Metadata-Version: 2.4
Name: comfy-env
Version: 0.0.43
Summary: Environment management for ComfyUI custom nodes - CUDA wheel resolution and process isolation
Project-URL: Homepage, https://github.com/PozzettiAndrea/comfy-env
Project-URL: Repository, https://github.com/PozzettiAndrea/comfy-env
Project-URL: Issues, https://github.com/PozzettiAndrea/comfy-env/issues
Author: Andrea Pozzetti
License: MIT
License-File: LICENSE
Keywords: comfyui,cuda,environment,isolation,process,venv,wheels
Classifier: Development Status :: 3 - Alpha
Classifier: Intended Audience :: Developers
Classifier: License :: OSI Approved :: MIT License
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.13
Requires-Python: >=3.10
Requires-Dist: pyyaml>=6.0
Requires-Dist: tomli>=2.0.0; python_version < '3.11'
Requires-Dist: uv>=0.4.0
Provides-Extra: dev
Requires-Dist: mypy; extra == 'dev'
Requires-Dist: pytest; extra == 'dev'
Requires-Dist: ruff; extra == 'dev'
Description-Content-Type: text/markdown

# comfy-env

Environment management for ComfyUI custom nodes. Provides:

1. **CUDA Wheel Resolution** - Install pre-built CUDA wheels (nvdiffrast, pytorch3d) without compilation
2. **Process Isolation** - Run nodes in separate Python environments with their own dependencies

## Why?

ComfyUI custom nodes face two challenges:

**Type 1: Dependency Conflicts**
- Node A needs `torch==2.1.0` with CUDA 11.8
- Node B needs `torch==2.8.0` with CUDA 12.8

**Type 2: CUDA Package Installation**
- Users don't have compilers installed
- Building from source takes forever
- pip install fails with cryptic errors

This package solves both problems.

## Installation

```bash
pip install comfy-env
```

Requires [uv](https://github.com/astral-sh/uv) for fast environment creation:

```bash
curl -LsSf https://astral.sh/uv/install.sh | sh
```

## Quick Start

### In-Place Installation (Type 2 - CUDA Wheels)

Create a `comfy-env.toml` in your node directory:

```toml
[cuda]
nvdiffrast = "0.4.0"
pytorch3d = "0.7.9"

[packages]
requirements = ["transformers>=4.56", "pillow"]
```

Then in your `__init__.py`:

```python
from comfy_env import install

# Install CUDA wheels into current environment
install()
```

### Process Isolation (Type 1 - Separate Venv)

For nodes that need completely separate dependencies:

```python
from comfy_env import isolated

@isolated(env="my-node")
class MyNode:
    FUNCTION = "process"
    RETURN_TYPES = ("IMAGE",)

    def process(self, image):
        # Runs in isolated subprocess with its own venv
        import conflicting_package
        return (result,)
```

## CLI

```bash
# Show detected environment
comfy-env info

# Install from config
comfy-env install

# Dry run (show what would be installed)
comfy-env install --dry-run

# Resolve wheel URLs without installing
comfy-env resolve nvdiffrast==0.4.0

# List all packages in the built-in registry
comfy-env list-packages

# Verify installation
comfy-env doctor
```

## Configuration

### Simple Format (comfy-env.toml)

```toml
# CUDA packages (uses built-in registry)
[cuda]
nvdiffrast = "0.4.0"
pytorch3d = "0.7.9"
torch-scatter = "2.1.2"

# Regular pip packages
[packages]
requirements = ["transformers>=4.56", "pillow"]
```

### Full Format

```toml
[system]
linux = ["libgl1", "libopengl0"]  # apt packages

[local.cuda]
nvdiffrast = "0.4.0"

[local.packages]
requirements = ["pillow", "numpy"]

# For isolated environments (creates separate venv)
[myenv]
python = "3.10"
cuda = "12.8"

[myenv.cuda]
torch-scatter = "2.1.2"

[myenv.packages]
requirements = ["transformers>=4.56"]

# Custom wheel templates (override built-in registry)
[wheel_sources]
my-custom-pkg = "https://my-server.com/my-pkg-{version}+cu{cuda_short}-{py_tag}-{platform}.whl"
```

