convert_url_to_diffusers_flux_gr.py

import json
import torch
from safetensors.torch import load_file, save_file
from pathlib import Path
import gc
import gguf
from dequant import dequantize_tensor # https://github.com/city96/ComfyUI-GGUF

import os
import argparse
import gradio as gr
# also requires aria, gdown, peft, huggingface_hub, safetensors, transformers, accelerate, pytorch_lightning

import subprocess
subprocess.run('pip cache purge', shell=True)

import spaces
@spaces.GPU()
def spaces_dummy():
    pass

flux_dev_repo = "ChuckMcSneed/FLUX.1-dev"
flux_schnell_repo = "black-forest-labs/FLUX.1-schnell"
system_temp_dir = "temp"

device = "cuda" if torch.cuda.is_available() else "cpu"
torch.set_grad_enabled(False)

GGUF_QTYPE = [gguf.GGMLQuantizationType.Q8_0, gguf.GGMLQuantizationType.Q5_1,
    gguf.GGMLQuantizationType.Q5_0, gguf.GGMLQuantizationType.Q4_1,
    gguf.GGMLQuantizationType.Q4_0, gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16]

TORCH_DTYPE = [torch.float32, torch.float, torch.float64, torch.double, torch.float16, torch.half,
    torch.bfloat16, torch.complex32, torch.chalf, torch.complex64, torch.cfloat,
    torch.complex128, torch.cdouble, torch.uint8, torch.uint16, torch.uint32, torch.uint64,
    torch.int8, torch.int16, torch.short, torch.int32, torch.int, torch.int64, torch.long,
    torch.bool, torch.float8_e4m3fn, torch.float8_e5m2]

TORCH_QUANTIZED_DTYPE = [torch.quint8, torch.qint8, torch.qint32, torch.quint4x2]

def list_sub(a, b):
    return [e for e in a if e not in b]

def is_repo_name(s):
    import re
    return re.fullmatch(r'^[^/,\s\"\']+/[^/,\s\"\']+$', s)

def clear_cache():
    torch.cuda.empty_cache()
    gc.collect()

def clear_sd(sd: dict):
    for k in list(sd.keys()):
        sd.pop(k)
    del sd
    torch.cuda.empty_cache()
    gc.collect()

def clone_sd(sd: dict):
    from copy import deepcopy
    print("Cloning state dict.")
    for k in list(sd.keys()):
        sd[k] = deepcopy(sd.pop(k))
        #sd[k] = sd.pop(k).detach().clone().to(device="cpu")
    torch.cuda.empty_cache()
    gc.collect()

def print_resource_usage():
    import psutil
    cpu_usage = psutil.cpu_percent()
    ram_usage = psutil.virtual_memory().used / psutil.virtual_memory().total * 100
    print(f"CPU usage: {cpu_usage}% / RAM usage: {ram_usage}%")

def download_thing(directory, url, civitai_api_key="", progress=gr.Progress(track_tqdm=True)):
    progress(0, desc="Start downloading...")
    url = url.strip()
    if "drive.google.com" in url:
        original_dir = os.getcwd()
        os.chdir(directory)
        os.system(f"gdown --fuzzy {url}")
        os.chdir(original_dir)
    elif "huggingface.co" in url:
        url = url.replace("?download=true", "")
        if "/blob/" in url:
            url = url.replace("/blob/", "/resolve/")
            os.system(f"aria2c --console-log-level=error --summary-interval=10 -c -x 16 -k 1M -s 16 {url} -d {directory}  -o {url.split('/')[-1]}")
        else:
            os.system (f"aria2c --optimize-concurrent-downloads --console-log-level=error --summary-interval=10 -c -x 16 -k 1M -s 16 {url} -d {directory}  -o {url.split('/')[-1]}")
    elif "civitai.com" in url:
        if "?" in url:
            url = url.split("?")[0]
        if civitai_api_key:
            url = url + f"?token={civitai_api_key}"
            os.system(f"aria2c --console-log-level=error --summary-interval=10 -c -x 16 -k 1M -s 16 -d {directory} {url}")
        else:
            print("You need an API key to download Civitai models.")
    else:
        os.system(f"aria2c --console-log-level=error --summary-interval=10 -c -x 16 -k 1M -s 16 -d {directory} {url}")

def get_local_model_list(dir_path):
    model_list = []
    valid_extensions = ('.safetensors')
    for file in Path(dir_path).glob("*"):
        if file.suffix in valid_extensions:
            file_path = str(Path(f"{dir_path}/{file.name}"))
            model_list.append(file_path)
    return model_list

def get_download_file(temp_dir, url, civitai_key, progress=gr.Progress(track_tqdm=True)):
    if not "http" in url and is_repo_name(url) and not Path(url).exists():
        print(f"Use HF Repo: {url}")
        new_file = url
    elif not "http" in url and Path(url).exists():
        print(f"Use local file: {url}")
        new_file = url
    elif Path(f"{temp_dir}/{url.split('/')[-1]}").exists():
        print(f"File to download alreday exists: {url}")
        new_file = f"{temp_dir}/{url.split('/')[-1]}"
    else:
        print(f"Start downloading: {url}")
        before = get_local_model_list(temp_dir)
        try:
            download_thing(temp_dir, url.strip(), civitai_key)
        except Exception:
            print(f"Download failed: {url}")
            return ""
        after = get_local_model_list(temp_dir)
        new_file = list_sub(after, before)[0] if list_sub(after, before) else ""
    if not new_file:
        print(f"Download failed: {url}")
        return ""
    print(f"Download completed: {url}")
    return new_file

def save_readme_md(dir, url):
    orig_url = ""
    if "http" in url:
        orig_url = url
    if orig_url:
       md = f"""---
license: other
license_name: flux-1-dev-non-commercial-license
license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.
language:
- en
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-to-image
- Flux
---
Converted from [{orig_url}]({orig_url}).
"""
    else:
        md = f"""---
license: other
license_name: flux-1-dev-non-commercial-license
license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.
language:
- en
library_name: diffusers
pipeline_tag: text-to-image
tags:
- text-to-image
- Flux
---
"""
    path = str(Path(dir, "README.md"))
    with open(path, mode='w', encoding="utf-8") as f:
        f.write(md)

