_rate
-                    if hasattr(self.config.dit_config.cfm_config, "inference_cfg_rate")
-                    else 10
-                ),
             )
             feat_pred = rearrange(feat_pred_seq.transpose(1, 2), "(b t) d p -> b d (t p)", b=B, p=self.patch_size)
 
@@ -463,7 +454,7 @@ class VoxCPM2Model(nn.Module):
         return tokens, feats, t_mask, a_mask
 
     def generate(self, *args, **kwargs) -> torch.Tensor:
-        return next(self._generate(*args, streaming=False, **kwargs))
+        return next_and_close(self._generate(*args, streaming=False, **kwargs))
 
     def generate_streaming(self, *args, **kwargs) -> Generator[torch.Tensor, None, None]:
         return self._generate(*args, streaming=True, **kwargs)
@@ -656,14 +647,14 @@ class VoxCPM2Model(nn.Module):
                 streaming_prefix_len=streaming_prefix_len,
             )
             if streaming:
-                patch_len = self.patch_size * self.chunk_size
+                with self.audio_vae.streaming_decode() as vae_dec:
-                for latent_pred, _ in inference_result:
+                    for latent_pred, _, _ctx in inference_result:
-                    decode_audio = self.audio_vae.decode(latent_pred.to(torch.float32))
+                        decode_audio = vae_dec.decode_chunk(latent_pred.to(torch.float32))
-                    decode_audio = decode_audio[..., -patch_len:].squeeze(1).cpu()
+                        decode_audio = decode_audio.squeeze(1).cpu()
                         yield decode_audio
                 break
             else:
-                latent_pred, pred_audio_feat = next(inference_result)
+                latent_pred, pred_audio_feat, context_len = next_and_close(inference_result)
                 if retry_badcase:
                     if pred_audio_feat.shape[0] >= target_text_length * retry_badcase_ratio_threshold:
                         print(
@@ -679,10 +670,9 @@ class VoxCPM2Model(nn.Module):
 
         if not streaming:
             decode_audio = self.audio_vae.decode(latent_pred.to(torch.float32))
-            patch_len = self.patch_size * self.chunk_size
+            decode_patch_len = self.patch_size * self._decode_chunk_size
-            has_continuation = bool(prompt_wav_path)
+            if context_len > 0:
-            if has_continuation:
+                decode_audio = decode_audio[..., decode_patch_len * context_len:].squeeze(1).cpu()
-                decode_audio = decode_audio[..., patch_len * (streaming_prefix_len - 1):].squeeze(1).cpu()
             else:
                 decode_audio = decode_audio.squeeze(1).cpu()
             yield decode_audio
@@ -782,7 +772,7 @@ class VoxCPM2Model(nn.Module):
         return merged
 
     def generate_with_prompt_cache(self, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
-        return next(self._generate_with_prompt_cache(*args, streaming=False, **kwargs))
+        return next_and_close(self._generate_with_prompt_cache(*args, streaming=False, **kwargs))
 
     def generate_with_prompt_cache_streaming(
         self, *args, **kwargs
@@ -944,14 +934,14 @@ class VoxCPM2Model(nn.Module):
                 streaming_prefix_len=streaming_prefix_len,
             )
             if streaming:
-                patch_len = self.patch_size * self.chunk_size
+                with self.audio_vae.streaming_decode() as vae_dec:
-                for latent_pred, pred_audio_feat in inference_result:
+                    for latent_pred, pred_audio_feat, _ctx in inference_result:
-                    decode_audio = self.audio_vae.decode(latent_pred.to(torch.float32))
+                        decode_audio = vae_dec.decode_chunk(latent_pred.to(torch.float32))
-                    decode_audio = decode_audio[..., -patch_len:].squeeze(1).cpu()
+                        decode_audio = decode_audio.squeeze(1).cpu()
                         yield (decode_audio, target_text_token, pred_audio_feat)
                 break
             else:
-                latent_pred, pred_audio_feat = next(inference_result)
+                latent_pred, pred_audio_feat, context_len = next_and_close(inference_result)
                 if retry_badcase:
                     if pred_audio_feat.shape[0] >= target_text_length * retry_badcase_ratio_threshold:
                         print(
@@ -966,18 +956,20 @@ class VoxCPM2Model(nn.Module):
                     break
         if not streaming:
             decode_audio = self.audio_vae.decode(latent_pred.to(torch.float32))
-            patch_len = self.patch_size * self.chunk_size
+            decode_patch_len = self.patch_size * self._decode_chunk_size
-            if mode in ("continuation", "ref_continuation"):
+            if context_len > 0:
-                decode_audio = decode_audio[..., patch_len * (streaming_prefix_len - 1) :].squeeze(1).cpu()
+                decode_audio = decode_audio[..., decode_patch_len * context_len:].squeeze(1).cpu()
             else:
-                decode_audio = decode_audio[..., :].squeeze(1).cpu()
+                decode_audio = decode_audio.squeeze(1).cpu()
             yield (decode_audio, target_text_token, pred_audio_feat)
 
     def inference(self, *args, **kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
-        return next(self._inference(*args, streaming=False, **kwargs))
+        feat_pred, generated_feat, _ = next_and_close(self._inference(*args, streaming=False, **kwargs))
+        return feat_pred, generated_feat
 
     def inference_streaming(self, *args, **kwargs) -> Generator[Tuple[torch.Tensor, List[torch.Tensor]], None, None]:
-        return self._inference(*args, streaming=True, **kwargs)
+        for feat_pred, pred_feat_seq, _ in self._inference(*args, streaming=True, **kwargs):
+            yield feat_pred, pred_feat_seq
 
     @torch.inference_mode()
     def _inference(
@@ -1016,7 +1008,8 @@ class VoxCPM2Model(nn.Module):
         """
         B, T, P, D = feat.shape
 
-        feat_embed = self.feat_encoder(feat)  # [b, t, h_feat]
+        prefill_encoder = getattr(self, "_feat_encoder_raw", self.feat_encoder)
+        feat_embed = prefill_encoder(feat)  # [b, t, h_feat]
         feat_embed = self.enc_to_lm_proj(feat_embed)
 
         if self.config.lm_config.use_mup:
@@ -1036,6 +1029,7 @@ class VoxCPM2Model(nn.Module):
         #   trailing audio patches as initial context so the VAE can decode smoothly.
         # - Reference-only / zero-shot (feat_mask ends with 0): start from scratch.
         has_continuation_audio = feat_mask[0, -1].item() == 1
+        context_len = 0
         if has_continuation_audio:
             audio_indices = feat_mask.squeeze(0).nonzero(as_tuple=True)[0]
             context_len = min(streaming_prefix_len - 1, len(audio_indices))
@@ -1085,11 +1079,13 @@ class VoxCPM2Model(nn.Module):
             prefix_feat_cond = pred_feat
 
             if streaming:
-                # return the last three predicted latent features to provide enough context for smooth decoding
+                # Yield only the newest patch latent for stateful VAE decode
-                pred_feat_chunk = torch.cat(pred_feat_seq[-streaming_prefix_len:], dim=1)
+                feat_pred = rearrange(pred_feat.unsqueeze(1), "b t p d -> b d (t p)", b=B, p=self.patch_size)
-                feat_pred = rearrange(pred_feat_chunk, "b t p d -> b d (t p)", b=B, p=self.patch_size)
 
