comfy/diffusers_convert.py

import re
import torch
import logging

# conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py

# ================#
# VAE Conversion #
# ================#

vae_conversion_map = [
    # (stable-diffusion, HF Diffusers)
    ("nin_shortcut", "conv_shortcut"),
    ("norm_out", "conv_norm_out"),
    ("mid.attn_1.", "mid_block.attentions.0."),
]

for i in range(4):
    # down_blocks have two resnets
    for j in range(2):
        hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
        sd_down_prefix = f"encoder.down.{i}.block.{j}."
        vae_conversion_map.append((sd_down_prefix, hf_down_prefix))

    if i < 3:
        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
        sd_downsample_prefix = f"down.{i}.downsample."
        vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))

        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
        sd_upsample_prefix = f"up.{3 - i}.upsample."
        vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))

    # up_blocks have three resnets
    # also, up blocks in hf are numbered in reverse from sd
    for j in range(3):
        hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
        sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
        vae_conversion_map.append((sd_up_prefix, hf_up_prefix))

# this part accounts for mid blocks in both the encoder and the decoder
for i in range(2):
    hf_mid_res_prefix = f"mid_block.resnets.{i}."
    sd_mid_res_prefix = f"mid.block_{i + 1}."
    vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))

vae_conversion_map_attn = [
    # (stable-diffusion, HF Diffusers)
    ("norm.", "group_norm."),
    ("q.", "query."),
    ("k.", "key."),
    ("v.", "value."),
    ("q.", "to_q."),
    ("k.", "to_k."),
    ("v.", "to_v."),
    ("proj_out.", "to_out.0."),
    ("proj_out.", "proj_attn."),
]


def reshape_weight_for_sd(w, conv3d=False):
    # convert HF linear weights to SD conv2d weights
    if conv3d:
        return w.reshape(*w.shape, 1, 1, 1)
    else:
        return w.reshape(*w.shape, 1, 1)


def convert_vae_state_dict(vae_state_dict):
    mapping = {k: k for k in vae_state_dict.keys()}
    conv3d = False
    for k, v in mapping.items():
        for sd_part, hf_part in vae_conversion_map:
            v = v.replace(hf_part, sd_part)
        if v.endswith(".conv.weight"):
            if not conv3d and vae_state_dict[k].ndim == 5:
                conv3d = True
        mapping[k] = v
    for k, v in mapping.items():
        if "attentions" in k:
            for sd_part, hf_part in vae_conversion_map_attn:
                v = v.replace(hf_part, sd_part)
            mapping[k] = v
    new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
    weights_to_convert = ["q", "k", "v", "proj_out"]
    for k, v in new_state_dict.items():
        for weight_name in weights_to_convert:
            if f"mid.attn_1.{weight_name}.weight" in k:
                logging.debug(f"Reshaping {k} for SD format")
                new_state_dict[k] = reshape_weight_for_sd(v, conv3d=conv3d)
    return new_state_dict


# =========================#
# Text Encoder Conversion #
# =========================#


textenc_conversion_lst = [
    # (stable-diffusion, HF Diffusers)
    ("resblocks.", "text_model.encoder.layers."),
    ("ln_1", "layer_norm1"),
    ("ln_2", "layer_norm2"),
    (".c_fc.", ".fc1."),
    (".c_proj.", ".fc2."),
    (".attn", ".self_attn"),
    ("ln_final.", "transformer.text_model.final_layer_norm."),
    ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
    ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
]
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
textenc_pattern = re.compile("|".join(protected.keys()))

# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
code2idx = {"q": 0, "k": 1, "v": 2}


# This function exists because at the time of writing torch.cat can't do fp8 with cuda
def cat_tensors(tensors):
    x = 0
    for t in tensors:
        x += t.shape[0]

    shape = [x] + list(tensors[0].shape)[1:]
    out = torch.empty(shape, device=tensors[0].device, dtype=tensors[0].dtype)

    x = 0
    for t in tensors:
        out[x:x + t.shape[0]] = t
        x += t.shape[0]

    return out


def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
    new_state_dict = {}
    capture_qkv_weight = {}
    capture_qkv_bias = {}
    for k, v in text_enc_dict.items():
        if not k.startswith(prefix):
            continue
        if (
                k.endswith(".self_attn.q_proj.weight")
                or k.endswith(".self_attn.k_proj.weight")
                or k.endswith(".self_attn.v_proj.weight")
        ):
            k_pre = k[: -len(".q_proj.weight")]
            k_code = k[-len("q_proj.weight")]
            if k_pre not in capture_qkv_weight:
                capture_qkv_weight[k_pre] = [None, None, None]
            capture_qkv_weight[k_pre][code2idx[k_code]] = v
            continue

        if (
                k.endswith(".self_attn.q_proj.bias")
                or k.endswith(".self_attn.k_proj.bias")
                or k.endswith(".self_attn.v_proj.bias")
        ):
            k_pre = k[: -len(".q_proj.bias")]
            k_code = k[-len("q_proj.bias")]
            if k_pre not in capture_qkv_bias:
                capture_qkv_bias[k_pre] = [None, None, None]
            capture_qkv_bias[k_pre][code2idx[k_code]] = v
            continue

        text_proj = "transformer.text_projection.weight"
        if k.endswith(text_proj):
            new_state_dict[k.replace(text_proj, "text_projection")] = v.transpose(0, 1).contiguous()
        else:
            relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
            new_state_dict[relabelled_key] = v

    for k_pre, tensors in capture_qkv_weight.items():
        if None in tensors:
            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
        new_state_dict[relabelled_key + ".in_proj_weight"] = cat_tensors(tensors)

    for k_pre, tensors in capture_qkv_bias.items():
        if None in tensors:
            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
        new_state_dict[relabelled_key + ".in_proj_bias"] = cat_tensors(tensors)

    return new_state_dict


def convert_text_enc_state_dict(text_enc_dict):
    return text_enc_dict