def is_repo_exists(repo_id):
    from huggingface_hub import HfApi
    api = HfApi()
    try:
        if api.repo_exists(repo_id=repo_id): return True
        else: return False
    except Exception as e:
        print(f"Error: Failed to connect {repo_id}. ")
        return True # for safe

def create_diffusers_repo(new_repo_id, diffusers_folder, is_private, is_overwrite, progress=gr.Progress(track_tqdm=True)):
    from huggingface_hub import HfApi
    import os
    hf_token = os.environ.get("HF_TOKEN")
    api = HfApi()
    try:
        progress(0, desc="Start uploading...")
        api.create_repo(repo_id=new_repo_id, token=hf_token, private=is_private, exist_ok=is_overwrite)
        for path in Path(diffusers_folder).glob("*"):
            if path.is_dir():
                api.upload_folder(repo_id=new_repo_id, folder_path=str(path), path_in_repo=path.name, token=hf_token)
            elif path.is_file():
                api.upload_file(repo_id=new_repo_id, path_or_fileobj=str(path), path_in_repo=path.name, token=hf_token)
        progress(1, desc="Uploaded.")
        url = f"https://huggingface.co/{new_repo_id}"
    except Exception as e:
        print(f"Error: Failed to upload to {new_repo_id}. ")
        print(e)
        return ""
    return url

# https://github.com/huggingface/diffusers/blob/main/scripts/convert_flux_to_diffusers.py 
# in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
# while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
with torch.no_grad(), torch.autocast(device):
    @torch.jit.script
    def swap_scale_shift(weight):
        shift, scale = weight.chunk(2, dim=0)
        new_weight = torch.cat([scale, shift], dim=0)
        return new_weight

with torch.no_grad(), torch.autocast(device):
    def convert_flux_transformer_checkpoint_to_diffusers(
        original_state_dict, num_layers, num_single_layers, inner_dim, mlp_ratio=4.0,
        progress=gr.Progress(track_tqdm=True)):
        def conv(cdict: dict, odict: dict, ckey: str, okey: str):
            if okey in odict.keys():
                progress(0, desc=f"Converting {okey} => {ckey}")
                print(f"Converting {okey} => {ckey}")
                cdict[ckey] = odict.pop(okey)
                gc.collect()

        def convswap(cdict: dict, odict: dict, ckey: str, okey: str):
            if okey in odict.keys():
                progress(0, desc=f"Converting (swap) {okey} => {ckey}")
                print(f"Converting {okey} => {ckey} (swap)")
                cdict[ckey] = swap_scale_shift(odict.pop(okey))
                gc.collect()

        def convqkv(cdict: dict, odict: dict, i: int):
            keys = odict.keys()
            if (f"double_blocks.{i}.img_attn.qkv.weight" in keys or f"double_blocks.{i}.txt_attn.qkv.weight" in keys\
                or f"double_blocks.{i}.img_attn.qkv.bias" in keys or f"double_blocks.{i}.txt_attn.qkv.bias" in keys)\
                and (f"double_blocks.{i}.img_attn.qkv.weight" not in keys or f"double_blocks.{i}.txt_attn.qkv.weight" not in keys\
                     or f"double_blocks.{i}.img_attn.qkv.bias" not in keys or f"double_blocks.{i}.txt_attn.qkv.bias" not in keys):
                progress(0, desc=f"Key error in converting Q, K, V (double_blocks.{i}).")
                print(f"Key error in converting Q, K, V (double_blocks.{i}).")
                return
            progress(0, desc=f"Converting Q, K, V (double_blocks.{i}).")
            print(f"Converting Q, K, V (double_blocks.{i}).")
            sample_q, sample_k, sample_v = torch.chunk(
                odict.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0
            )
            context_q, context_k, context_v = torch.chunk(
                odict.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0
            )
            sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
                odict.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0
            )
            context_q_bias, context_k_bias, context_v_bias = torch.chunk(
                odict.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0
            )
            cdict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q])
            cdict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias])
            cdict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k])
            cdict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias])
            cdict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v])
            cdict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias])
            cdict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q])
            cdict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias])
            cdict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k])
            cdict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias])
            cdict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v])
            cdict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias])
            gc.collect()

        def convqkvmlp(cdict: dict, odict: dict, i: int, inner_dim: int, mlp_ratio: float):
            keys = odict.keys()
            if (f"single_blocks.{i}.linear1.weight" in keys or f"single_blocks.{i}.linear1.bias" in keys)\
                and (f"single_blocks.{i}.linear1.weight" not in keys or f"single_blocks.{i}.linear1.bias" not in keys):
                progress(0, desc=f"Key error in converting Q, K, V, mlp (single_blocks.{i}).")
                print(f"Key error in converting Q, K, V, mlp (single_blocks.{i}).")
                return
            progress(0, desc=f"Converting Q, K, V, mlp (single_blocks.{i}).")
            print(f"Converting Q, K, V, mlp (single_blocks.{i}).")
            mlp_hidden_dim = int(inner_dim * mlp_ratio)
            split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
            q, k, v, mlp = torch.split(odict.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0)
            q_bias, k_bias, v_bias, mlp_bias = torch.split(
                odict.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0
            )
            cdict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q])
            cdict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias])
            cdict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k])
            cdict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias])
            cdict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v])
            cdict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias])
            cdict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp])
            cdict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias])
            gc.collect()

        converted_state_dict = {}
        progress(0, desc="Converting FLUX.1 state dict to Diffusers format.")