-                yield feat_pred, pred_feat_seq
+                yield feat_pred, pred_feat_seq, context_len
+
+                if len(pred_feat_seq) > streaming_prefix_len:
+                    pred_feat_seq = pred_feat_seq[-streaming_prefix_len:]
 
             stop_flag = self.stop_head(self.stop_actn(self.stop_proj(lm_hidden))).argmax(dim=-1)[0].cpu().item()
             if i > min_len and stop_flag == 1:
@@ -1108,11 +1104,20 @@ class VoxCPM2Model(nn.Module):
         if not streaming:
             pred_feat_seq = torch.cat(pred_feat_seq, dim=1)  # b, t, p, d
             feat_pred = rearrange(pred_feat_seq, "b t p d -> b d (t p)", b=B, p=self.patch_size)
-            yield feat_pred, pred_feat_seq.squeeze(0).cpu()
+            generated_feat = pred_feat_seq[:, context_len:, :, :].squeeze(0).cpu()
+            yield feat_pred, generated_feat, context_len
 
     @classmethod
-    def from_local(cls, path: str, optimize: bool = True, training: bool = False, lora_config: LoRAConfig = None):
+    def from_local(
-        config = VoxCPMConfig.model_validate_json(open(os.path.join(path, "config.json")).read())
+        cls,
+        path: str,
+        optimize: bool = True,
+        training: bool = False,
+        device: str | None = None,
+        lora_config: LoRAConfig = None,
+    ):
+        with open(os.path.join(path, "config.json"), "r", encoding="utf-8") as _cfg_f:
+            config = VoxCPMConfig.model_validate_json(_cfg_f.read())
         tokenizer = LlamaTokenizerFast.from_pretrained(path)
         audio_vae_config = getattr(config, "audio_vae_config", None)
         audio_vae = AudioVAEV2(config=audio_vae_config) if audio_vae_config else AudioVAEV2()
@@ -1134,7 +1139,7 @@ class VoxCPM2Model(nn.Module):
             raise FileNotFoundError(
                 f"AudioVAE checkpoint not found. Expected either {audiovae_safetensors_path} or {audiovae_pth_path}"
             )
-        model = cls(config, tokenizer, audio_vae, lora_config)
+        model = cls(config, tokenizer, audio_vae, lora_config, device=device)
         if not training:
             lm_dtype = get_dtype(model.config.dtype)
             model = model.to(lm_dtype)
@@ -1216,7 +1221,7 @@ class VoxCPM2Model(nn.Module):
         if safetensors_file and safetensors_file.exists() and SAFETENSORS_AVAILABLE:
             state_dict = load_file(str(safetensors_file), device=device)
         elif ckpt_file and ckpt_file.exists():
-            ckpt = torch.load(ckpt_file, map_location=device, weights_only=False)
+            ckpt = torch.load(ckpt_file, map_location=device, weights_only=True)
             state_dict = ckpt.get("state_dict", ckpt)
         else:
             raise FileNotFoundError(f"LoRA checkpoint not found. Expected either {safetensors_file} or {ckpt_file}")

src/voxcpm/modules/audiovae/audio_vae_v2.py

@@ -436,6 +436,7 @@ class AudioVAE(nn.Module):
         self.out_sample_rate = out_sample_rate
         self.sr_bin_boundaries = sr_bin_boundaries
         self.chunk_size = math.prod(encoder_rates)
+        self.decode_chunk_size = math.prod(decoder_rates)
 
     def preprocess(self, audio_data, sample_rate):
         if sample_rate is None:
@@ -471,6 +472,20 @@ class AudioVAE(nn.Module):
                 sr_cond = torch.tensor([self.out_sample_rate], device=z.device, dtype=torch.int32)
         return self.decoder(z, sr_cond)
 
+    def streaming_decode(self):
+        """Return a ``StreamingVAEDecoder`` context manager for stateful
+        chunk-by-chunk decoding.  Each call to ``decode_chunk`` processes only
+        the new latent patch and carries causal-conv state internally, avoiding
+        the redundant overlap decode used previously.
+
+        Usage::
+
+            with vae.streaming_decode() as dec:
+                for patch in patches:
+                    audio_chunk = dec.decode_chunk(patch)
+        """
+        return StreamingVAEDecoder(self)
+
     def encode(self, audio_data: torch.Tensor, sample_rate: int):
         """
         Args:
@@ -484,3 +499,82 @@ class AudioVAE(nn.Module):
 
         audio_data = self.preprocess(audio_data, sample_rate)
         return self.encoder(audio_data)["mu"]
+
+
+class StreamingVAEDecoder:
+    """Stateful streaming wrapper for :class:`AudioVAE`.
+
+    Carries causal-convolution padding buffers between calls so that each
+    ``decode_chunk`` processes only the new latent patch — no overlap needed.
+    """
+
+    def __init__(self, vae: AudioVAE):
+        self._vae = vae
+        self._states: dict = {}
+        self._originals: list = []
+
+    # -- context manager --------------------------------------------------
+    def __enter__(self):
+        self._states.clear()
+        self._install()
+        return self
+
+    def __exit__(self, *exc):
+        self._restore()
+        self._states.clear()
+
+    # -- public API --------------------------------------------------------
+    def decode_chunk(self, z_chunk: torch.Tensor) -> torch.Tensor:
+        """Decode a single latent chunk and return the audio waveform."""
+        return self._vae.decode(z_chunk)
+
+    # -- internals ---------------------------------------------------------
+    def _install(self):
+        for name, mod in self._vae.decoder.named_modules():
+            if isinstance(mod, CausalConv1d):
+                pad = mod._CausalConv1d__padding * 2 - mod._CausalConv1d__output_padding
+                if pad > 0:
+                    self._patch_causal_conv(mod, pad)
+            elif isinstance(mod, CausalTransposeConv1d):
+                trim = mod._CausalTransposeConv1d__padding * 2 - mod._CausalTransposeConv1d__output_padding
+                ctx = (mod.kernel_size[0] - 1) // mod.stride[0]
+                if ctx > 0:
+                    self._patch_transpose_conv(mod, ctx, trim)
+
+    def _patch_causal_conv(self, mod, pad_size):
+        states = self._states
+        key = id(mod)
+        orig = mod.forward
+
+        def fwd(x, _k=key, _p=pad_size, _m=mod):
+            x_pad = torch.cat([states[_k], x], dim=-1) if _k in states else F.pad(x, (_p, 0))
+            if x.shape[-1] >= _p:
+                states[_k] = x[:, :, -_p:].detach()
+            else:
+                prev = states.get(_k, torch.zeros(x.shape[0], x.shape[1], _p,
+                                                  device=x.device, dtype=x.dtype))
+                states[_k] = torch.cat([prev, x], dim=-1)[:, :, -_p:].detach()
+            return nn.Conv1d.forward(_m, x_pad)
+
+        mod.forward = fwd
+        self._originals.append((mod, orig))
+
+    def _patch_transpose_conv(self, mod, ctx, trim):
+        states = self._states
+        key = id(mod)
+        orig = mod.forward
+
+        def fwd(x, _k=key, _c=ctx, _t=trim, _m=mod):
+            x_full = torch.cat([states[_k], x], dim=-1) if _k in states else F.pad(x, (_c, 0))
+            states[_k] = x[:, :, -_c:].detach()
+            out = nn.ConvTranspose1d.forward(_m, x_full)
+            left = _c * _m.stride[0]
+            return out[..., left:-_t] if _t > 0 else out[..., left:]
+
+        mod.forward = fwd
+        self._originals.append((mod, orig))
+
+    def _restore(self):
+        for mod, orig in self._originals:
+            mod.forward = orig
+        self._originals.clear()

src/voxcpm/modules/locdit/unified_cfm.py

@@ -225,7 +225,7 @@ class UnifiedCFM(torch.nn.Module):
         losses = F.mse_loss(u_pred, u_tgt.detach(), reduction="none").mean(dim=1)
         if tgt_mask is not None:
             weights = self.adaptive_loss_weighting(losses, tgt_mask.squeeze(1))
-            loss = (weights * losses).sum() / torch.sum(tgt_mask)
+            loss = (weights * losses).sum() / torch.clamp(torch.sum(tgt_mask), min=1.0)
         else:
             loss = losses.mean()
 

src/voxcpm/modules/minicpm4/model.py

@@ -196,7 +196,9 @@ class MiniCPMAttention(nn.Module):
         key_cache[:, :, position_id, :] = key_states
         value_cache[:, :, position_id, :] = value_states
 
