tsi-org/tango

1# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.2# *Only* converts the UNet, VAE, and Text Encoder.3# Does not convert optimizer state or any other thing.4 5import argparse6import os.path as osp7import re8 9import torch10from safetensors.torch import load_file, save_file11 12 13# =================#14# UNet Conversion #15# =================#16 17unet_conversion_map = [18    # (stable-diffusion, HF Diffusers)19    ("time_embed.0.weight", "time_embedding.linear_1.weight"),20    ("time_embed.0.bias", "time_embedding.linear_1.bias"),21    ("time_embed.2.weight", "time_embedding.linear_2.weight"),22    ("time_embed.2.bias", "time_embedding.linear_2.bias"),23    ("input_blocks.0.0.weight", "conv_in.weight"),24    ("input_blocks.0.0.bias", "conv_in.bias"),25    ("out.0.weight", "conv_norm_out.weight"),26    ("out.0.bias", "conv_norm_out.bias"),27    ("out.2.weight", "conv_out.weight"),28    ("out.2.bias", "conv_out.bias"),29]30 31unet_conversion_map_resnet = [32    # (stable-diffusion, HF Diffusers)33    ("in_layers.0", "norm1"),34    ("in_layers.2", "conv1"),35    ("out_layers.0", "norm2"),36    ("out_layers.3", "conv2"),37    ("emb_layers.1", "time_emb_proj"),38    ("skip_connection", "conv_shortcut"),39]40 41unet_conversion_map_layer = []42# hardcoded number of downblocks and resnets/attentions...43# would need smarter logic for other networks.44for i in range(4):45    # loop over downblocks/upblocks46 47    for j in range(2):48        # loop over resnets/attentions for downblocks49        hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."50        sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."51        unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))52 53        if i < 3:54            # no attention layers in down_blocks.355            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."56            sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."57            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))58 59    for j in range(3):60        # loop over resnets/attentions for upblocks61        hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."62        sd_up_res_prefix = f"output_blocks.{3*i + j}.0."63        unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))64 65        if i > 0:66            # no attention layers in up_blocks.067            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."68            sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."69            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))70 71    if i < 3:72        # no downsample in down_blocks.373        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."74        sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."75        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))76 77        # no upsample in up_blocks.378        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."79        sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."80        unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))81 82hf_mid_atn_prefix = "mid_block.attentions.0."83sd_mid_atn_prefix = "middle_block.1."84unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))85 86for j in range(2):87    hf_mid_res_prefix = f"mid_block.resnets.{j}."88    sd_mid_res_prefix = f"middle_block.{2*j}."89    unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))90 91 92def convert_unet_state_dict(unet_state_dict):93    # buyer beware: this is a *brittle* function,94    # and correct output requires that all of these pieces interact in95    # the exact order in which I have arranged them.96    mapping = {k: k for k in unet_state_dict.keys()}97    for sd_name, hf_name in unet_conversion_map:98        mapping[hf_name] = sd_name99    for k, v in mapping.items():100        if "resnets" in k:101            for sd_part, hf_part in unet_conversion_map_resnet:102                v = v.replace(hf_part, sd_part)103            mapping[k] = v104    for k, v in mapping.items():105        for sd_part, hf_part in unet_conversion_map_layer:106            v = v.replace(hf_part, sd_part)107        mapping[k] = v108    new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}109    return new_state_dict110 111 112# ================#113# VAE Conversion #114# ================#115 116vae_conversion_map = [117    # (stable-diffusion, HF Diffusers)118    ("nin_shortcut", "conv_shortcut"),119    ("norm_out", "conv_norm_out"),120    ("mid.attn_1.", "mid_block.attentions.0."),121]122 123for i in range(4):124    # down_blocks have two resnets125    for j in range(2):126        hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."127        sd_down_prefix = f"encoder.down.{i}.block.{j}."128        vae_conversion_map.append((sd_down_prefix, hf_down_prefix))129 130    if i < 3:131        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."132        sd_downsample_prefix = f"down.{i}.downsample."133        vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))134 135        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."136        sd_upsample_prefix = f"up.{3-i}.upsample."137        vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))138 139    # up_blocks have three resnets140    # also, up blocks in hf are numbered in reverse from sd141    for j in range(3):142        hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."143        sd_up_prefix = f"decoder.up.{3-i}.block.{j}."144        vae_conversion_map.append((sd_up_prefix, hf_up_prefix))145 146# this part accounts for mid blocks in both the encoder and the decoder147for i in range(2):148    hf_mid_res_prefix = f"mid_block.resnets.{i}."149    sd_mid_res_prefix = f"mid.block_{i+1}."150    vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))151 152 153vae_conversion_map_attn = [154    # (stable-diffusion, HF Diffusers)155    ("norm.", "group_norm."),156    ("q.", "query."),157    ("k.", "key."),158    ("v.", "value."),159    ("proj_out.", "proj_attn."),160]161 162 163def reshape_weight_for_sd(w):164    # convert HF linear weights to SD conv2d weights165    return w.reshape(*w.shape, 1, 1)166 167 168def convert_vae_state_dict(vae_state_dict):169    mapping = {k: k for k in vae_state_dict.keys()}170    for k, v in mapping.items():171        for sd_part, hf_part in vae_conversion_map:172            v = v.replace(hf_part, sd_part)173        mapping[k] = v174    for k, v in mapping.items():175        if "attentions" in k:176            for sd_part, hf_part in vae_conversion_map_attn:177                v = v.replace(hf_part, sd_part)178            mapping[k] = v179    new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}180    weights_to_convert = ["q", "k", "v", "proj_out"]181    for k, v in new_state_dict.items():182        for weight_name in weights_to_convert:183            if f"mid.attn_1.