anonymous1436778242134213535654/Wan2.1-T2V-1.3B-Distill-Comfy

This model is a straight conversion of the original

Here is the code I used to convert it. Tried it in the latest comfyui and it works. See the included workflow file for an example. On my 3090, and using the lightvaew2_1 autoencoder it takes 25 seconds to run the workflow.

import torch
from safetensors.torch import save_file

SRC = '/home/user/convert model/model_iter6000.pt'
DST = '/home/user/convert model/wan2.1-t2v-1.3b-4step-distill.safetensors'

sd = torch.load(SRC, map_location='cpu')

def conv(k):
    if k.startswith('blocks.'):
        idx, rest = k[len('blocks.'):].split('.', 1)
        if rest.startswith('attn1.'):
            s = rest[len('attn1.'):].replace('to_out.0', 'o').replace('to_q', 'q').replace('to_k', 'k').replace('to_v', 'v')
            return f'blocks.{idx}.self_attn.{s}'
        if rest.startswith('attn2.'):
            s = rest[len('attn2.'):].replace('to_out.0', 'o').replace('to_q', 'q').replace('to_k', 'k').replace('to_v', 'v')
            return f'blocks.{idx}.cross_attn.{s}'
        if rest.startswith('ffn.net.0.proj.'):
            return f'blocks.{idx}.ffn.0.' + rest[len('ffn.net.0.proj.'):]
        if rest.startswith('ffn.net.2.'):
            return f'blocks.{idx}.ffn.2.' + rest[len('ffn.net.2.'):]
        if rest.startswith('norm2.'):
            return f'blocks.{idx}.norm3.' + rest[len('norm2.'):]
        if rest == 'scale_shift_table':
            return f'blocks.{idx}.modulation'
        return None
    if k in ('patch_embedding.weight', 'patch_embedding.bias'):
        return k
    if k == 'proj_out.weight':
        return 'head.head.weight'
    if k == 'proj_out.bias':
        return 'head.head.bias'
    if k == 'scale_shift_table':
        return 'head.modulation'
    if k.startswith('condition_embedder.time_embedder.linear_1.'):
        return 'time_embedding.0.' + k.split('.')[-1]
    if k.startswith('condition_embedder.time_embedder.linear_2.'):
        return 'time_embedding.2.' + k.split('.')[-1]
    if k.startswith('condition_embedder.time_proj.'):
        return 'time_projection.1.' + k.split('.')[-1]
    if k.startswith('condition_embedder.text_embedder.linear_1.'):
        return 'text_embedding.0.' + k.split('.')[-1]
    if k.startswith('condition_embedder.text_embedder.linear_2.'):
        return 'text_embedding.2.' + k.split('.')[-1]
    return None

out = {}
for k, v in sd.items():
    nk = conv(k)
    if nk is None:
        raise SystemExit(f'UNMAPPED: {k}')
    out[nk] = v.contiguous()

save_file(out, DST, metadata={'format': 'pt', 'framework': 'pt'})
print(f'Saved {len(out)} tensors -> {DST}')

GGUFs made with city96's stuff

Q3_K_S is for the desperate

Models from Q3_K_M upward start looking 'fine'.

The ggufs work great with comfyui. They are also compatible with leejet's stable-diffusion.cpp, but theyre buggy, video outputs seem messed up somewhat, and ram usage spikes on VAE decode to like 60 gb+ on a 5 second video using the TAE, if you use VAE and tile it its fine and vram usage is low. Might look into it later.