Soon_Merger / appMerger01.py
AlekseyCalvin's picture
Rename app.py to appMerger01.py
086cad3 verified
import gradio as gr
import torch
import os
import gc
import shutil
import requests
import json
import struct
import numpy as np
import yaml
import subprocess
import sys
import tempfile
import re
from pathlib import Path
from typing import Dict, Any, Optional, List, Iterable
from huggingface_hub import HfApi, hf_hub_download, list_repo_files, login
from safetensors.torch import load_file, save_file
from tqdm import tqdm
# --- Import MergeKit Config ---
try:
from mergekit.config import MergeConfiguration
except ImportError:
# Fallback if installation fails temporarily
class MergeConfiguration:
@staticmethod
def model_validate(config): pass
# --- Constants & Setup ---
try:
TempDir = Path("/tmp/temp_tool")
os.makedirs(TempDir, exist_ok=True)
except:
TempDir = Path("./temp_tool")
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()
# =================================================================================
# SHARED HELPERS (Tabs 1-4 & 10)
# =================================================================================
def parse_hf_url(url):
if "huggingface.co" in url and "resolve" in url:
try:
parts = url.split("huggingface.co/")[-1].split("/")
repo_id = f"{parts[0]}/{parts[1]}"
filename = "/".join(parts[4:]).split("?")[0]
return repo_id, filename
except:
return None, None
return None, None
def download_lora_smart(input_str, token):
local_path = TempDir / "adapter.safetensors"
if local_path.exists(): os.remove(local_path)
repo_id, filename = parse_hf_url(input_str)
if repo_id and filename:
try:
hf_hub_download(repo_id=repo_id, filename=filename, token=token, local_dir=TempDir)
found = list(TempDir.rglob(filename.split("/")[-1]))[0]
if found != local_path: shutil.move(found, local_path)
return local_path
except: pass
try:
if ".safetensors" in input_str and input_str.count("/") >= 2:
parts = input_str.split("/")
repo_id = f"{parts[0]}/{parts[1]}"
filename = "/".join(parts[2:])
hf_hub_download(repo_id=repo_id, filename=filename, token=token, local_dir=TempDir)
found = list(TempDir.rglob(filename.split("/")[-1]))[0]
if found != local_path: shutil.move(found, local_path)
return local_path
candidates = ["adapter_model.safetensors", "model.safetensors"]
files = list_repo_files(repo_id=input_str, token=token)
target = next((f for f in files if f in candidates), None)
if not target:
safes = [f for f in files if f.endswith(".safetensors")]
if safes: target = safes[0]
if not target: raise ValueError("No safetensors found")
hf_hub_download(repo_id=input_str, filename=target, token=token, local_dir=TempDir)
found = list(TempDir.rglob(target.split("/")[-1]))[0]
if found != local_path: shutil.move(found, local_path)
return local_path
except Exception as e:
if input_str.startswith("http"):
try:
headers = {"Authorization": f"Bearer {token}"} if token else {}
r = requests.get(input_str, stream=True, headers=headers, timeout=60)
r.raise_for_status()
with open(local_path, 'wb') as f:
for chunk in r.iter_content(chunk_size=8192): f.write(chunk)
return local_path
except: pass
raise e
def load_lora_to_memory(lora_path, precision_dtype=torch.bfloat16):
state_dict = load_file(lora_path, device="cpu")
pairs = {}
alphas = {}
for k, v in state_dict.items():
stem = get_key_stem(k)
if "alpha" in k:
alphas[stem] = v.item() if isinstance(v, torch.Tensor) else v
else:
if stem not in pairs: pairs[stem] = {}
if "lora_down" in k or "lora_A" in k:
pairs[stem]["down"] = v.to(dtype=precision_dtype)
pairs[stem]["rank"] = v.shape[0]
elif "lora_up" in k or "lora_B" in k:
pairs[stem]["up"] = v.to(dtype=precision_dtype)
for stem in pairs:
pairs[stem]["alpha"] = alphas.get(stem, float(pairs[stem].get("rank", 1.0)))
return pairs
def get_key_stem(key):
key = key.replace(".weight", "").replace(".bias", "")
key = key.replace(".lora_down", "").replace(".lora_up", "")
key = key.replace(".lora_A", "").replace(".lora_B", "")
key = key.replace(".alpha", "")
prefixes = [
"model.diffusion_model.", "diffusion_model.", "model.",
"transformer.", "text_encoder.", "lora_unet_", "lora_te_", "base_model.model."