        ## time_text_embed.timestep_embedder <-  time_in
        conv(converted_state_dict, original_state_dict, "time_text_embed.timestep_embedder.linear_1.weight", "time_in.in_layer.weight")
        conv(converted_state_dict, original_state_dict, "time_text_embed.timestep_embedder.linear_1.bias", "time_in.in_layer.bias")
        conv(converted_state_dict, original_state_dict, "time_text_embed.timestep_embedder.linear_2.weight", "time_in.out_layer.weight")
        conv(converted_state_dict, original_state_dict, "time_text_embed.timestep_embedder.linear_2.bias", "time_in.out_layer.bias")

        ## time_text_embed.text_embedder <- vector_in
        conv(converted_state_dict, original_state_dict, "time_text_embed.text_embedder.linear_1.weight", "vector_in.in_layer.weight")
        conv(converted_state_dict, original_state_dict, "time_text_embed.text_embedder.linear_1.bias", "vector_in.in_layer.bias")
        conv(converted_state_dict, original_state_dict, "time_text_embed.text_embedder.linear_2.weight", "vector_in.out_layer.weight")
        conv(converted_state_dict, original_state_dict, "time_text_embed.text_embedder.linear_2.bias", "vector_in.out_layer.bias")

        # guidance
        has_guidance = any("guidance" in k for k in original_state_dict)
        if has_guidance:
            conv(converted_state_dict, original_state_dict, "time_text_embed.guidance_embedder.linear_1.weight", "guidance_in.in_layer.weight")
            conv(converted_state_dict, original_state_dict, "time_text_embed.guidance_embedder.linear_1.bias", "guidance_in.in_layer.bias")
            conv(converted_state_dict, original_state_dict, "time_text_embed.guidance_embedder.linear_2.weight", "guidance_in.out_layer.weight")
            conv(converted_state_dict, original_state_dict, "time_text_embed.guidance_embedder.linear_2.bias", "guidance_in.out_layer.bias")

        # context_embedder
        conv(converted_state_dict, original_state_dict, "context_embedder.weight", "txt_in.weight")
        conv(converted_state_dict, original_state_dict, "context_embedder.bias", "txt_in.bias")

        # x_embedder
        conv(converted_state_dict, original_state_dict, "x_embedder.weight", "img_in.weight")
        conv(converted_state_dict, original_state_dict, "x_embedder.bias", "img_in.bias")

        progress(0.25, desc="Converting FLUX.1 state dict to Diffusers format.")
        # double transformer blocks
        for i in range(num_layers):
            block_prefix = f"transformer_blocks.{i}."
            # norms.
            ## norm1
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm1.linear.weight", f"double_blocks.{i}.img_mod.lin.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm1.linear.bias", f"double_blocks.{i}.img_mod.lin.bias")
            ## norm1_context
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm1_context.linear.weight", f"double_blocks.{i}.txt_mod.lin.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm1_context.linear.bias", f"double_blocks.{i}.txt_mod.lin.bias")
            # Q, K, V
            convqkv(converted_state_dict, original_state_dict, i)
            # qk_norm
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_q.weight", f"double_blocks.{i}.img_attn.norm.query_norm.scale")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_k.weight", f"double_blocks.{i}.img_attn.norm.key_norm.scale")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_added_q.weight", f"double_blocks.{i}.txt_attn.norm.query_norm.scale")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_added_k.weight", f"double_blocks.{i}.txt_attn.norm.key_norm.scale")
            # ff img_mlp
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff.net.0.proj.weight", f"double_blocks.{i}.img_mlp.0.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff.net.0.proj.bias", f"double_blocks.{i}.img_mlp.0.bias")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff.net.2.weight", f"double_blocks.{i}.img_mlp.2.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff.net.2.bias", f"double_blocks.{i}.img_mlp.2.bias")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff_context.net.0.proj.weight", f"double_blocks.{i}.txt_mlp.0.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff_context.net.0.proj.bias", f"double_blocks.{i}.txt_mlp.0.bias")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff_context.net.2.weight", f"double_blocks.{i}.txt_mlp.2.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}ff_context.net.2.bias", f"double_blocks.{i}.txt_mlp.2.bias")
            # output projections.
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.to_out.0.weight", f"double_blocks.{i}.img_attn.proj.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.to_out.0.bias", f"double_blocks.{i}.img_attn.proj.bias")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.to_add_out.weight", f"double_blocks.{i}.txt_attn.proj.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.to_add_out.bias", f"double_blocks.{i}.txt_attn.proj.bias")

        progress(0.5, desc="Converting FLUX.1 state dict to Diffusers format.")
        # single transfomer blocks
        for i in range(num_single_layers):
            block_prefix = f"single_transformer_blocks.{i}."
            # norm.linear  <- single_blocks.0.modulation.lin
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm.linear.weight", f"single_blocks.{i}.modulation.lin.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}norm.linear.bias", f"single_blocks.{i}.modulation.lin.bias")
            # Q, K, V, mlp
            convqkvmlp(converted_state_dict, original_state_dict, i, inner_dim, mlp_ratio)
            # qk norm
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_q.weight", f"single_blocks.{i}.norm.query_norm.scale")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}attn.norm_k.weight", f"single_blocks.{i}.norm.key_norm.scale")
            # output projections.
            conv(converted_state_dict, original_state_dict, f"{block_prefix}proj_out.weight", f"single_blocks.{i}.linear2.weight")
            conv(converted_state_dict, original_state_dict, f"{block_prefix}proj_out.bias", f"single_blocks.{i}.linear2.bias")

        progress(0.75, desc="Converting FLUX.1 state dict to Diffusers format.")
        conv(converted_state_dict, original_state_dict, "proj_out.weight", "final_layer.linear.weight")
        conv(converted_state_dict, original_state_dict, "proj_out.bias", "final_layer.linear.bias")
        convswap(converted_state_dict, original_state_dict, "norm_out.linear.weight", "final_layer.adaLN_modulation.1.weight")
        convswap(converted_state_dict, original_state_dict, "norm_out.linear.bias", "final_layer.adaLN_modulation.1.bias")

        progress(1, desc="Converting FLUX.1 state dict to Diffusers format.")
        return converted_state_dict

