Comfy Kitchen

0.2.8 · active · verified Sat Apr 11

Comfy Kitchen is a fast kernel library designed for ComfyUI, offering optimized computational primitives with support for multiple compute backends such as PyTorch, JAX, and CuPy. It provides low-level, high-performance operations for AI workloads, allowing users to leverage different hardware and frameworks. The current version is 0.2.8, and it appears to follow a rapid iteration release cadence, indicating active development.

Warnings

gotcha Comfy Kitchen relies on optional backend dependencies (e.g., 'torch', 'jax', 'cupy') which must be installed separately using pip extra commands. Attempting to use a backend without its corresponding dependency will result in an `ImportError` or runtime errors.
Fix: Install with extras: `pip install comfy-kitchen[backend_name]` where `backend_name` is 'torch', 'jax', or 'cupy'.
gotcha Before loading any kernels, you must explicitly set the desired compute backend using `K.set_backend("torch")` (or "jax", "cupy", etc.). The library does not default to a backend, and attempting to load kernels without this step will raise an error.
Fix: Call `K.set_backend("your_backend_choice")` early in your code, after importing `comfy_kitchen.kernels`.
gotcha Comfy Kitchen is a low-level kernel library providing optimized primitives. It is not a high-level wrapper for ComfyUI itself or a complete ComfyUI API client. Users might incorrectly assume it provides tools for general ComfyUI orchestration.
Fix: Align expectations; this library is designed for specific kernel acceleration and high-performance operations, not for managing ComfyUI workflows or interacting with its UI layer.

Install

pip install comfy-kitchen Base installation
pip install comfy-kitchen[torch] With PyTorch backend
pip install comfy-kitchen[jax] With JAX backend
pip install comfy-kitchen[cupy] With CuPy backend

Imports

K
```
import comfy_kitchen.kernels as K
```
Kernels functionality is exposed through the `comfy_kitchen.kernels` submodule, commonly aliased as K.

Quickstart

This quickstart demonstrates how to initialize Comfy Kitchen, set a compute backend (e.g., 'torch'), list available kernels, and load a specific kernel. It highlights the importance of installing optional backend dependencies and explicitly setting the backend before use.

import comfy_kitchen.kernels as K
import os

# --- Important: Choose and install a backend first ---
# For this example, we default to 'torch'.
# Ensure you have installed the corresponding extra, e.g., `pip install comfy-kitchen[torch]`
try:
    backend_choice = os.environ.get('COMFY_KITCHEN_BACKEND', 'torch')
    K.set_backend(backend_choice)
    print(f"Successfully set Comfy Kitchen backend to: {backend_choice}")

    # List available kernels for the selected backend
    available_kernels = K.list_kernels()
    print(f"Available kernels for '{backend_choice}' backend: {', '.join(available_kernels)}")

    # Attempt to load an example kernel (e.g., a diffusion related one)
    kernel_name = "diffusion_forward" # Common kernel, but might vary per version/backend
    
    if kernel_name in available_kernels:
        kernel = K.load_kernel(kernel_name)
        print(f"Loaded kernel: {kernel.name} from backend: {K.get_backend()}")
        print(f"Kernel signature (expected inputs): {kernel.signature()}")
        
        # A full example would involve preparing inputs as per kernel.signature()
        # For demonstration, we just show successful loading and signature.
        # Example: to use, you'd prepare tensors (e.g., torch.Tensor) matching the signature
        # output = kernel.forward(*your_prepared_tensors_here)
        print("Kernel is ready for execution with appropriate input tensors.")
    else:
        print(f"Kernel '{kernel_name}' not found for backend '{backend_choice}'.")

except ImportError as e:
    print(f"ERROR: Failed to import a backend. Did you run `pip install comfy-kitchen[{backend_choice}]`? Details: {e}")
except Exception as e:
    print(f"An error occurred during Comfy Kitchen setup or kernel loading: {e}")

view raw JSON →