-        attn_mask = torch.arange(key_cache.size(2), device=key_cache.device) <= position_id
+        # Use an explicit broadcastable mask shape for SDPA. A 1D mask can
+        # trigger a CPU-side dimension bug in some PyTorch versions.
+        attn_mask = (torch.arange(key_cache.size(2), device=key_cache.device) <= position_id).view(1, 1, 1, -1)
 
         # ref: https://github.com/pytorch/pytorch/issues/163597
         # there is a bug in MPS for non-contiguous tensors, so we need to make them contiguous

src/voxcpm/training/__init__.py

@@ -15,6 +15,7 @@ from .data import (
     BatchProcessor,
 )
 from .state import TrainingState
+from .validate import validate_manifest, ValidationResult
 
 __all__ = [
     "Accelerator",
@@ -24,4 +25,6 @@ __all__ = [
     "TrainingState",
     "load_audio_text_datasets",
     "build_dataloader",
+    "validate_manifest",
+    "ValidationResult",
 ]

src/voxcpm/training/data.py

@@ -12,6 +12,7 @@ from .packers import AudioFeatureProcessingPacker
 
 DEFAULT_TEXT_COLUMN = "text"
 DEFAULT_AUDIO_COLUMN = "audio"
+DEFAULT_REF_AUDIO_COLUMN = "ref_audio"
 DEFAULT_ID_COLUMN = "dataset_id"
 
 
@@ -21,6 +22,7 @@ def load_audio_text_datasets(
     val_manifest: str = "",
     text_column: str = DEFAULT_TEXT_COLUMN,
     audio_column: str = DEFAULT_AUDIO_COLUMN,
+    ref_audio_column: str = DEFAULT_REF_AUDIO_COLUMN,
     dataset_id_column: str = DEFAULT_ID_COLUMN,
     sample_rate: int = 16_000,
     num_proc: int = 1,
@@ -34,14 +36,19 @@ def load_audio_text_datasets(
     def prepare(ds: Dataset) -> Dataset:
         if audio_column not in ds.column_names:
             raise ValueError(f"Expected '{audio_column}' column in manifest.")
-        # We cast to Audio to ensure proper handling during training,
-        # but for length calculation we might need raw path or duration if available.
-        # HF datasets usually don't compute duration automatically for 'Audio' column.
         ds = ds.cast_column(audio_column, Audio(sampling_rate=sample_rate))
         if audio_column != DEFAULT_AUDIO_COLUMN:
             ds = ds.rename_column(audio_column, DEFAULT_AUDIO_COLUMN)
         if text_column != DEFAULT_TEXT_COLUMN:
             ds = ds.rename_column(text_column, DEFAULT_TEXT_COLUMN)
+
+        # ref_audio is optional — cast to Audio if the column exists
+        ref_col = ref_audio_column if ref_audio_column in ds.column_names else DEFAULT_REF_AUDIO_COLUMN
+        if ref_col in ds.column_names:
+            ds = ds.cast_column(ref_col, Audio(sampling_rate=sample_rate))
+            if ref_col != DEFAULT_REF_AUDIO_COLUMN:
+                ds = ds.rename_column(ref_col, DEFAULT_REF_AUDIO_COLUMN)
+
         if dataset_id_column and dataset_id_column in ds.column_names:
             if dataset_id_column != DEFAULT_ID_COLUMN:
                 ds = ds.rename_column(dataset_id_column, DEFAULT_ID_COLUMN)
@@ -67,11 +74,11 @@ def compute_sample_lengths(
     - 音频长度:
         duration(s) * audio_vae_fps -> 近似 VAE 帧数 t_vae
         t_seq = ceil(t_vae / patch_size)
-    - 序列总长约为: text_len + t_seq + 2
+    - 无 ref_audio: text_len + t_seq + 2
+    - 有 ref_audio: text_len + t_seq + ref_seq + 4
 
     Optimized: Use batch column access instead of iterating item by item.
     """
-    # Batch access columns - much faster than per-item access
     text_ids_list = ds["text_ids"]
     text_lens = [len(t) for t in text_ids_list]
 
@@ -79,18 +86,35 @@ def compute_sample_lengths(
     if has_duration:
         durations = ds["duration"]
     else:
-        # Fallback: need to compute from audio (slow, but unavoidable without duration column)
         durations = []
         for i in range(len(ds)):
             audio = ds[i][DEFAULT_AUDIO_COLUMN]
             durations.append(len(audio["array"]) / float(audio["sampling_rate"]))
 
-    # Vectorized length computation
+    has_ref_audio = DEFAULT_REF_AUDIO_COLUMN in ds.column_names
+    if has_ref_audio:
+        ref_duration_col = "ref_duration" if "ref_duration" in ds.column_names else None
+
     lengths = []
-    for text_len, duration in zip(text_lens, durations):
+    for i, (text_len, duration) in enumerate(zip(text_lens, durations)):
         t_vae = math.ceil(float(duration) * audio_vae_fps)
         t_seq = math.ceil(t_vae / patch_size)
-        total_len = text_len + t_seq + 2
+
+        ref_seq = 0
+        if has_ref_audio:
+            # Estimate ref_audio length; ref_audio is None for samples without it
+            if ref_duration_col:
+                ref_dur = ds[i].get(ref_duration_col)
+            else:
+                ref_item = ds[i].get(DEFAULT_REF_AUDIO_COLUMN)
+                ref_dur = len(ref_item["array"]) / float(ref_item["sampling_rate"]) if ref_item else None
+            if ref_dur is not None and float(ref_dur) > 0:
+                ref_vae = math.ceil(float(ref_dur) * audio_vae_fps)
+                ref_seq = math.ceil(ref_vae / patch_size)
+
+        # +2 for 101/102; +2 more for 103/104 when ref_audio present
+        overhead = 4 if ref_seq > 0 else 2
+        total_len = text_len + t_seq + ref_seq + overhead
         lengths.append(total_len)
 
     return lengths
@@ -102,8 +126,11 @@ class HFVoxCPMDataset(TorchDataset):
     PyTorch-friendly samples.
     """
 
+    _SENTINEL = [-100.0]
+
     def __init__(self, dataset: Dataset):
         self.dataset = dataset
+        self.has_ref_audio = DEFAULT_REF_AUDIO_COLUMN in dataset.column_names
 
     def __len__(self):
         return len(self.dataset)
@@ -111,13 +138,17 @@ class HFVoxCPMDataset(TorchDataset):
     def __getitem__(self, idx: int):
         item = self.dataset[idx]
         audio = item[DEFAULT_AUDIO_COLUMN]
-        return {
+        sample = {
             "text_ids": item["text_ids"],
             "audio_array": audio["array"],
             "audio_sampling_rate": audio["sampling_rate"],
             "dataset_id": item.get(DEFAULT_ID_COLUMN, 0),
             "is_prompt": item.get("is_prompt", False),
         }
+        if self.has_ref_audio:
+            ref = item.get(DEFAULT_REF_AUDIO_COLUMN)
+            sample["ref_audio_array"] = ref["array"] if ref else self._SENTINEL
+        return sample
 