# read safetensors metadata
def read_safetensors_metadata(path):
    with open(path, 'rb') as f:       
        header_size = int.from_bytes(f.read(8), 'little')
        header_json = f.read(header_size).decode('utf-8')
        header = json.loads(header_json)
        metadata = header.get('__metadata__', {})
        return metadata.copy()

def normalize_key(k: str):
    return k.replace("vae.", "").replace("model.diffusion_model.", "")\
        .replace("text_encoders.clip_l.transformer.", "")\
        .replace("text_encoders.t5xxl.transformer.", "")

def load_json_list(path: str):
    try:
        with open(path, encoding='utf-8') as f:
            return list(json.load(f))
    except Exception as e:
        print(e)
        return []

# https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/modeling_utils.py
# https://huggingface.co/docs/huggingface_hub/v0.24.5/package_reference/serialization
# https://huggingface.co/docs/huggingface_hub/index
with torch.no_grad():
    def to_safetensors(sd: dict, path: str, pattern: str, size: str, progress=gr.Progress(track_tqdm=True)):
        from huggingface_hub import save_torch_state_dict
        print(f"Saving a temporary file to disk: {path}")
        os.makedirs(path, exist_ok=True)
        try:
            for k, v in sd.items():
                sd[k] = v.to(device="cpu")
            save_torch_state_dict(sd, path, filename_pattern=pattern, max_shard_size=size)
        except Exception as e:
            print(e)

# https://discuss.huggingface.co/t/t5forconditionalgeneration-checkpoint-size-mismatch-19418/24119
# https://github.com/huggingface/transformers/issues/13769
# https://github.com/huggingface/optimum-quanto/issues/278
# https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/serialization/_torch.py
# https://huggingface.co/docs/accelerate/usage_guides/big_modeling
with torch.no_grad():
    def to_safetensors_flux_module(sd: dict, path: str, pattern: str, size: str,
                                        quantization: bool=False, name: str = "",
                                        metadata: dict | None = None, progress=gr.Progress(track_tqdm=True)):
        from huggingface_hub import save_torch_state_dict, save_torch_model
        from accelerate import init_empty_weights
        try:
            progress(0, desc=f"Preparing to save FLUX.1 {name} to Diffusers format.")
            print(f"Preparing to save FLUX.1 {name} to Diffusers format.")
            for k, v in sd.items():
                sd[k] = v.to(device="cpu")
            progress(0, desc=f"Loading FLUX.1 {name}.")
            print(f"Loading FLUX.1 {name}.")
            os.makedirs(path, exist_ok=True)
            if quantization:
                progress(0.5, desc=f"Saving quantized FLUX.1 {name} to {path}")
                print(f"Saving quantized FLUX.1 {name} to {path}")
            else:
                progress(0.5, desc=f"Saving FLUX.1 {name} to: {path}")
                if False and path.endswith("/transformer"):
                    from diffusers import FluxTransformer2DModel
                    has_guidance = any("guidance" in k for k in sd)
                    with init_empty_weights():
                        model = FluxTransformer2DModel(guidance_embeds=has_guidance)
                    model.to("cpu")
                    model.load_state_dict(sd, strict=True)
                    print(f"Saving FLUX.1 {name} to: {path} (FluxTransformer2DModel)")
                    if metadata is not None:
                        progress(0.5, desc=f"Saving FLUX.1 {name} metadata to: {path}")
                        save_torch_model(model=model, save_directory=path,
                                            filename_pattern=pattern, max_shard_size=size, metadata=metadata)
                    else:
                        save_torch_model(model=model, save_directory=path,
                                        filename_pattern=pattern, max_shard_size=size)
                else:
                    print(f"Saving FLUX.1 {name} to: {path}")
                    if metadata is not None:
                        progress(0.5, desc=f"Saving FLUX.1 {name} metadata to: {path}")
                        save_torch_state_dict(state_dict=sd, save_directory=path,
                                            filename_pattern=pattern, max_shard_size=size, metadata=metadata)
                    else:
                        save_torch_state_dict(state_dict=sd, save_directory=path,
                                        filename_pattern=pattern, max_shard_size=size)
                progress(1, desc=f"Saved FLUX.1 {name} to: {path}")
                print(f"Saved FLUX.1 {name} to: {path}")
        except Exception as e:
            print(e)
        finally:
            gc.collect()

flux_transformer_json = "flux_transformer_keys.json"
flux_t5xxl_json = "flux_t5xxl_keys.json"
flux_clip_json = "flux_clip_keys.json"
flux_vae_json = "flux_vae_keys.json"
keys_flux_t5xxl = set(load_json_list(flux_t5xxl_json))
keys_flux_transformer = set(load_json_list(flux_transformer_json))
keys_flux_clip = set(load_json_list(flux_clip_json))
keys_flux_vae = set(load_json_list(flux_vae_json))

with torch.no_grad():
    def dequant_tensor(v: torch.Tensor, dtype: torch.dtype, dequant: bool):
        try:
            #print(f"shape: {v.shape} / dim: {v.ndim}")
            if dequant:
                qtype = v.tensor_type
                if v.dtype in TORCH_DTYPE: return v.to(dtype) if v.dtype != dtype else v
                elif qtype in GGUF_QTYPE: return dequantize_tensor(v, dtype)
                elif torch.dtype in TORCH_QUANTIZED_DTYPE: return torch.dequantize(v).to(dtype)
                else: return torch.dequantize(v).to(dtype)
            else: return v.to(dtype) if v.dtype != dtype else v
        except Exception as e:
            print(e)

with torch.no_grad():
    def normalize_flux_state_dict(path: str, savepath: str, dtype: torch.dtype = torch.bfloat16,
                                dequant: bool = False, progress=gr.Progress(track_tqdm=True)):
        progress(0, desc=f"Loading and normalizing FLUX.1 safetensors: {path}")
        print(f"Loading and normalizing FLUX.1 safetensors: {path}")
        new_sd = dict()
        state_dict = load_file(path, device="cpu")
        try:
            for k in list(state_dict.keys()):
                v = state_dict.pop(k)
                nk = normalize_key(k)
                print(f"{k} => {nk}") #
                new_sd[nk] = dequant_tensor(v, dtype, dequant)
        except Exception as e:
            print(e)
            return
        finally:
            clear_sd(state_dict)
        new_path = str(Path(savepath, Path(path).stem + "_fixed" + Path(path).suffix))
        metadata = read_safetensors_metadata(path)
        progress(0.5, desc=f"Saving FLUX.1 safetensors: {new_path}")
        print(f"Saving FLUX.1 safetensors: {new_path}")
        os.makedirs(savepath, exist_ok=True)
        save_file(new_sd, new_path, metadata={"format": "pt", **metadata})
        progress(1, desc=f"Saved FLUX.1 safetensors: {new_path}")
        print(f"Saved FLUX.1 safetensors: {new_path}")
        clear_sd(new_sd)

