fix: add install.py to prevent omnivoice from overwriting ComfyUI's torch

README.md

@@ -18,15 +18,14 @@ A ComfyUI custom node for [OmniVoice](https://github.com/k2-fsa/OmniVoice) — a
    git clone https://github.com/ethanfel/ComfyUI-Omnivoice.git
    ```
 
-2. Install dependencies:
+2. Install dependencies **without overwriting ComfyUI's torch**:
    ```bash
-   pip install omnivoice
+   python install.py
    ```
+   > **Warning:** Do NOT run `pip install omnivoice` directly — it pins `torch==2.8.*` from a CUDA 12.8 index and will overwrite ComfyUI's torch installation.
 
 3. Restart ComfyUI. The nodes will appear under the **OmniVoice** category.
 
-> **Note:** OmniVoice requires PyTorch 2.8+ and a CUDA-capable GPU (or Apple Silicon).
-
 ## Nodes
 
 ### OmniVoice Model Loader

install.py

@@ -0,0 +1,32 @@
+"""
+Installation script for ComfyUI-Omnivoice.
+
+ComfyUI Manager runs this file instead of pip-installing requirements.txt directly.
+We install omnivoice with --no-deps to avoid overwriting ComfyUI's torch installation.
+omnivoice pins torch==2.8.* from a CUDA 12.8 custom index which would break ComfyUI.
+"""
+import subprocess
+import sys
+
+
+def pip(*args):
+    subprocess.check_call([sys.executable, "-m", "pip", "install", *args])
+
+
+# Install omnivoice itself without pulling in its torch/torchaudio pins.
+# ComfyUI ships its own torch build — let it manage torch.
+pip("omnivoice", "--no-deps")
+
+# Install omnivoice's runtime inference dependencies (excludes torch, torchaudio,
+# gradio, tensorboardX, webdataset which are training/demo-only tools).
+pip(
+    "transformers>=5.0.0",
+    "accelerate",
+    "pydub",
+    "soundfile",
+    "numpy",
+)
+
+print("\n[ComfyUI-Omnivoice] Installation complete.")
+print("[ComfyUI-Omnivoice] NOTE: omnivoice was installed without its pinned torch.")
+print("[ComfyUI-Omnivoice] If you encounter errors, ensure torch>=2.0 is installed.")

requirements.txt

@@ -1 +1,3 @@
-omnivoice
+# Dependencies are managed by install.py to avoid overwriting ComfyUI's torch.
+# omnivoice pins torch==2.8.* (CUDA 12.8) which would break ComfyUI's torch build.
+# Do not add omnivoice here — install.py handles it with --no-deps.