comfy-env 0.3.70

Environment management for ComfyUI custom nodes - CUDA wheel resolution and process isolation

comfy-env

Environment management for ComfyUI custom nodes.

The Problem

ComfyUI custom nodes share a single Python environment. This breaks when:

• Node A needs torch 2.4, Node B needs torch 2.8
• Two packages bundle conflicting native libraries (libomp, CUDA runtimes)
• A node requires a specific Python version (e.g., Blender API needs 3.11)

The Solution

comfy-env provides process isolation -- nodes that need conflicting dependencies run in their own Python environments as persistent subprocesses, transparent to ComfyUI.

Two config files, two roles:

• comfy-env-root.toml (root level): System packages (apt/brew) and ComfyUI node dependency management. Never touches the Python environment -- PyPI deps stay in requirements.txt.
• comfy-env.toml (subdirectories): Each subdirectory with this file gets its own isolated Python environment via pixi, with a separate interpreter, conda packages, pip packages, and pre-built CUDA wheels.

Architecture

ComfyUI-MyPack/
+-- comfy-env-root.toml            # System packages + node deps
+-- install.py                     # from comfy_env import install; install()
+-- prestartup_script.py           # from comfy_env import setup_env; setup_env()
+-- __init__.py                    # from comfy_env import register_nodes
`-- nodes/
    +-- main/                      # No config -> imported in main process
    |   `-- __init__.py
    `-- cgal/                      # Has config -> isolated subprocess
        +-- comfy-env.toml
        +-- _env_a1b2c3 ---------> <drive>/ce/_env_a1b2c3/.pixi/envs/default
        `-- __init__.py

<drive>/ce/                        # Central build cache (same drive as ComfyUI)
`-- _env_a1b2c3/                   # SHA256(config + comfy-env version)[:8]
    +-- .pixi/envs/default/        # Complete Python environment
    |   +-- bin/python
    |   `-- lib/python3.11/site-packages/
    +-- pixi.toml                  # Generated from comfy-env.toml
    `-- .done                      # Build complete marker

How It Works

Build time (install.py): For each subdirectory with a comfy-env.toml, comfy-env hashes the config contents + its own package version, checks the central build cache (<drive>/ce on Windows, ~/.ce on Unix; override with COMFY_ENV_BUILD_BASE), and builds a pixi environment if needed. The result is linked into the node directory as _env_<hash> -- symlink on Unix, NTFS junction on Windows (no admin required). Identical configs across different node packs share the same cached build.

Runtime (register_nodes()): Discovers all node subdirectories. Those with a built _env_* run in persistent subprocess workers using the isolated Python interpreter. Those without a config are imported normally. Workers communicate via Unix domain sockets (TCP on Windows) and support bidirectional callbacks for VRAM budget negotiation and progress reporting.

CUDA packages: Listed in [cuda], installed from the PyTorch wheel index or cuda-wheels -- pre-built wheels for nvdiffrast, pytorch3d, gsplat, etc. No CUDA toolkit or C++ compiler needed.

Future work: replace share_torch with pixi's Multi-Environment feature

Today, when the host ComfyUI venv and an isolation env agree on torch version, share_torch installs torch into the isolation env, then uv pip uninstalls it, and redirects the worker's sys.path to the host's torch at runtime. This is fragile on Windows: the uninstall leaves torch/lib/*.dll files on disk (pip/uv strip the dist-info but not the DLLs), which causes OSError: [WinError 127] in cold-install scenarios because Python finds the half-gutted torch/ folder and tries to load its stranded DLLs.

The intended fix is to drop the strip-after-install approach entirely and use pixi's Multi-Environment feature. Pixi lets you declare [feature.X] tables (e.g., a shared host-torch feature with the same torch/torchvision the host uses) and compose them into named environments. Packages present in multiple environments at the same version are stored once on disk and hardlinked into each environment -- no physical duplication, no runtime sys.path hacks, no uninstall step that Windows can't do cleanly.

Concrete migration sketch:

1. Generate a single pixi.toml for the whole node pack with a host-torch feature (torch, torchvision, CUDA wheels) and one feature per isolation env.
2. Define one environment per isolation env as [host-torch, <node-feature>].
3. Drop the share_torch codepath; workers use their composed env directly.

Research on alternatives (April 2026): pip/uv/pixi/poetry overrides only change versions, not install decisions -- "constraints and overrides cannot exclude packages from installation" (uv docs). conda's --stack mutates sys.path at activate time and is known to break (base-env packages sometimes unimportable). Multi-Environment is the only first-class primitive that actually composes envs with real on-disk dedup.

Usage

# install.py
from comfy_env import install
install()

# prestartup_script.py
from comfy_env import setup_env
setup_env()

# __init__.py
from comfy_env import register_nodes
NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS = register_nodes()

pip install comfy-env

Docs

• Getting Started -- setup guide and walkthrough
• Config Reference -- all config options for both files
• CLI -- command-line tools
• Build Internals -- how isolated environments are built
• Worker Architecture -- subprocess IPC, serialization, memory, lifecycle

Example

ComfyUI-GeometryPack -- multiple isolated environments (CGAL, Blender, GPU) with per-subdirectory configs and different Python versions.