     @staticmethod
     def pad_sequences(seqs: List[torch.Tensor], pad_value: float):
@@ -143,7 +174,7 @@ class HFVoxCPMDataset(TorchDataset):
         audio_padded = cls.pad_sequences(audio_tensors, pad_value=-100.0)
         task_ids = torch.ones(text_padded.size(0), dtype=torch.int32)
 
-        return {
+        result = {
             "text_tokens": text_padded,
             "audio_tokens": audio_padded,
             "task_ids": task_ids,
@@ -151,6 +182,12 @@ class HFVoxCPMDataset(TorchDataset):
             "is_prompts": is_prompts,
         }
 
+        if "ref_audio_array" in batch[0]:
+            ref_tensors = [torch.tensor(s["ref_audio_array"], dtype=torch.float32) for s in batch]
+            result["ref_audio_tokens"] = cls.pad_sequences(ref_tensors, pad_value=-100.0)
+
+        return result
+
 
 class BatchProcessor:
     """
@@ -184,12 +221,17 @@ class BatchProcessor:
         task_ids = batch["task_ids"].to(self.device)
         dataset_ids = batch["dataset_ids"].to(self.device)
 
+        ref_audio_tokens = None
+        if "ref_audio_tokens" in batch:
+            ref_audio_tokens = batch["ref_audio_tokens"].to(self.device)
+
         packed = self.packer(
             audio_tokens=audio_tokens,
             text_tokens=text_tokens,
             task_ids=task_ids,
             dataset_ids=dataset_ids,
             is_prompts=batch["is_prompts"],
+            ref_audio_tokens=ref_audio_tokens,
         )
         return packed
 

src/voxcpm/training/packers.py

@@ -1,4 +1,4 @@
-from typing import Dict, List
+from typing import Dict, List, Optional
 
 import torch
 import torch.nn as nn
@@ -14,7 +14,6 @@ class AudioFeatureProcessingPacker:
     def __init__(self, dataset_cnt: int, max_len: int, patch_size: int, feat_dim: int, audio_vae: nn.Module):
         self.audio_start_id = 101
         self.audio_end_id = 102
-        # unused now
         self.audio_prompt_start_id = 103
         self.audio_prompt_end_id = 104
         self.text_eos_token_id = 2
@@ -78,11 +77,16 @@ class AudioFeatureProcessingPacker:
         task_ids: torch.Tensor,
         dataset_ids: torch.Tensor,
         is_prompts: List[bool],
+        ref_audio_tokens: Optional[torch.Tensor] = None,
     ) -> Dict[str, torch.Tensor]:
         """
         Padding-based batching: each sample in the input batch is processed
         independently and then padded to a common length (capped by ``max_len``).
         The result tensors all have shape [B, T, ...].
+
+        If ``ref_audio_tokens`` is provided (same batch dim as ``audio_tokens``),
+        samples whose unpadded ref_audio length > 0 will be processed with the
+        reference-audio path (tokens 103/104 prepended, loss only on target audio).
         """
         device = audio_tokens.device
         max_dataset_id = int(dataset_ids.max().item()) if dataset_ids.numel() > 0 else -1
@@ -101,13 +105,33 @@ class AudioFeatureProcessingPacker:
         audio_duration_consumed = torch.zeros(dataset_cnt, dtype=torch.float32, device=device)
         text_token_consumed = torch.zeros(dataset_cnt, dtype=torch.float32, device=device)
 
-        for audio_token, text_token, task_id, dataset_idx, is_prompt in zip(
+        ref_iter = ref_audio_tokens if ref_audio_tokens is not None else [None] * audio_tokens.size(0)
-            audio_tokens, text_tokens, task_ids.tolist(), dataset_ids.tolist(), is_prompts
+
+        for audio_token, text_token, task_id, dataset_idx, is_prompt, ref_token in zip(
+            audio_tokens, text_tokens, task_ids.tolist(), dataset_ids.tolist(), is_prompts, ref_iter
         ):
             unpad_audio_token = self.unpad_audio_tokens(audio_token).to(torch.float32)
             unpad_text_token = self.unpad_text_tokens(text_token)
             usage = self.id_to_task[task_id]
 
+            has_ref = False
+            if ref_token is not None:
+                unpad_ref_token = self.unpad_audio_tokens(ref_token).to(torch.float32)
+                if unpad_ref_token.numel() > 0:
+                    has_ref = True
+
+            if has_ref:
+                (
+                    packed_text,
+                    audio_feat,
+                    text_mask,
+                    audio_mask,
+                    loss_mask,
+                    labels,
+                    audio_duration,
+                    text_token_count,
+                ) = self.process_tts_data_with_ref(unpad_ref_token, unpad_audio_token, unpad_text_token)
+            else:
                 (
                     packed_text,
                     audio_feat,
@@ -294,3 +318,89 @@ class AudioFeatureProcessingPacker:
             audio_duration,
             text_token_count,
         )
+
+    def process_tts_data_with_ref(
+        self,
+        ref_audio_token: torch.Tensor,
+        target_audio_token: torch.Tensor,
+        text_token: torch.Tensor,
+    ):
+        """
+        Build a training sequence with reference audio prepended:
+
+            [103, ref_feats, 104, text, 101, target_feats, 102]
+
+        Loss is computed only on the target audio segment.
+        """
+        device = text_token.device
+        txt_len = len(text_token)
+
+        ref_feats, ref_duration = self.extract_audio_feats(ref_audio_token)
+        ref_feats = ref_feats.squeeze(0)  # [R, P, D]
+        ref_len = ref_feats.shape[0]
+
+        tgt_feats, tgt_duration = self.extract_audio_feats(target_audio_token)
+        tgt_feats = tgt_feats.squeeze(0)  # [A, P, D]
+        tgt_len = tgt_feats.shape[0]
+
+        feat_shape = (self.patch_size, ref_feats.size(-1))
+
+        def _tok(ids):
+            return torch.tensor(ids, dtype=torch.int32, device=device)
+
+        # -- text token track --
+        # [103, 0×R, 104, text_ids, 101, 0×A, 102]
+        text_token_info = torch.cat([
+            _tok([self.audio_prompt_start_id]),
+            torch.zeros(ref_len, dtype=torch.int32, device=device),
+            _tok([self.audio_prompt_end_id]),
+            text_token,
+            _tok([self.audio_start_id]),
+            torch.zeros(tgt_len, dtype=torch.int32, device=device),
+            _tok([self.audio_end_id]),
+        ])
+
+        # -- audio feature track --
+        zero_1 = torch.zeros((1,) + feat_shape, dtype=torch.float32, device=device)
+        zero_txt = torch.zeros((txt_len,) + feat_shape, dtype=torch.float32, device=device)
+        audio_feat_info = torch.cat([
+            zero_1, ref_feats, zero_1,      # 103, ref, 104
+            zero_txt,                        # text
+            zero_1, tgt_feats, zero_1,       # 101, target, 102
+        ], dim=0)
+
+        # -- masks --
+        text_mask = torch.cat([
+            torch.ones(1), torch.zeros(ref_len), torch.ones(1),
+            torch.ones(txt_len),
+            torch.ones(1), torch.zeros(tgt_len), torch.ones(1),
+        ]).to(torch.int32).to(device)
+
+        audio_mask = torch.cat([
+            torch.zeros(1), torch.ones(ref_len), torch.zeros(1),
+            torch.zeros(txt_len),
+            torch.zeros(1), torch.ones(tgt_len), torch.zeros(1),
+        ]).to(torch.int32).to(device)
+
+        loss_mask = torch.cat([
+            torch.zeros(1 + ref_len + 1),   # ref part: no loss
+            torch.zeros(txt_len),            # text: no loss
+            torch.zeros(1),                  # 101: no loss
+            torch.ones(tgt_len),             # target audio: LOSS
+            torch.zeros(1),                  # 102: no loss
+        ]).to(torch.int32).to(device)
+
+        total_len = 1 + ref_len + 1 + txt_len + 1 + tgt_len + 1
+        labels = torch.zeros(total_len, dtype=torch.int32, device=device)
+        labels[-2] = 1  # stop label at last target audio position
+
+        return (
+            text_token_info,
+            audio_feat_info,
+            text_mask,
+            audio_mask,
+            loss_mask,
+            labels,
+            ref_duration + tgt_duration,
+            txt_len,
+        )