## Writing Wheel Templates

### Template Variables

| Variable | Example | Description |
|----------|---------|-------------|
| `{version}` | `0.4.0` | Package version |
| `{cuda_version}` | `12.8` | Full CUDA version |
| `{cuda_short}` | `128` | CUDA without dot |
| `{cuda_major}` | `12` | CUDA major only |
| `{torch_version}` | `2.8.0` | Full PyTorch version |
| `{torch_mm}` | `28` | PyTorch major.minor no dot |
| `{torch_dotted_mm}` | `2.8` | PyTorch major.minor with dot |
| `{py_version}` | `3.10` | Python version |
| `{py_short}` | `310` | Python without dot |
| `{py_tag}` | `cp310` | Python wheel tag |
| `{platform}` | `linux_x86_64` | Platform tag |

### Common Wheel URL Patterns

**Pattern 1: Simple CUDA + Python**
```
https://example.com/{package}-{version}+cu{cuda_short}-{py_tag}-{py_tag}-{platform}.whl
```

**Pattern 2: CUDA + PyTorch**
```
https://example.com/{package}-{version}+cu{cuda_short}torch{torch_mm}-{py_tag}-{py_tag}-{platform}.whl
```

**Pattern 3: GitHub Releases**
```
https://github.com/org/repo/releases/download/v{version}/{package}-{version}+cu{cuda_short}-{py_tag}-{platform}.whl
```

### How to Find the Right Template

1. Download a wheel manually from the source
2. Look at the filename pattern: `nvdiffrast-0.4.0+cu128torch28-cp310-cp310-linux_x86_64.whl`
3. Replace values with variables: `nvdiffrast-{version}+cu{cuda_short}torch{torch_mm}-{py_tag}-{py_tag}-{platform}.whl`
4. Prepend the base URL

### Testing Your Template

```bash
comfy-env resolve my-package==1.0.0
```

This shows the resolved URL without installing.

### Adding Custom Wheel Sources

If a package isn't in the built-in registry, add it to your `comfy-env.toml`:

```toml
[cuda]
my-custom-pkg = "1.0.0"

[wheel_sources]
my-custom-pkg = "https://my-server.com/my-custom-pkg-{version}+cu{cuda_short}-{py_tag}-{platform}.whl"
```

Resolution order:
1. User's `[wheel_sources]` in comfy-env.toml (highest priority)
2. Built-in `wheel_sources.yml` registry
3. Error if not found

## API Reference

### install()

```python
from comfy_env import install

# Auto-discover config
install()

# Explicit config
install(config="comfy-env.toml")

# Dry run
install(dry_run=True)
```

### RuntimeEnv

```python
from comfy_env import RuntimeEnv

env = RuntimeEnv.detect()
print(env)
# Python 3.10, CUDA 12.8, PyTorch 2.8.0, GPU: NVIDIA GeForce RTX 4090

# Get template variables
vars_dict = env.as_dict()
# {'cuda_version': '12.8', 'cuda_short': '128', 'torch_mm': '28', ...}
```

### Workers (for isolation)

```python
from comfy_env import TorchMPWorker

# Same-venv isolation (zero-copy tensors)
worker = TorchMPWorker()
result = worker.call(my_function, image=tensor)
```

## GPU Detection

```python
from comfy_env import detect_cuda_version, get_gpu_summary

cuda = detect_cuda_version()  # "12.8", "12.4", or None
print(get_gpu_summary())
# GPU 0: NVIDIA GeForce RTX 5090 (sm_120) [Blackwell - CUDA 12.8]
```

## Built-in Package Registry

Run `comfy-env list-packages` to see all packages in the built-in registry.

The registry includes:
- PyTorch Geometric packages (torch-scatter, torch-cluster, torch-sparse)
- NVIDIA packages (nvdiffrast, pytorch3d, gsplat)
- Flash Attention (flash-attn)
- And more

## License

MIT - see LICENSE file.