with torch.no_grad():
    def extract_norm_flux_module_sd(path: str, dtype: torch.dtype = torch.bfloat16,
                                        dequant: bool = False, name: str = "", keys: set = {},
                                        progress=gr.Progress(track_tqdm=True)):
        progress(0, desc=f"Loading and normalizing FLUX.1 {name} safetensors: {path}")
        print(f"Loading and normalizing FLUX.1 {name} safetensors: {path}")
        new_sd = dict()
        state_dict = load_file(path, device="cpu")
        try:
            for k in list(state_dict.keys()):
                if k not in keys: state_dict.pop(k)
            gc.collect()
            for k in list(state_dict.keys()):
                v = state_dict.pop(k)
                if k in keys:
                    nk = normalize_key(k)
                    progress(0.5, desc=f"{k} => {nk}") #
                    print(f"{k} => {nk}") #
                    new_sd[nk] = dequant_tensor(v, dtype, dequant)
                    #print_resource_usage() #
        except Exception as e:
            print(e)
            return None
        finally:
            progress(1, desc=f"Normalized FLUX.1 {name} safetensors: {path}")
            print(f"Normalized FLUX.1 {name} safetensors: {path}")
            clear_sd(state_dict)
        return new_sd

with torch.no_grad():
    def convert_flux_transformer_sd_to_diffusers(sd: dict, progress=gr.Progress(track_tqdm=True)):
        progress(0, desc="Converting FLUX.1 state dict to Diffusers format.")
        print("Converting FLUX.1 state dict to Diffusers format.")
        num_layers = 19
        num_single_layers = 38
        inner_dim = 3072
        mlp_ratio = 4.0
        try:
            sd = convert_flux_transformer_checkpoint_to_diffusers(
                sd, num_layers, num_single_layers, inner_dim, mlp_ratio=mlp_ratio
            )
        except Exception as e:
            print(e)
        finally:
            progress(1, desc="Converted FLUX.1 state dict to Diffusers format.")
            print("Converted FLUX.1 state dict to Diffusers format.")
            gc.collect()
        return sd

with torch.no_grad():
    def load_sharded_safetensors(path: str):
        import glob
        sd = {}
        try:
            for filepath in glob.glob(f"{path}/*.safetensors"):
                sharded_sd = load_file(str(filepath), device="cpu")
                for k, v in sharded_sd.items():
                    sharded_sd[k] = v.to(device="cpu")
                sd = sd | sharded_sd.copy()
                clear_sd(sharded_sd)
        except Exception as e:
                print(e)
        return sd
    
# https://huggingface.co/docs/safetensors/api/torch
with torch.no_grad():
    def convert_flux_transformer_sd_to_diffusers_sharded(sd: dict, path: str, pattern: str,
                                                size: str, progress=gr.Progress(track_tqdm=True)):
        from huggingface_hub import save_torch_state_dict#, load_torch_model
        import glob
        try:
            progress(0, desc=f"Saving temporary files to disk: {path}")
            print(f"Saving temporary files to disk: {path}")
            os.makedirs(path, exist_ok=True)
            for k, v in sd.items():
                if k in set(keys_flux_transformer): sd[k] = v.to(device="cpu")
            save_torch_state_dict(sd, path, filename_pattern=pattern, max_shard_size=size)
            clear_sd(sd)
            progress(0.25, desc=f"Saved temporary files to disk: {path}")
            print(f"Saved temporary files to disk: {path}")
            for filepath in glob.glob(f"{path}/*.safetensors"):
                progress(0.25, desc=f"Processing temporary files: {str(filepath)}")
                print(f"Processing temporary files: {str(filepath)}")
                sharded_sd = load_file(str(filepath), device="cpu")
                sharded_sd = convert_flux_transformer_sd_to_diffusers(sharded_sd)
                for k, v in sharded_sd.items():
                    sharded_sd[k] = v.to(device="cpu")
                save_file(sharded_sd, str(filepath))
                clear_sd(sharded_sd)
            print(f"Loading temporary files from disk: {path}")
            sd = load_sharded_safetensors(path)
            print(f"Loaded temporary files from disk: {path}")
        except Exception as e:
            print(e)
        return sd

with torch.no_grad():
    def extract_normalized_flux_state_dict_sharded(loadpath: str, dtype: torch.dtype,
         dequant: bool, path: str, pattern: str, size: str, progress=gr.Progress(track_tqdm=True)):
        from huggingface_hub import save_torch_state_dict#, load_torch_model
        import glob
        try:
            progress(0, desc=f"Loading model file: {loadpath}")
            print(f"Loading model file: {loadpath}")
            sd = load_file(loadpath, device="cpu")
            progress(0, desc=f"Saving temporary files to disk: {path}")
            print(f"Saving temporary files to disk: {path}")
            os.makedirs(path, exist_ok=True)
            for k, v in sd.items():
                sd[k] = v.to(device="cpu")
            save_torch_state_dict(sd, path, filename_pattern=pattern, max_shard_size=size)
            clear_sd(sd)
            progress(0.25, desc=f"Saved temporary files to disk: {path}")
            print(f"Saved temporary files to disk: {path}")
            for filepath in glob.glob(f"{path}/*.safetensors"):
                progress(0.25, desc=f"Processing temporary files: {str(filepath)}")
                print(f"Processing temporary files: {str(filepath)}")
                sharded_sd = extract_norm_flux_module_sd(str(filepath), dtype, dequant,
                                                         "Transformer", keys_flux_transformer)
                for k, v in sharded_sd.items():
                    sharded_sd[k] = v.to(device="cpu")
                save_file(sharded_sd, str(filepath))
                clear_sd(sharded_sd)
                print(f"Processed temporary files: {str(filepath)}")
            print(f"Loading temporary files from disk: {path}")
            sd = load_sharded_safetensors(path)
            print(f"Loaded temporary files from disk: {path}")
        except Exception as e:
            print(e)
        return sd

