Soon_Merger / app5.py
AlekseyCalvin's picture
Rename app3.py to app5.py
8da757c verified
import gradio as gr
import torch
import os
import gc
import re
import shutil
import requests
import json
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download, list_repo_files, login
from safetensors.torch import load_file, save_file
from safetensors import safe_open
from tqdm import tqdm
# --- Constants & Setup ---
TempDir = Path("./temp_merge")
os.makedirs(TempDir, exist_ok=True)
api = HfApi()
def cleanup_temp():
if TempDir.exists():
shutil.rmtree(TempDir)
os.makedirs(TempDir, exist_ok=True)
gc.collect()
# --- Core Logic ---
def download_lora(lora_input, hf_token):
"""Downloads LoRA from a Repo ID or a direct URL."""
local_path = TempDir / "adapter.safetensors"
if lora_input.startswith("http"):
print(f"Downloading LoRA from URL: {lora_input}")
response = requests.get(lora_input, stream=True)
response.raise_for_status()
with open(local_path, 'wb') as f:
for chunk in response.iter_content(chunk_size=8192):
f.write(chunk)
return local_path
else:
print(f"Downloading LoRA from Repo: {lora_input}")
try:
return hf_hub_download(repo_id=lora_input, filename="adapter_model.safetensors", token=hf_token, local_dir=TempDir)
except:
files = list_repo_files(repo_id=lora_input, token=hf_token)
# Prioritize safetensors
safe_files = [f for f in files if f.endswith(".safetensors")]
if not safe_files:
raise ValueError("Could not find a .safetensors file in the LoRA repo.")
# Heuristic: pick the one that looks most like a model file
target_file = safe_files[0]
for f in safe_files:
if "fp16" in f or "rank" in f:
target_file = f
break
return hf_hub_download(repo_id=lora_input, filename=target_file, token=hf_token, local_dir=TempDir)
def standardize_lora_config(lora_state_dict):
"""
Analyzes the LoRA state dict and converts keys to a standardized Diffusers-compatible format.
Handles 'lora_down' -> 'lora_A', prefix stripping, and alpha scaling.
"""
standardized_dict = {}
alphas = {}
ranks = {}
keys = list(lora_state_dict.keys())
# 1. First Pass: Detect structure and Alphas
for key in keys:
if "alpha" in key:
# key example: diffusion_model.layers.24.feed_forward.w1.alpha
stem = key.replace(".alpha", "")
alphas[stem] = lora_state_dict[key].item() if isinstance(lora_state_dict[key], torch.Tensor) else lora_state_dict[key]
print(f"Found {len(alphas)} alpha keys in LoRA.")
# 2. Second Pass: Convert Weights
for key in keys:
if "alpha" in key:
continue
tensor = lora_state_dict[key]
new_key = key
# --- Conversion Logic (Inspired by Diffusers lora_conversion_utils.py) ---
# Strip common ComfyUI/Internal prefixes
prefixes_to_strip = ["diffusion_model.", "model.diffusion_model.", "lora_unet_"]
for p in prefixes_to_strip:
if new_key.startswith(p):
new_key = new_key[len(p):]
# Convert lora_down/up to lora_A/B
is_down = "lora_down.weight" in new_key
is_up = "lora_up.weight" in new_key
if is_down:
new_key = new_key.replace("lora_down.weight", "lora_A.weight")
stem = key.split(".lora_down.weight")[0]
ranks[stem] = tensor.shape[0] # Down projection output dim is rank
elif is_up:
new_key = new_key.replace("lora_up.weight", "lora_B.weight")
# Handling Z-Image specific "feed_forward" vs "ff" discrepancies if necessary
# (Based on your logs, Z-Image base uses 'feed_forward' so we might not need heavy mapping if we strip prefix)
standardized_dict[new_key] = tensor
# 3. Third Pass: Embed Scaling into Weights
# If we have alpha and rank, we can pre-multiply the weights so the merge function just needs to do B @ A
# Scale = alpha / rank
final_dict = {}
for key, tensor in standardized_dict.items():
# Find corresponding stem to check for alpha
# key is like: layers.24.feed_forward.w1.lora_A.weight
if "lora_A.weight" in key:
stem_suffix = ".lora_A.weight"
is_A = True
elif "lora_B.weight" in key:
stem_suffix = ".lora_B.weight"
is_A = False
else:
final_dict[key] = tensor
continue
# We need to map the "new key" stem back to the "old key" stem to find the alpha
# This is tricky because we stripped prefixes.
# Simpler approach: Calculate scale factor now if possible, or store metadata.
# Heuristic: Match alpha by checking if alpha key ends with the current key's structural part
# Current key struct: layers.24.feed_forward.w1
struct_part = key.replace(stem_suffix, "")
scale = 1.0
# Find matching alpha
# We look for an alpha key that ends with 'struct_part'
# e.g. alpha key "diffusion_model.layers.24...w1" ends with "layers.24...w1"
found_alpha = None
for a_key, a_val in alphas.items():
if a_key.endswith(struct_part):
found_alpha = a_val
break
if found_alpha:
# We need the rank.
# If it's lora_A, rank is tensor.shape[0]
# If it's lora_B, rank is tensor.shape[1]
rank = tensor.shape[0] if is_A else tensor.shape[1]
# Scale calculation: scale = alpha / rank
# We apply sqrt(scale) to both A and B so that A@B is scaled by (alpha/rank)
scale_factor = (found_alpha / rank) ** 0.5
tensor = tensor * scale_factor
final_dict[key] = tensor
return final_dict
def match_keys(base_key, lora_keys):
"""
Robust matching finding the best LoRA pair for a Base Key.