{weight_name}.weight" in k:184                print(f"Reshaping {k} for SD format")185                new_state_dict[k] = reshape_weight_for_sd(v)186    return new_state_dict187 188 189# =========================#190# Text Encoder Conversion #191# =========================#192 193 194textenc_conversion_lst = [195    # (stable-diffusion, HF Diffusers)196    ("resblocks.", "text_model.encoder.layers."),197    ("ln_1", "layer_norm1"),198    ("ln_2", "layer_norm2"),199    (".c_fc.", ".fc1."),200    (".c_proj.", ".fc2."),201    (".attn", ".self_attn"),202    ("ln_final.", "transformer.text_model.final_layer_norm."),203    ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),204    ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),205]206protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}207textenc_pattern = re.compile("|".join(protected.keys()))208 209# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp210code2idx = {"q": 0, "k": 1, "v": 2}211 212 213def convert_text_enc_state_dict_v20(text_enc_dict):214    new_state_dict = {}215    capture_qkv_weight = {}216    capture_qkv_bias = {}217    for k, v in text_enc_dict.items():218        if (219            k.endswith(".self_attn.q_proj.weight")220            or k.endswith(".self_attn.k_proj.weight")221            or k.endswith(".self_attn.v_proj.weight")222        ):223            k_pre = k[: -len(".q_proj.weight")]224            k_code = k[-len("q_proj.weight")]225            if k_pre not in capture_qkv_weight:226                capture_qkv_weight[k_pre] = [None, None, None]227            capture_qkv_weight[k_pre][code2idx[k_code]] = v228            continue229 230        if (231            k.endswith(".self_attn.q_proj.bias")232            or k.endswith(".self_attn.k_proj.bias")233            or k.endswith(".self_attn.v_proj.bias")234        ):235            k_pre = k[: -len(".q_proj.bias")]236            k_code = k[-len("q_proj.bias")]237            if k_pre not in capture_qkv_bias:238                capture_qkv_bias[k_pre] = [None, None, None]239            capture_qkv_bias[k_pre][code2idx[k_code]] = v240            continue241 242        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)243        new_state_dict[relabelled_key] = v244 245    for k_pre, tensors in capture_qkv_weight.items():246        if None in tensors:247            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")248        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)249        new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)250 251    for k_pre, tensors in capture_qkv_bias.items():252        if None in tensors:253            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")254        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)255        new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)256 257    return new_state_dict258 259 260def convert_text_enc_state_dict(text_enc_dict):261    return text_enc_dict262 263 264if __name__ == "__main__":265    parser = argparse.ArgumentParser()266 267    parser.add_argument("--model_path", default=None, type=str, required=True, help="Path to the model to convert.")268    parser.add_argument("--checkpoint_path", default=None, type=str, required=True, help="Path to the output model.")269    parser.add_argument("--half", action="store_true", help="Save weights in half precision.")270    parser.add_argument(271        "--use_safetensors", action="store_true", help="Save weights use safetensors, default is ckpt."272    )273 274    args = parser.parse_args()275 276    assert args.model_path is not None, "Must provide a model path!"277 278    assert args.checkpoint_path is not None, "Must provide a checkpoint path!"279 280    # Path for safetensors281    unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.safetensors")282    vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.safetensors")283    text_enc_path = osp.join(args.model_path, "text_encoder", "model.safetensors")284 285    # Load models from safetensors if it exists, if it doesn't pytorch286    if osp.exists(unet_path):287        unet_state_dict = load_file(unet_path, device="cpu")288    else:289        unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.bin")290        unet_state_dict = torch.load(unet_path, map_location="cpu")291 292    if osp.exists(vae_path):293        vae_state_dict = load_file(vae_path, device="cpu")294    else:295        vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.bin")296        vae_state_dict = torch.load(vae_path, map_location="cpu")297 298    if osp.exists(text_enc_path):299        text_enc_dict = load_file(text_enc_path, device="cpu")300    else:301        text_enc_path = osp.join(args.model_path, "text_encoder", "pytorch_model.bin")302        text_enc_dict = torch.load(text_enc_path, map_location="cpu")303 304    # Convert the UNet model305    unet_state_dict = convert_unet_state_dict(unet_state_dict)306    unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}307 308    # Convert the VAE model309    vae_state_dict = convert_vae_state_dict(vae_state_dict)310    vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}311 312    # Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper313    is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict314 315    if is_v20_model:316        # Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm317        text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}318        text_enc_dict = convert_text_enc_state_dict_v20(text_enc_dict)319        text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}320    else:321        text_enc_dict = convert_text_enc_state_dict(text_enc_dict)322        text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}323 324    # Put together new checkpoint325    state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}326    if args.half:327        state_dict = {k: v.half() for k, v in state_dict.items()}328 329    if args.use_safetensors:330        save_file(state_dict, args.checkpoint_path)331    else:332        state_dict = {"state_dict": state_dict}333        torch.save(state_dict, args.checkpoint_path)334