def download_repo(repo_name, path, use_original=["vae", "text_encoder"], progress=gr.Progress(track_tqdm=True)):
    from huggingface_hub import snapshot_download
    print(f"Downloading {repo_name}.")
    try:
        if "text_encoder_2" in use_original:
            snapshot_download(repo_id=repo_name, local_dir=path, ignore_patterns=["transformer/diffusion*.*", "*.sft", ".*", "README*", "*.md", "*.index", "*.jpg", "*.png", "*.webp"])
        else:
            snapshot_download(repo_id=repo_name, local_dir=path, ignore_patterns=["transformer/diffusion*.*", "text_encoder_2/model*.*", "*.sft", ".*", "README*", "*.md", "*.index", "*.jpg", "*.png", "*.webp"])
    except Exception as e:
        print(e)

def copy_nontensor_files(from_path, to_path, use_original=["vae", "text_encoder"]):
    import shutil
    if "text_encoder_2" in use_original:
        te_from = str(Path(from_path, "text_encoder_2"))
        te_to = str(Path(to_path, "text_encoder_2"))
        print(f"Copying Text Encoder 2 files {te_from} to {te_to}")
        shutil.copytree(te_from, te_to, ignore=shutil.ignore_patterns(".*", "README*", "*.md", "*.jpg", "*.png", "*.webp"), dirs_exist_ok=True)
    if "text_encoder" in use_original:
        te1_from = str(Path(from_path, "text_encoder"))
        te1_to = str(Path(to_path, "text_encoder"))
        print(f"Copying Text Encoder 1 files {te1_from} to {te1_to}")
        shutil.copytree(te1_from, te1_to, ignore=shutil.ignore_patterns(".*", "README*", "*.md", "*.jpg", "*.png", "*.webp"), dirs_exist_ok=True)
    if "vae" in use_original:
        vae_from = str(Path(from_path, "vae"))
        vae_to = str(Path(to_path, "vae"))
        print(f"Copying VAE files {vae_from} to {vae_to}")
        shutil.copytree(vae_from, vae_to, ignore=shutil.ignore_patterns(".*", "README*", "*.md", "*.jpg", "*.png", "*.webp"), dirs_exist_ok=True)
    tn2_from = str(Path(from_path, "tokenizer_2"))
    tn2_to = str(Path(to_path, "tokenizer_2"))
    print(f"Copying Tokenizer 2 files {tn2_from} to {tn2_to}")
    shutil.copytree(tn2_from, tn2_to, ignore=shutil.ignore_patterns(".*", "README*", "*.md", "*.jpg", "*.png", "*.webp"), dirs_exist_ok=True)
    print(f"Copying non-tensor files {from_path} to {to_path}")
    shutil.copytree(from_path, to_path, ignore=shutil.ignore_patterns("*.safetensors", "*.bin", "*.sft", ".*", "README*", "*.md", "*.index", "*.jpg", "*.png", "*.webp", "*.index.json"), dirs_exist_ok=True)
    
def save_flux_other_diffusers(path: str, model_type: str = "dev", use_original: list = ["vae", "text_encoder"], progress=gr.Progress(track_tqdm=True)):
    import shutil
    progress(0, desc="Loading FLUX.1 Components.")
    print("Loading FLUX.1 Components.")
    temppath = system_temp_dir
    if model_type == "schnell": repo = flux_schnell_repo
    else: repo = flux_dev_repo
    os.makedirs(temppath, exist_ok=True)
    os.makedirs(path, exist_ok=True)
    download_repo(repo, temppath, use_original)
    progress(0.5, desc="Saving FLUX.1 Components.")
    print("Saving FLUX.1 Components.")
    copy_nontensor_files(temppath, path, use_original)
    shutil.rmtree(temppath)

with torch.no_grad():
    def fix_flux_safetensors(loadpath: str, savepath: str, dtype: torch.dtype = torch.bfloat16,
                          quantization: bool = False, model_type: str = "dev", dequant: bool = False):
        save_flux_other_diffusers(savepath, model_type)
        normalize_flux_state_dict(loadpath, savepath, dtype, dequant)
        clear_cache()

with torch.no_grad(): # Much lower memory consumption, but higher disk load
    def flux_to_diffusers_lowmem(loadpath: str, savepath: str, dtype: torch.dtype = torch.bfloat16,
                          quantization: bool = False, model_type: str = "dev",
                          dequant: bool = False, use_original: list = ["vae", "text_encoder"],
                          new_repo_id: str = "", local: bool = False, progress=gr.Progress(track_tqdm=True)):
        unet_sd_path = savepath.removesuffix("/") + "/transformer"
        unet_sd_pattern = "diffusion_pytorch_model{suffix}.safetensors"
        unet_sd_size = "9.5GB"
        te_sd_path = savepath.removesuffix("/") + "/text_encoder_2"
        te_sd_pattern = "model{suffix}.safetensors"
        te_sd_size = "5GB"
        clip_sd_path = savepath.removesuffix("/") + "/text_encoder"
        clip_sd_pattern = "model{suffix}.safetensors"
        clip_sd_size = "9.5GB"
        vae_sd_path = savepath.removesuffix("/") + "/vae"
        vae_sd_pattern = "diffusion_pytorch_model{suffix}.safetensors"
        vae_sd_size = "9.5GB"
        print_resource_usage() #
        metadata = {"format": "pt", **read_safetensors_metadata(loadpath)}
        clear_cache()
        print_resource_usage() #
        if "vae" not in use_original:
            vae_sd = extract_norm_flux_module_sd(loadpath, torch.bfloat16, dequant, "VAE",
                                                keys_flux_vae)
            to_safetensors_flux_module(vae_sd, vae_sd_path, vae_sd_pattern, vae_sd_size,
                                       quantization, "VAE", None)
            clear_sd(vae_sd)
            print_resource_usage() #
        if "text_encoder" not in use_original:
            clip_sd = extract_norm_flux_module_sd(loadpath, torch.bfloat16, dequant, "Text Encoder",
                                                keys_flux_clip)
            to_safetensors_flux_module(clip_sd, clip_sd_path, clip_sd_pattern, clip_sd_size,
                                       quantization, "Text Encoder", None)
            clear_sd(clip_sd)
            print_resource_usage() #
        if "text_encoder_2" not in use_original:
            te_sd = extract_norm_flux_module_sd(loadpath, dtype, dequant, "Text Encoder 2",
                                                keys_flux_t5xxl)
            to_safetensors_flux_module(te_sd, te_sd_path, te_sd_pattern, te_sd_size,
                                       quantization, "Text Encoder 2", None)
            clear_sd(te_sd)
            print_resource_usage() #
        unet_sd = extract_norm_flux_module_sd(loadpath, dtype, dequant, "Transformer",
                                              keys_flux_transformer)
        clear_cache()
        print_resource_usage() #
        if not local:
            os.remove(loadpath)
            print("Deleted downloaded file.")
        clear_cache()
        print_resource_usage() #
        unet_sd = convert_flux_transformer_sd_to_diffusers(unet_sd)
        clear_cache()
        print_resource_usage() #
        to_safetensors_flux_module(unet_sd, unet_sd_path, unet_sd_pattern, unet_sd_size,
                                        quantization, "Transformer", metadata)
        clear_sd(unet_sd)
        print_resource_usage() #
        save_flux_other_diffusers(savepath, model_type, use_original)
        print_resource_usage() #