"""
# base_key example: layers.24.feed_forward.w1.weight
# lora_key example: layers.24.feed_forward.w1.lora_A.weight
base_stem = base_key.replace(".weight", "")
pair_A = None
pair_B = None
# Exact stem match check
candidate_A = f"{base_stem}.lora_A.weight"
candidate_B = f"{base_stem}.lora_B.weight"
if candidate_A in lora_keys and candidate_B in lora_keys:
return candidate_A, candidate_B
# Fuzzy match if exact fails
# This handles slight naming diffs like "processor" inclusion
matches = [k for k in lora_keys if base_stem in k]
for k in matches:
if "lora_A" in k:
pair_A = k
elif "lora_B" in k:
pair_B = k
if pair_A and pair_B:
# Verify they belong to the same block
# e.g. ensure we don't match layer.24 to layer.2
prefix_A = pair_A.split(".lora_A")[0]
prefix_B = pair_B.split(".lora_B")[0]
if prefix_A == prefix_B:
return pair_A, pair_B
return None, None
def copy_auxiliary_files(src_repo, tgt_repo, token):
print(f"Copying infrastructure from {src_repo} to {tgt_repo}...")
try:
files = list_repo_files(repo_id=src_repo, token=token)
files_to_copy = [
f for f in files
if not f.endswith(".safetensors")
and not f.endswith(".bin")
and not f.endswith(".pt")
and not f.endswith(".pth")
and not f.endswith(".msgpack")
and not f.endswith(".h5")
]
for f in tqdm(files_to_copy, desc="Copying configs"):
try:
local = hf_hub_download(repo_id=src_repo, filename=f, token=token)
api.upload_file(
path_or_fileobj=local,
path_in_repo=f,
repo_id=tgt_repo,
repo_type="model",
token=token
)
os.remove(local)
except Exception as e:
print(f"Skipped {f}: {e}")
except Exception as e:
print(f"Error copying config files: {e}")
def run_merge(
hf_token,
base_repo,
base_subfolder,
structure_repo,
lora_input,
user_scale,
output_repo,
is_private,
progress=gr.Progress()
):
cleanup_temp()
logs = []
try:
login(hf_token)
logs.append(f"Logged in. Target: {output_repo}")
# 1. Create Output Repo
try:
api.create_repo(repo_id=output_repo, private=is_private, exist_ok=True, token=hf_token)
logs.append("Output repository ready.")
except Exception as e:
return "\n".join(logs) + f"\nError creating repo: {e}"
# 2. Replicate Structure
if structure_repo.strip():
progress(0.1, desc="Cloning Model Structure...")
logs.append(f"Cloning configuration from {structure_repo}...")
copy_auxiliary_files(structure_repo, output_repo, hf_token)
logs.append("Configuration files copied.")
# 3. Load and Standardize LoRA
progress(0.2, desc="Downloading & Processing LoRA...")
logs.append(f"Fetching LoRA: {lora_input}")
lora_path = download_lora(lora_input, hf_token)
raw_lora_state = load_file(lora_path, device="cpu")
# STANDARDIZE: Convert Comfy/Kohya keys to Diffusers keys & apply Alpha
lora_state = standardize_lora_config(raw_lora_state)
lora_keys = list(lora_state.keys())
logs.append(f"LoRA loaded & standardized. Found {len(lora_keys)} tensors.")
if len(lora_keys) > 0:
logs.append(f"Sample key: {lora_keys[0]}")
# 4. Identify Base Shards
progress(0.3, desc="Analyzing Base Model...")
all_files = list_repo_files(repo_id=base_repo, token=hf_token)
target_shards = []
for f in all_files:
if not f.endswith(".safetensors"):
continue
if base_subfolder.strip() and not f.startswith(base_subfolder.strip("/")):
continue
target_shards.append(f)
logs.append(f"Found {len(target_shards)} matching safetensors shards in base.")
if not target_shards:
raise ValueError("No safetensors found in the specified base repo/subfolder.")
