8

How to Contribute Code

Alex "mcmonkey" Goodwin edited this page 2024-07-19 11:27:36 -07:00

Table of Contents

• General
• Special Cases
• Frontend Changes
• Adding support for new model architectures

General

For general improvements/bug fixes just make a pull request.

• Before doing anything, make sure your change is wanted. Make sure there's an open Feature Request or Bug Report on the issues page.
• Try to make a single pull request for each change to make reviewing easier, as opposed to large/bulky PRs.
  • Especially first time contributors should focus on very simple and small tasks, and take on tougher ones after you've had a PR or two successfully pulled.
• avoid adding "sensitive" code, eg eval(...), unless absolutely unavoidable
• When you submit a pull request, please make sure you write a clear title and good description text.
  • Description text should be detailed but concise. What issue are you addressing, how does this PR address it, what have you done to test the change, what potential concerns or side effects may apply?

Special Cases

Frontend Changes

• No new frontend features should be pushed to the ComfyUI/web folder. Instead they should be submitted to https://github.com/Comfy-Org/ComfyUI_frontend
• Bugfixes are still welcome in the main repo

Adding support for new model architectures

Checklist of requirements for a PR that adds support for a new model architecture:

• Have a minimal implementation of the model code that only depends on pytorch under a license compatible with the GPL license that ComfyUI uses.
• Provide a reference image with sampling settings/seed/etc. so that we can make sure the ComfyUI implementation matches the reference one.
• Replace all attention functions with the comfyui optimized_attention attention function.
• If you are the part of the team that authored the model architecture itself:
  • Please release your primary models in .safetensors file format, not legacy .ckpt pickle files.
  • Please include proper identifying metadata in the header of the file.

Example of the SDPA implementation:

def optimized_attention(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False):
    if skip_reshape:
        b, _, _, dim_head = q.shape
    else:
        b, _, dim_head = q.shape
        dim_head //= heads
        q, k, v = map(
            lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
            (q, k, v),
        )

    out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
    out = (
        out.transpose(1, 2).reshape(b, -1, heads * dim_head)
    )
    return out

Examples in the Code:

Audio

SDXL Cascade

mmdit

After that you can just ping comfyanonymous on matrix or @comfyanon on discord and he will take a look.