with torch.no_grad(): # lowest memory consumption, but higheest disk load
    def flux_to_diffusers_lowmem2(loadpath: str, savepath: str, dtype: torch.dtype = torch.bfloat16,
                          quantization: bool = False, model_type: str = "dev",
                          dequant: bool = False, use_original: list = ["vae", "text_encoder"],
                          new_repo_id: str = "", progress=gr.Progress(track_tqdm=True)):
        unet_sd_path = savepath.removesuffix("/") + "/transformer"
        unet_temp_path = system_temp_dir.removesuffix("/") + "/sharded"
        unet_sd_pattern = "diffusion_pytorch_model{suffix}.safetensors"
        unet_sd_size = "10GB"
        unet_temp_size = "5GB"
        te_sd_path = savepath.removesuffix("/") + "/text_encoder_2"
        te_sd_pattern = "model{suffix}.safetensors"
        te_sd_size = "5GB"
        clip_sd_path = savepath.removesuffix("/") + "/text_encoder"
        clip_sd_pattern = "model{suffix}.safetensors"
        clip_sd_size = "10GB"
        vae_sd_path = savepath.removesuffix("/") + "/vae"
        vae_sd_pattern = "diffusion_pytorch_model{suffix}.safetensors"
        vae_sd_size = "10GB"
        print_resource_usage() #
        metadata = {"format": "pt", **read_safetensors_metadata(loadpath)}
        clear_cache()
        print_resource_usage() #
        if "vae" not in use_original:
            vae_sd = extract_norm_flux_module_sd(loadpath, torch.bfloat16, dequant, "VAE",
                                                keys_flux_vae)
            to_safetensors_flux_module(vae_sd, vae_sd_path, vae_sd_pattern, vae_sd_size,
                                       quantization, "VAE", None)
            clear_sd(vae_sd)
            print_resource_usage() #
        if "text_encoder" not in use_original:
            clip_sd = extract_norm_flux_module_sd(loadpath, torch.bfloat16, dequant, "Text Encoder",
                                                keys_flux_clip)
            to_safetensors_flux_module(clip_sd, clip_sd_path, clip_sd_pattern, clip_sd_size,
                                       quantization, "Text Encoder", None)
            clear_sd(clip_sd)
            print_resource_usage() #
        if "text_encoder_2" not in use_original:
            te_sd = extract_norm_flux_module_sd(loadpath, dtype, dequant, "Text Encoder 2",
                                                keys_flux_t5xxl)
            to_safetensors_flux_module(te_sd, te_sd_path, te_sd_pattern, te_sd_size,
                                       quantization, "Text Encoder 2", None)
            clear_sd(te_sd)
            print_resource_usage() #
        unet_sd = extract_normalized_flux_state_dict_sharded(loadpath, dtype, dequant,
                                        unet_temp_path, unet_sd_pattern, unet_temp_size)
        clear_cache()
        print_resource_usage() #
        unet_sd = convert_flux_transformer_sd_to_diffusers_sharded(unet_sd, unet_temp_path,
                                                                   unet_sd_pattern, unet_temp_size)
        clear_cache()
        print_resource_usage() #
        to_safetensors_flux_module(unet_sd, unet_sd_path, unet_sd_pattern, unet_sd_size,
                                        quantization, "Transformer", metadata)
        clear_sd(unet_sd)
        print_resource_usage() #
        save_flux_other_diffusers(savepath, model_type, use_original)
        print_resource_usage() #

def convert_url_to_diffusers_flux(url, civitai_key="", is_upload_sf=False, data_type="bf16",
                                  model_type="dev", dequant=False, use_original=["vae", "text_encoder"],
                                  hf_user="", hf_repo="", q=None, progress=gr.Progress(track_tqdm=True)):
    progress(0, desc="Start converting...")
    temp_dir = "."
    print_resource_usage() #
    new_file = get_download_file(temp_dir, url, civitai_key)
    if not new_file:
        print(f"Not found: {url}")
        return ""
    new_repo_name = Path(new_file).stem.replace(" ", "_").replace(",", "_").replace(".", "_") #

    dtype = torch.bfloat16
    quantization = False
    if data_type == "fp8": dtype = torch.float8_e4m3fn
    elif data_type == "fp16": dtype = torch.float16
    elif data_type == "qfloat8":
        dtype = torch.bfloat16
        quantization = True
    else: dtype = torch.bfloat16

    new_repo_id = f"{hf_user}/{Path(new_repo_name).stem}"
    if hf_repo != "": new_repo_id = f"{hf_user}/{hf_repo}"
    flux_to_diffusers_lowmem(new_file, new_repo_name, dtype, quantization, model_type, dequant, use_original, new_repo_id)