src/voxcpm/training/validate.py

@@ -0,0 +1,309 @@
+"""
+Pre-flight validation for VoxCPM training data manifests.
+
+Validates JSONL manifest files before starting expensive fine-tuning jobs,
+catching format issues, missing files, and data quality problems early.
+"""
+
+import json
+import os
+import sys
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import List, Optional
+
+
+@dataclass
+class ValidationResult:
+    """Structured result of a manifest validation run."""
+
+    total_samples: int = 0
+    valid_samples: int = 0
+    errors: List[str] = field(default_factory=list)
+    warnings: List[str] = field(default_factory=list)
+    audio_durations: List[float] = field(default_factory=list)
+    text_lengths: List[int] = field(default_factory=list)
+    has_ref_audio: int = 0
+
+    @property
+    def is_valid(self) -> bool:
+        return len(self.errors) == 0 and self.valid_samples > 0
+
+
+def _check_audio_file(audio_path: str, sample_rate: int) -> Optional[str]:
+    """Check if an audio file exists, is readable, and matches expected sample rate.
+
+    Returns an error message, or None if the file is valid.
+    """
+    if not os.path.isfile(audio_path):
+        return f"Audio file not found: {audio_path}"
+    try:
+        import soundfile as sf
+
+        info = sf.info(audio_path)
+        if info.frames == 0:
+            return f"Audio file is empty: {audio_path}"
+        if info.samplerate != sample_rate:
+            return (
+                f"Sample rate mismatch in {audio_path}: "
+                f"expected {sample_rate} Hz, got {info.samplerate} Hz"
+            )
+        return None
+    except ImportError:
+        # soundfile not available; just check existence
+        return None
+    except Exception as e:
+        return f"Cannot read audio file {audio_path}: {e}"
+
+
+def _get_audio_duration(audio_path: str) -> Optional[float]:
+    """Get audio duration in seconds. Returns None if unavailable."""
+    try:
+        import soundfile as sf
+
+        info = sf.info(audio_path)
+        return info.duration
+    except Exception:
+        return None
+
+
+def validate_manifest(
+    manifest_path: str,
+    sample_rate: int = 16_000,
+    max_samples: int = 0,
+    verbose: bool = False,
+) -> ValidationResult:
+    """Validate a JSONL training manifest file.
+
+    Checks:
+        1. File exists and is readable
+        2. Each line is valid JSON
+        3. Required columns present (text, audio)
+        4. Audio files exist and are readable
+        5. Text content is non-empty
+        6. Collects duration and text length statistics
+        7. Validates optional ref_audio column
+
+    Args:
+        manifest_path: Path to the JSONL manifest file.
+        sample_rate: Expected audio sample rate (for informational purposes).
+        max_samples: Maximum number of samples to validate (0 = all).
+        verbose: Print per-sample progress.
+
+    Returns:
+        ValidationResult with errors, warnings, and statistics.
+    """
+    result = ValidationResult()
+    path = Path(manifest_path)
+
+    if not path.exists():
+        result.errors.append(f"Manifest file not found: {manifest_path}")
+        return result
+
+    if not path.is_file():
+        result.errors.append(f"Manifest path is not a file: {manifest_path}")
+        return result
+
+    manifest_dir = path.parent
+
+    try:
+        with open(path, "r", encoding="utf-8") as f:
+            lines = f.readlines()
+    except Exception as e:
+        result.errors.append(f"Cannot read manifest file: {e}")
+        return result
+
+    if not lines:
+        result.errors.append("Manifest file is empty")
+        return result
+
+    samples_to_check = len(lines)
+    if max_samples > 0:
+        samples_to_check = min(samples_to_check, max_samples)
+
+    missing_audio_count = 0
+    empty_text_count = 0
+
+    for i, line in enumerate(lines[:samples_to_check]):
+        line = line.strip()
+        if not line:
+            continue
+
+        result.total_samples += 1
+
+        # Check JSON validity
+        try:
+            entry = json.loads(line)
+        except json.JSONDecodeError as e:
+            result.errors.append(f"Line {i + 1}: Invalid JSON — {e}")
+            continue
+
+        if not isinstance(entry, dict):
+            result.errors.append(f"Line {i + 1}: Expected JSON object, got {type(entry).__name__}")
+            continue
+
+        # Check required columns
+        has_error = False
+
+        if "text" not in entry:
+            result.errors.append(f"Line {i + 1}: Missing required column 'text'")
+            has_error = True
+
+        if "audio" not in entry:
+            result.errors.append(f"Line {i + 1}: Missing required column 'audio'")
+            has_error = True
+
+        if has_error:
+            continue
+
+        # Validate text
+        text = entry["text"]
+        if not isinstance(text, str) or not text.strip():
+            empty_text_count += 1
+            if empty_text_count <= 5:
+                result.warnings.append(f"Line {i + 1}: Empty or non-string text")
+        else:
+            result.text_lengths.append(len(text))
+
+        # Validate audio path
+        audio_path = entry["audio"]
+        if isinstance(audio_path, dict):
+            # HuggingFace Audio format with {"path": ..., "array": ...}
+            audio_path = audio_path.get("path", "")
+
+        if isinstance(audio_path, str) and audio_path:
+            # Resolve relative paths against manifest directory
+            if not os.path.isabs(audio_path):
+                audio_path = str(manifest_dir / audio_path)
+
+            audio_error = _check_audio_file(audio_path, sample_rate)
+            if audio_error:
+                missing_audio_count += 1
+                if missing_audio_count <= 5:
+                    result.errors.append(f"Line {i + 1}: {audio_error}")
+                has_error = True
+            else:
+                duration = _get_audio_duration(audio_path)
+                if duration is not None:
+                    result.audio_durations.append(duration)
+                    if duration < 0.3:
+                        result.warnings.append(
+                            f"Line {i + 1}: Very short audio ({duration:.2f}s)"
+                        )
+                    elif duration > 30.0:
+                        result.warnings.append(
+                            f"Line {i + 1}: Very long audio ({duration:.1f}s), may cause OOM"
+                        )
+        else:
+            result.errors.append(f"Line {i + 1}: Invalid audio path")
+            has_error = True
+
+        # Validate optional ref_audio
+        if "ref_audio" in entry:
+            ref_path = entry["ref_audio"]
+            if isinstance(ref_path, dict):
+                ref_path = ref_path.get("path", "")
+            if isinstance(ref_path, str) and ref_path:
+                if not os.path.isabs(ref_path):
+                    ref_path = str(manifest_dir / ref_path)
+                if os.path.isfile(ref_path):
+                    result.has_ref_audio += 1
+                else:
+                    result.warnings.append(
+                        f"Line {i + 1}: ref_audio file not found: {ref_path}"
+                    )
+
+        if not has_error:
+            result.valid_samples += 1
+
+        if verbose and (i + 1) % 100 == 0:
+            print(f"  Validated {i + 1}/{samples_to_check} samples...", file=sys.stderr)
+
+    # Summarize truncated errors
+    if missing_audio_count > 5:
+        result.errors.append(
+            f"... and {missing_audio_count - 5} more missing audio files "
+            f"({missing_audio_count} total)"
+        )
+    if empty_text_count > 5:
+        result.warnings.append(
+            f"... and {empty_text_count - 5} more empty text entries "
+            f"({empty_text_count} total)"
+        )
+
+    return result
+
+
+def print_validation_report(result: ValidationResult, manifest_path: str) -> None:
+    """Print a human-readable validation report to stderr."""
+    print(f"\n{'=' * 60}", file=sys.stderr)
+    print(f"  VoxCPM Training Data Validation Report", file=sys.stderr)
+    print(f"{'=' * 60}", file=sys.stderr)
+    print(f"  Manifest : {manifest_path}", file=sys.stderr)
+    print(f"  Samples  : {result.valid_samples}/{result.total_samples} valid", file=sys.stderr)
+
+    if result.has_ref_audio > 0:
+        print(
+            f"  Ref Audio: {result.has_ref_audio} samples with reference audio",
+            file=sys.stderr,
+        )
+
+    # Audio duration statistics
+    if result.audio_durations:
+        durations = sorted(result.audio_durations)
+        total_hrs = sum(durations) / 3600
+        print(f"\n  Audio Duration Statistics:", file=sys.stderr)
+        print(f"    Total    : {total_hrs:.2f} hours", file=sys.stderr)
+        print(
+            f"    Range    : {durations[0]:.2f}s — {durations[-1]:.1f}s",
+            file=sys.stderr,
+        )
+        print(
+            f"    Mean     : {sum(durations) / len(durations):.2f}s",
+            file=sys.stderr,
+        )
+        median_idx = len(durations) // 2
+        print(f"    Median   : {durations[median_idx]:.2f}s", file=sys.stderr)
+
+    # Text length statistics
+    if result.text_lengths:
+        lengths = sorted(result.text_lengths)
+        print(f"\n  Text Length Statistics (characters):", file=sys.stderr)
+        print(
+            f"    Range    : {lengths[0]} — {lengths[-1]}",
+            file=sys.stderr,
+        )
+        print(
+            f"    Mean     : {sum(lengths) / len(lengths):.0f}",
+            file=sys.stderr,
+        )
+
+    # Errors
+    if result.errors:
+        print(f"\n  ERRORS ({len(result.errors)}):", file=sys.stderr)
+        for err in result.errors[:20]:
+            print(f"    x {err}", file=sys.stderr)
+        if len(result.errors) > 20:
+            print(
+                f"    ... ({len(result.errors) - 20} more errors omitted)",
+                file=sys.stderr,
+            )
+
+    # Warnings
+    if result.warnings:
+        print(f"\n  WARNINGS ({len(result.warnings)}):", file=sys.stderr)
+        for warn in result.warnings[:10]:
+            print(f"    ! {warn}", file=sys.stderr)
+        if len(result.warnings) > 10:
+            print(
+                f"    ... ({len(result.warnings) - 10} more warnings omitted)",
+                file=sys.stderr,
+            )
+
+    # Summary
+    print(f"\n{'=' * 60}", file=sys.stderr)
+    if result.is_valid:
+        print("  PASSED: Manifest is valid for training.", file=sys.stderr)
+    else:
+        print("  FAILED: Fix errors above before starting training.", file=sys.stderr)
+    print(f"{'=' * 60}\n", file=sys.stderr)