# 5. Process Shards
total_shards = len(target_shards)
merged_count = 0
for idx, shard_file in enumerate(target_shards):
progress(0.3 + (0.6 * (idx / total_shards)), desc=f"Processing Shard {idx+1}/{total_shards}")
logs.append(f"--- Processing {shard_file} ---")
local_shard = hf_hub_download(repo_id=base_repo, filename=shard_file, token=hf_token, local_dir=TempDir)
# Load base to CPU
base_tensors = load_file(local_shard, device="cpu")
modified_tensors = {}
has_changes = False
for key, tensor in base_tensors.items():
pair_A, pair_B = match_keys(key, lora_keys)
if pair_A and pair_B:
w_a = lora_state[pair_A].float()
w_b = lora_state[pair_B].float()
current_tensor = tensor.float()
# Apply merge
# Note: Alpha scaling is already embedded in w_a/w_b by standardize_lora_config
# We just apply the user_scale here
# Check shapes for Transpose requirement
# Standard LoRA: B @ A
try:
delta = (w_b @ w_a) * user_scale
except RuntimeError:
# Shape mismatch fallback
# Sometimes LoRA weights are transposed relative to base
if w_a.shape[0] == w_b.shape[1]:
delta = (w_a @ w_b) * user_scale
else:
# Last ditch: try transposing B
delta = (w_b.T @ w_a) * user_scale
if delta.shape != current_tensor.shape:
if delta.T.shape == current_tensor.shape:
delta = delta.T
else:
# Log only once per shard to avoid spam
if not has_changes:
logs.append(f"Warning: Shape mismatch for {key}. Base: {current_tensor.shape}, Delta: {delta.shape}. Skipping.")
modified_tensors[key] = tensor
continue
modified_tensors[key] = (current_tensor + delta).to(tensor.dtype)
merged_count += 1
has_changes = True
else:
modified_tensors[key] = tensor
if has_changes:
logs.append(f"Merging complete for shard. Saving...")
output_path = TempDir / "processed.safetensors"
save_file(modified_tensors, output_path)
api.upload_file(path_or_fileobj=output_path, path_in_repo=shard_file, repo_id=output_repo, repo_type="model", token=hf_token)
logs.append(f"Uploaded {shard_file}")
else:
logs.append(f"No LoRA matches in this shard. Copying original...")
api.upload_file(path_or_fileobj=local_shard, path_in_repo=shard_file, repo_id=output_repo, repo_type="model", token=hf_token)
# cleanup
del base_tensors
del modified_tensors
if 'delta' in locals(): del delta
gc.collect()
os.remove(local_shard)
if os.path.exists(TempDir / "processed.safetensors"):
os.remove(TempDir / "processed.safetensors")
progress(1.0, desc="Done!")
logs.append(f"\nSUCCESS. Merged {merged_count} layers total.")
logs.append(f"New model available at: https://huggingface.co/{output_repo}")
except Exception as e:
import traceback
logs.append(f"\nCRITICAL ERROR: {str(e)}")
logs.append(traceback.format_exc())
finally:
cleanup_temp()
return "\n".join(logs)
# --- UI ---
css = """
.container { max-width: 900px; margin: auto; }
.header { text-align: center; margin-bottom: 20px; }
"""
with gr.Blocks() as demo:
gr.Markdown(
"""
# ⚡ soonMERGE® for Weights & Adapters
Merge LoRA adapters into **any** base model (LLM, Diffusion, Audio) and reconstruct the repository structure.
**New:** Auto-converts ComfyUI/Kohya LoRA formats (e.g. Z-Image) to match Diffusers base models on the fly.
"""
)
with gr.Group():
gr.Markdown("### 1. Authentication & Output")
with gr.Row():
hf_token = gr.Textbox(label="HF Write Token", type="password", placeholder="hf_...")
output_repo = gr.Textbox(label="Target Output Repo", placeholder="username/Z-Image-Turbo-Merged")
is_private = gr.Checkbox(label="Private Repo", value=True)
with gr.Group():
gr.Markdown("### 2. Base Weights (The Target)")
with gr.Row():
base_repo = gr.Textbox(label="Base Model Repo", placeholder="e.g. ostris/Z-Image-De-Turbo")
base_subfolder = gr.Textbox(label="Subfolder (Optional)", placeholder="e.g. transformer", info="Only merge weights found inside this folder.")
with gr.Group():
gr.Markdown("### 3. LoRA Configuration")
with gr.Row():
lora_input = gr.Textbox(label="LoRA Source", placeholder="Repo ID OR Direct URL (http...)", info="Accepts direct .safetensors resolve links.")
scale = gr.Slider(label="Scale", minimum=-2.0, maximum=2.0, value=1.0, step=0.1, info="Global multiplier (applied on top of LoRA's internal alpha)")
with gr.Group():
gr.Markdown("### 4. Repository Reconstruction (Optional)")
gr.Markdown("*Use this to fill in missing files (Scheduler, VAE, Tokenizer, model_index.json) from a different source repo.*")
structure_repo = gr.Textbox(label="Structure Source Repo", placeholder="e.g. Tongyi-MAI/Z-Image-Turbo", info="Copies all NON-weight files from here to output.")
submit_btn = gr.Button("🚀 Start Merge & Upload", variant="primary")
output_log = gr.Textbox(label="Process Log", lines=20, interactive=False)
submit_btn.click(
fn=run_merge,
inputs=[hf_token, base_repo, base_subfolder, structure_repo, lora_input, scale, output_repo, is_private],
outputs=output_log
)
if __name__ == "__main__":
demo.queue(max_size=1).launch(css=css)