    """if is_upload_sf:
        import shutil
        shutil.move(str(Path(new_file).resolve()), str(Path(new_repo_name, Path(new_file).name).resolve()))
    else: os.remove(new_file)"""

    progress(1, desc="Converted.")
    q.put(new_repo_name)
    return new_repo_name

def convert_url_to_fixed_flux_safetensors(url, civitai_key="", is_upload_sf=False, data_type="bf16",
                                  model_type="dev", dequant=False, q=None, progress=gr.Progress(track_tqdm=True)):
    progress(0, desc="Start converting...")
    temp_dir = "."
    print_resource_usage() #
    new_file = get_download_file(temp_dir, url, civitai_key)
    if not new_file:
        print(f"Not found: {url}")
        return ""
    new_repo_name = Path(new_file).stem.replace(" ", "_").replace(",", "_").replace(".", "_") #

    dtype = torch.bfloat16
    quantization = False
    if data_type == "fp8": dtype = torch.float8_e4m3fn
    elif data_type == "fp16": dtype = torch.float16
    elif data_type == "qfloat8":
        dtype = torch.bfloat16
        quantization = True
    else: dtype = torch.bfloat16

    fix_flux_safetensors(new_file, new_repo_name, dtype, model_type, dequant)

    os.remove(new_file)

    progress(1, desc="Converted.")
    q.put(new_repo_name)
    return new_repo_name

def convert_url_to_diffusers_repo_flux(dl_url, hf_user, hf_repo, hf_token, civitai_key="", is_private=True, is_overwrite=False,
                                       is_upload_sf=False, data_type="bf16", model_type="dev", dequant=False,
                                       repo_urls=[], fix_only=False, use_original=["vae", "text_encoder"],
                                       progress=gr.Progress(track_tqdm=True)):
    import multiprocessing as mp
    import shutil
    if not hf_user:
        print(f"Invalid user name: {hf_user}")
        progress(1, desc=f"Invalid user name: {hf_user}")
        return gr.update(value=repo_urls, choices=repo_urls), gr.update(value="")
    if hf_token and not os.environ.get("HF_TOKEN"): os.environ['HF_TOKEN'] = hf_token
    if not civitai_key and os.environ.get("CIVITAI_API_KEY"): civitai_key = os.environ.get("CIVITAI_API_KEY")
    q = mp.Queue()
    if fix_only:
        p = mp.Process(target=convert_url_to_fixed_flux_safetensors, args=(dl_url, civitai_key,
                    is_upload_sf, data_type, model_type, dequant, q))
        #new_path = convert_url_to_fixed_flux_safetensors(dl_url, civitai_key, is_upload_sf, data_type, model_type, dequant)
    else:
        p = mp.Process(target=convert_url_to_diffusers_flux, args=(dl_url, civitai_key,
                    is_upload_sf, data_type, model_type, dequant, use_original, hf_user, hf_repo, q))
        #new_path = convert_url_to_diffusers_flux(dl_url, civitai_key, is_upload_sf, data_type, model_type, dequant)
    p.start()
    new_path = q.get()
    p.join()
    if not new_path: return ""
    new_repo_id = f"{hf_user}/{Path(new_path).stem}"
    if hf_repo != "": new_repo_id = f"{hf_user}/{hf_repo}"
    if not is_repo_name(new_repo_id):
        print(f"Invalid repo name: {new_repo_id}")
        progress(1, desc=f"Invalid repo name: {new_repo_id}")
        return gr.update(value=repo_urls, choices=repo_urls), gr.update(value="")
    if not is_overwrite and is_repo_exists(new_repo_id):
        print(f"Repo already exists: {new_repo_id}")
        progress(1, desc=f"Repo already exists: {new_repo_id}")
        return gr.update(value=repo_urls, choices=repo_urls), gr.update(value="")
    #save_readme_md(new_path, dl_url)
    repo_url = create_diffusers_repo(new_repo_id, new_path, is_private, is_overwrite)
    shutil.rmtree(new_path)
    if not repo_urls: repo_urls = []
    repo_urls.append(repo_url)
    md = "Your new repo:<br>"
    for u in repo_urls:
        md += f"[{str(u).split('/')[-2]}/{str(u).split('/')[-1]}]({str(u)})<br>"
    return gr.update(value=repo_urls, choices=repo_urls), gr.update(value=md)

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--url", default=None, type=str, required=False, help="URL of the model to convert.")
    parser.add_argument("--file", default=None, type=str, required=False, help="Filename of the model to convert.")
    parser.add_argument("--fix", action="store_true", help="Only fix the keys of the local model.")
    parser.add_argument("--civitai_key", default=None, type=str, required=False, help="Civitai API Key (If you want to download file from Civitai).")
    parser.add_argument("--dtype", type=str, default="fp8")
    parser.add_argument("--model", type=str, default="dev")
    parser.add_argument("--dequant", action="store_true", help="Dequantize model.")
    args = parser.parse_args()
    assert (args.url, args.file) != (None, None), "Must provide --url or --file!"

    dtype = torch.bfloat16
    quantization = False
    if args.dtype == "fp8": dtype = torch.float8_e4m3fn
    elif args.dtype == "fp16": dtype = torch.float16
    elif args.dtype == "qfloat8":
        dtype = torch.bfloat16
        quantization = True
    else: dtype = torch.bfloat16

    use_original = ["vae", "text_encoder"]
    new_repo_id = ""
    use_local = True

    if args.file is not None and Path(args.file).exists():
        if args.fix: normalize_flux_state_dict(args.file, ".", dtype, args.dequant)
        else: flux_to_diffusers_lowmem(args.file, Path(args.file).stem, dtype, quantization,
                                       args.model, args.dequant, use_original, new_repo_id, use_local)
    elif args.url is not None:
        convert_url_to_diffusers_flux(args.url, args.civitai_key, False, args.dtype, args.model,
                                      args.dequant)