tests/test_cli.py

@@ -58,6 +58,7 @@ def test_parser_defaults_to_voxcpm2():
     parser = cli._build_parser()
     args = parser.parse_args(["design", "--text", "hello", "--output", "out.wav"])
     assert args.hf_model_id == "openbmb/VoxCPM2"
+    assert args.device == "auto"
     assert args.no_optimize is False
 
 
@@ -85,6 +86,7 @@ def test_load_model_respects_no_optimize_for_local_model(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is False
 
 
@@ -110,6 +112,7 @@ def test_load_model_defaults_optimize_for_hf(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is True
 
 
@@ -136,9 +139,37 @@ def test_load_model_respects_no_optimize_for_hf(monkeypatch):
 
     cli.load_model(args)
 
+    assert calls["kwargs"]["device"] == "auto"
     assert calls["kwargs"]["optimize"] is False
 
 
+def test_load_model_passes_explicit_device_to_hf(monkeypatch):
+    calls = {}
+
+    class FakeVoxCPM:
+        @classmethod
+        def from_pretrained(cls, **kwargs):
+            calls["kwargs"] = kwargs
+            return DummyModel()
+
+    monkeypatch.setattr(cli, "VoxCPM", FakeVoxCPM)
+    args = cli._build_parser().parse_args(
+        [
+            "design",
+            "--text",
+            "hello",
+            "--output",
+            "out.wav",
+            "--device",
+            "mps",
+        ]
+    )
+
+    cli.load_model(args)
+
+    assert calls["kwargs"]["device"] == "mps"
+
+
 def test_design_subcommand_applies_control(monkeypatch, tmp_path):
     dummy_model = DummyModel()
     monkeypatch.setattr(cli, "load_model", lambda args: dummy_model)

tests/test_lora_checkpoint_loading.py

@@ -0,0 +1,150 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+import torch
+
+ROOT = Path(__file__).resolve().parents[1]
+SRC = ROOT / "src"
+
+
+def _load_module(name: str, path: Path):
+    spec = importlib.util.spec_from_file_location(name, path)
+    module = importlib.util.module_from_spec(spec)
+    assert spec.loader is not None
+    sys.modules[name] = module
+    spec.loader.exec_module(module)
+    return module
+
+
+def bootstrap_repo_modules(monkeypatch):
+    for name, path in [
+        ("voxcpm", SRC / "voxcpm"),
+        ("voxcpm.model", SRC / "voxcpm" / "model"),
+        ("voxcpm.modules", SRC / "voxcpm" / "modules"),
+    ]:
+        pkg = types.ModuleType(name)
+        pkg.__path__ = [str(path)]
+        monkeypatch.setitem(sys.modules, name, pkg)
+
+    hh = types.ModuleType("huggingface_hub")
+    hh.snapshot_download = lambda *a, **k: "/tmp/fake"
+    monkeypatch.setitem(sys.modules, "huggingface_hub", hh)
+
+    pydantic = types.ModuleType("pydantic")
+
+    class BaseModel:
+        @classmethod
+        def model_rebuild(cls):
+            return None
+
+        @classmethod
+        def model_validate_json(cls, s):
+            return cls()
+
+        def model_dump(self):
+            return {}
+
+    pydantic.BaseModel = BaseModel
+    monkeypatch.setitem(sys.modules, "pydantic", pydantic)
+
+    torchaudio = types.ModuleType("torchaudio")
+    monkeypatch.setitem(sys.modules, "torchaudio", torchaudio)
+
+    librosa = types.ModuleType("librosa")
+    librosa.effects = types.SimpleNamespace(trim=lambda *a, **k: (None, (0, 0)))
+    monkeypatch.setitem(sys.modules, "librosa", librosa)
+
+    einops = types.ModuleType("einops")
+    einops.rearrange = lambda x, *a, **k: x
+    monkeypatch.setitem(sys.modules, "einops", einops)
+
+    tqdm_pkg = types.ModuleType("tqdm")
+    tqdm_pkg.__path__ = ["/nonexistent"]
+    tqdm_pkg.tqdm = lambda x, *a, **k: x
+    monkeypatch.setitem(sys.modules, "tqdm", tqdm_pkg)
+
+    tqdm_auto = types.ModuleType("tqdm.auto")
+    tqdm_auto.tqdm = lambda x, *a, **k: x
+    monkeypatch.setitem(sys.modules, "tqdm.auto", tqdm_auto)
+
+    transformers = types.ModuleType("transformers")
+
+    class LlamaTokenizerFast:
+        pass
+
+    class PreTrainedTokenizer:
+        pass
+
+    transformers.LlamaTokenizerFast = LlamaTokenizerFast
+    transformers.PreTrainedTokenizer = PreTrainedTokenizer
+    monkeypatch.setitem(sys.modules, "transformers", transformers)
+
+    internal_mods = {
+        "voxcpm.modules.audiovae": ["AudioVAE", "AudioVAEConfig", "AudioVAEV2", "AudioVAEConfigV2"],
+        "voxcpm.modules.layers": ["ScalarQuantizationLayer"],
+        "voxcpm.modules.locdit": ["CfmConfig", "UnifiedCFM", "VoxCPMLocDiT", "VoxCPMLocDiTV2"],
+        "voxcpm.modules.locenc": ["VoxCPMLocEnc"],
+        "voxcpm.modules.minicpm4": ["MiniCPM4Config", "MiniCPMModel"],
+        "voxcpm.modules.layers.lora": ["apply_lora_to_named_linear_modules", "LoRALinear"],
+    }
+    for modname, names in internal_mods.items():
+        module = types.ModuleType(modname)
+        for name in names:
+            if name == "apply_lora_to_named_linear_modules":
+                setattr(module, name, lambda *a, **k: None)
+            else:
+                setattr(module, name, type(name, (), {}))
+        monkeypatch.setitem(sys.modules, modname, module)
+
+    _load_module("voxcpm.model.utils", SRC / "voxcpm" / "model" / "utils.py")
+    voxcpm = _load_module("voxcpm.model.voxcpm", SRC / "voxcpm" / "model" / "voxcpm.py")
+    voxcpm2 = _load_module("voxcpm.model.voxcpm2", SRC / "voxcpm" / "model" / "voxcpm2.py")
+    return voxcpm.VoxCPMModel, voxcpm2.VoxCPM2Model
+
+
+class DummyModel:
+    device = "cpu"
+
+    def named_parameters(self):
+        return []
+
+
+@pytest.mark.parametrize("module_name", ["v1", "v2"])
+def test_load_lora_weights_accepts_tensor_only_legacy_checkpoints(monkeypatch, tmp_path, module_name):
+    VoxCPMModel, VoxCPM2Model = bootstrap_repo_modules(monkeypatch)
+    cls = VoxCPMModel if module_name == "v1" else VoxCPM2Model
+
+    ckpt_path = tmp_path / "lora_weights.ckpt"
+    torch.save({"state_dict": {"fake": torch.zeros(1)}}, ckpt_path)
+
+    loaded, skipped = cls.load_lora_weights(DummyModel(), str(ckpt_path), device="cpu")
+
+    assert loaded == []
+    assert skipped == ["fake"]
+
+
+@pytest.mark.parametrize("module_name", ["v1", "v2"])
+def test_load_lora_weights_rejects_malicious_pickle_payloads(monkeypatch, tmp_path, module_name):
+    VoxCPMModel, VoxCPM2Model = bootstrap_repo_modules(monkeypatch)
+    cls = VoxCPMModel if module_name == "v1" else VoxCPM2Model
+
+    ckpt_path = tmp_path / "lora_weights.ckpt"
+    marker_path = tmp_path / f"{module_name}-marker.txt"
+
+    class Exploit:
+        def __reduce__(self):
+            import pathlib
+
+            return (pathlib.Path.write_text, (marker_path, f"{module_name} executed\n"))
+
+    torch.save({"state_dict": {"fake": torch.zeros(1)}, "boom": Exploit()}, ckpt_path)
+
+    with pytest.raises(Exception, match="Weights only load failed"):
+        cls.load_lora_weights(DummyModel(), str(ckpt_path), device="cpu")
+
+    assert not marker_path.exists()

tests/test_model_utils.py

@@ -0,0 +1,49 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+UTILS_PATH = ROOT / "src" / "voxcpm" / "model" / "utils.py"
+
+transformers_stub = types.ModuleType("transformers")
+transformers_stub.PreTrainedTokenizer = object
+sys.modules.setdefault("transformers", transformers_stub)
+
+spec = importlib.util.spec_from_file_location("voxcpm.model.utils", UTILS_PATH)
+utils = importlib.util.module_from_spec(spec)
+assert spec.loader is not None
+spec.loader.exec_module(utils)
+
+
+def test_resolve_runtime_device_auto_falls_back_to_cpu(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: False)
+
+    assert utils.resolve_runtime_device(None, "cuda") == "cpu"
+
+
+def test_resolve_runtime_device_auto_uses_mps_when_available(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    assert utils.resolve_runtime_device("auto", "cuda") == "mps"
+
+
+def test_resolve_runtime_device_respects_explicit_cpu(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: True)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    assert utils.resolve_runtime_device("cpu", "cuda") == "cpu"
+
+
+def test_resolve_runtime_device_rejects_unavailable_explicit_cuda(monkeypatch):
+    monkeypatch.setattr(utils.torch.cuda, "is_available", lambda: False)
+    monkeypatch.setattr(utils, "_has_mps", lambda: True)
+
+    with pytest.raises(ValueError, match="CUDA is not available"):
+        utils.resolve_runtime_device("cuda:0", "cuda")

tests/test_validate.py

@@ -0,0 +1,252 @@
+"""Tests for the training data validation module."""
+
+from __future__ import annotations
+
+import json
+import os
+import sys
+import tempfile
+import types
+from pathlib import Path
+
+import pytest
+
+ROOT = Path(__file__).resolve().parents[1]
+
+# Stub voxcpm package so imports work without full dependencies
+pkg = types.ModuleType("voxcpm")
+pkg.__path__ = [str(ROOT / "src" / "voxcpm")]
+sys.modules.setdefault("voxcpm", pkg)
+
+training_pkg = types.ModuleType("voxcpm.training")
+training_pkg.__path__ = [str(ROOT / "src" / "voxcpm" / "training")]
+sys.modules.setdefault("voxcpm.training", training_pkg)
+
+from voxcpm.training.validate import ValidationResult, validate_manifest
+
+
+@pytest.fixture
+def tmp_dir():
+    with tempfile.TemporaryDirectory() as d:
+        yield Path(d)
+
+
+def _create_wav(path: Path, duration_s: float = 1.0, sr: int = 16000):
+    """Create a minimal valid WAV file."""
+    try:
+        import soundfile as sf
+        import numpy as np
+
+        samples = int(duration_s * sr)
+        data = np.zeros(samples, dtype=np.float32)
+        sf.write(str(path), data, sr)
+    except ImportError:
+        # If soundfile is not available, create a minimal WAV header
+        import struct
+
+        samples = int(duration_s * sr)
+        data_size = samples * 2  # 16-bit PCM
+        with open(path, "wb") as f:
+            f.write(b"RIFF")
+            f.write(struct.pack("<I", 36 + data_size))
+            f.write(b"WAVEfmt ")
+            f.write(struct.pack("<IHHIIHH", 16, 1, 1, sr, sr * 2, 2, 16))
+            f.write(b"data")
+            f.write(struct.pack("<I", data_size))
+            f.write(b"\x00" * data_size)
+
+
+def _write_manifest(path: Path, entries: list[dict]):
+    with open(path, "w", encoding="utf-8") as f:
+        for entry in entries:
+            f.write(json.dumps(entry, ensure_ascii=False) + "\n")
+
+
+class TestValidateManifest:
+    def test_valid_manifest(self, tmp_dir):
+        audio1 = tmp_dir / "audio1.wav"
+        audio2 = tmp_dir / "audio2.wav"
+        _create_wav(audio1, 2.0)
+        _create_wav(audio2, 3.0)
+
+        manifest = tmp_dir / "train.jsonl"
+        _write_manifest(
+            manifest,
+            [
+                {"text": "Hello world", "audio": str(audio1)},
+                {"text": "Goodbye world", "audio": str(audio2)},
+            ],
+        )
+
+        result = validate_manifest(str(manifest))
+        assert result.total_samples == 2
+        assert result.valid_samples == 2
+        assert result.is_valid
+        assert len(result.errors) == 0
+
+    def test_missing_manifest(self):
+        result = validate_manifest("/nonexistent/path.jsonl")
+        assert not result.is_valid
+        assert any("not found" in e for e in result.errors)
+
+    def test_empty_manifest(self, tmp_dir):
+        manifest = tmp_dir / "empty.jsonl"
+        manifest.write_text("")
+        result = validate_manifest(str(manifest))
+        assert not result.is_valid
+
+    def test_invalid_json(self, tmp_dir):
+        manifest = tmp_dir / "bad.jsonl"
+        manifest.write_text("not json\n{bad json}\n")
+        result = validate_manifest(str(manifest))
+        assert len(result.errors) >= 2
+        assert any("Invalid JSON" in e for e in result.errors)
+
+    def test_missing_columns(self, tmp_dir):
+        manifest = tmp_dir / "missing.jsonl"
+        _write_manifest(
+            manifest,
+            [
+                {"text": "hello"},  # missing audio
+                {"audio": "test.wav"},  # missing text
+            ],
+        )
+        result = validate_manifest(str(manifest))
+        assert len(result.errors) >= 2
+        assert any("'audio'" in e for e in result.errors)
+        assert any("'text'" in e for e in result.errors)
+
+    def test_missing_audio_file(self, tmp_dir):
+        manifest = tmp_dir / "missing_audio.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": "hello", "audio": "/nonexistent/audio.wav"}],
+        )
+        result = validate_manifest(str(manifest))
+        assert not result.is_valid
+        assert any("not found" in e for e in result.errors)
+
+    def test_empty_text_warning(self, tmp_dir):
+        audio = tmp_dir / "audio.wav"
+        _create_wav(audio)
+        manifest = tmp_dir / "empty_text.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": "", "audio": str(audio)}],
+        )
+        result = validate_manifest(str(manifest))
+        assert len(result.warnings) > 0
+        assert any("Empty" in w for w in result.warnings)
+
+    def test_relative_audio_path(self, tmp_dir):
+        audio = tmp_dir / "audio.wav"
+        _create_wav(audio)
+        manifest = tmp_dir / "rel.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": "hello", "audio": "audio.wav"}],
+        )
+        result = validate_manifest(str(manifest))
+        assert result.valid_samples == 1
+        assert result.is_valid
+
+    def test_max_samples_limit(self, tmp_dir):
+        audio = tmp_dir / "audio.wav"
+        _create_wav(audio)
+        manifest = tmp_dir / "many.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": f"sample {i}", "audio": str(audio)} for i in range(100)],
+        )
+        result = validate_manifest(str(manifest), max_samples=10)
+        assert result.total_samples == 10
+
+    def test_ref_audio_counted(self, tmp_dir):
+        audio = tmp_dir / "audio.wav"
+        ref = tmp_dir / "ref.wav"
+        _create_wav(audio)
+        _create_wav(ref)
+        manifest = tmp_dir / "ref.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": "hello", "audio": str(audio), "ref_audio": str(ref)}],
+        )
+        result = validate_manifest(str(manifest))
+        assert result.has_ref_audio == 1
+
+    def test_validation_result_properties(self):
+        r = ValidationResult(total_samples=5, valid_samples=5)
+        assert r.is_valid
+
+        r2 = ValidationResult(total_samples=5, valid_samples=5, errors=["err"])
+        assert not r2.is_valid
+
+        r3 = ValidationResult(total_samples=0, valid_samples=0)
+        assert not r3.is_valid
+
+    def test_invalid_audio_not_counted_as_valid(self, tmp_dir):
+        """A row with a bad audio path must not increment valid_samples."""
+        manifest = tmp_dir / "bad_audio.jsonl"
+        _write_manifest(
+            manifest,
+            [{"text": "hello", "audio": "/nonexistent/audio.wav"}],
+        )
+        result = validate_manifest(str(manifest))
+        assert result.total_samples == 1
+        assert result.valid_samples == 0
+        assert not result.is_valid
+        assert any("not found" in e for e in result.errors)
+
+    def test_sample_rate_mismatch(self, tmp_dir):
+        """A file with a different sample rate should be reported as an error."""
+        try:
+            import soundfile as sf
+            import numpy as np
+        except ImportError:
+            pytest.skip("soundfile not available")
+
+        audio = tmp_dir / "audio_8k.wav"
+        import numpy as np
+        samples = np.zeros(8000, dtype=np.float32)
+        sf.write(str(audio), samples, 8000)
+
+        manifest = tmp_dir / "sr_mismatch.jsonl"
+        _write_manifest(manifest, [{"text": "hello", "audio": str(audio)}])
+
+        result = validate_manifest(str(manifest), sample_rate=16000)
+        assert result.valid_samples == 0
+        assert not result.is_valid
+        assert any("Sample rate mismatch" in e or "sample rate" in e.lower() for e in result.errors)
+
+    def test_mixed_ref_audio_warns_for_each_missing(self, tmp_dir):
+        """Missing ref_audio entries should each generate a warning independently."""
+        audio = tmp_dir / "audio.wav"
+        ref_good = tmp_dir / "ref_good.wav"
+        _create_wav(audio)
+        _create_wav(ref_good)
+
+        manifest = tmp_dir / "mixed_ref.jsonl"
+        _write_manifest(
+            manifest,
+            [
+                {"text": "row1", "audio": str(audio), "ref_audio": str(ref_good)},
+                {"text": "row2", "audio": str(audio), "ref_audio": "/nonexistent/ref.wav"},
+            ],
+        )
+        result = validate_manifest(str(manifest))
+        assert result.has_ref_audio == 1
+        assert any("ref_audio file not found" in w for w in result.warnings)
+
+    def test_cli_validate_exit_code(self, tmp_dir):
+        """validate subcommand must exit 1 on validation error (missing audio)."""
+        import subprocess
+        manifest = tmp_dir / "bad.jsonl"
+        _write_manifest(manifest, [{"text": "hi", "audio": "/nonexistent/x.wav"}])
+
+        proc = subprocess.run(
+            [sys.executable, "-m", "voxcpm.cli", "validate", "--manifest", str(manifest)],
+            capture_output=True,
+            text=True,
+        )
+        assert proc.returncode == 1, f"Expected exit 1, got {proc.returncode}"
+        assert "FAILED" in proc.stderr or "Audio file not found" in proc.stderr