]
changed = True
while changed:
changed = False
for p in prefixes:
if key.startswith(p):
key = key[len(p):]
changed = True
return key
# =================================================================================
# TABS 1-4 LOGIC (RESTORED)
# =================================================================================
class MemoryEfficientSafeOpen:
def __init__(self, filename):
self.filename = filename
self.file = open(filename, "rb")
self.header, self.header_size = self._read_header()
def __enter__(self): return self
def __exit__(self, exc_type, exc_val, exc_tb): self.file.close()
def keys(self) -> list[str]: return [k for k in self.header.keys() if k != "__metadata__"]
def metadata(self) -> Dict[str, str]: return self.header.get("__metadata__", {})
def get_tensor(self, key):
if key not in self.header: raise KeyError(f"Tensor '{key}' not found")
metadata = self.header[key]
start, end = metadata["data_offsets"]
self.file.seek(self.header_size + 8 + start)
return self._deserialize_tensor(self.file.read(end - start), metadata)
def _read_header(self):
header_size = struct.unpack("<Q", self.file.read(8))[0]
return json.loads(self.file.read(header_size).decode("utf-8")), header_size
def _deserialize_tensor(self, tensor_bytes, metadata):
dtype_map = {"F32": torch.float32, "F16": torch.float16, "BF16": torch.bfloat16, "I64": torch.int64, "I32": torch.int32, "I16": torch.int16, "I8": torch.int8, "U8": torch.uint8, "BOOL": torch.bool}
dtype = dtype_map[metadata["dtype"]]
shape = metadata["shape"]
return torch.frombuffer(tensor_bytes, dtype=torch.uint8).view(dtype).reshape(shape)
class ShardBuffer:
def __init__(self, max_size_gb, output_dir, output_repo, subfolder, hf_token, filename_prefix="model"):
self.max_bytes = int(max_size_gb * 1024**3)
self.output_dir, self.output_repo, self.subfolder, self.hf_token, self.filename_prefix = output_dir, output_repo, subfolder, hf_token, filename_prefix
self.buffer, self.current_bytes, self.shard_count, self.index_map, self.total_size = [], 0, 0, {}, 0
def add_tensor(self, key, tensor):
if tensor.dtype == torch.bfloat16: raw, dt = tensor.view(torch.int16).numpy().tobytes(), "BF16"
elif tensor.dtype == torch.float16: raw, dt = tensor.numpy().tobytes(), "F16"
else: raw, dt = tensor.numpy().tobytes(), "F32"
self.buffer.append({"key": key, "data": raw, "dtype": dt, "shape": tensor.shape})
self.current_bytes += len(raw)
self.total_size += len(raw)
if self.current_bytes >= self.max_bytes: self.flush()
def flush(self):
if not self.buffer: return
self.shard_count += 1
fname = f"{self.filename_prefix}-{self.shard_count:05d}.safetensors"
header = {"__metadata__": {"format": "pt"}}
curr_off = 0
for i in self.buffer:
header[i["key"]] = {"dtype": i["dtype"], "shape": i["shape"], "data_offsets": [curr_off, curr_off + len(i["data"])]}
curr_off += len(i["data"])
self.index_map[i["key"]] = fname
out = self.output_dir / fname
header_json = json.dumps(header).encode('utf-8')
with open(out, 'wb') as f:
f.write(struct.pack('<Q', len(header_json)))
f.write(header_json)
for i in self.buffer: f.write(i["data"])
api.upload_file(path_or_fileobj=out, path_in_repo=f"{self.subfolder}/{fname}" if self.subfolder else fname, repo_id=self.output_repo, token=self.hf_token)
os.remove(out)
self.buffer, self.current_bytes = [], 0
gc.collect()
def task_merge_legacy(hf_token, base, sub, lora, scale, prec, shard, out, struct_s, priv, progress=gr.Progress()):
cleanup_temp()
if hf_token: login(hf_token.strip())
try: api.create_repo(repo_id=out, private=priv, exist_ok=True, token=hf_token)
except Exception as e: return f"Error: {e}"
if struct_s:
try:
files = api.list_repo_files(repo_id=struct_s, token=hf_token)
for f in tqdm(files, desc="Copying Structure"):
if sub and f.startswith(sub): continue
if not sub and any(f.endswith(x) for x in ['.safetensors', '.bin', '.pt', '.pth']): continue
l = hf_hub_download(repo_id=struct_s, filename=f, token=hf_token, local_dir=TempDir)
api.upload_file(path_or_fileobj=l, path_in_repo=f, repo_id=out, token=hf_token)
except: pass
files = [f for f in list_repo_files(repo_id=base, token=hf_token) if f.endswith(".safetensors")]
if sub: files = [f for f in files if f.startswith(sub)]
if not files: return "No safetensors found"
prefix = "diffusion_pytorch_model" if (sub in ["transformer", "unet"] or "diffusion_pytorch_model" in os.path.basename(files[0])) else "model"
dtype = torch.bfloat16 if prec == "bf16" else torch.float16 if prec == "fp16" else torch.float32
try: lora_pairs = load_lora_to_memory(download_lora_smart(lora, hf_token), dtype)
except Exception as e: return f"LoRA Error: {e}"
buf = ShardBuffer(shard, TempDir, out, sub, hf_token, prefix)
for i, fpath in enumerate(files):
local = hf_hub_download(repo_id=base, filename=fpath, token=hf_token, local_dir=TempDir)
with MemoryEfficientSafeOpen(local) as f:
for k in f.keys():
v = f.get_tensor(k)
stem = get_key_stem(k)
match = lora_pairs.get(stem) or lora_pairs.get(stem.replace("to_q", "qkv")) or lora_pairs.get(stem.replace("to_k", "qkv")) or lora_pairs.get(stem.replace("to_v", "qkv"))
if match:
d, u = match["down"], match["up"]
s = scale * (match["alpha"] / match["rank"])
if len(v.shape)==4 and len(d.shape)==2: d, u = d.unsqueeze(-1).unsqueeze(-1), u.unsqueeze(-1).unsqueeze(-1)
delta = (up.squeeze() @ down.squeeze()).reshape(up.shape[0], down.shape[1], 1, 1) if len(up.shape)==4 else u @ d
v = v.to(dtype).add_((delta * s).to(dtype))
buf.add_tensor(k, v.to(dtype))
os.remove(local)
buf.flush()
idx = {"metadata": {"total_size": buf.total_size}, "weight_map": buf.index_map}
idx_n = f"{prefix}.safetensors.index.json"
with open(TempDir/idx_n, "w") as f: json.dump(idx, f, indent=4)
api.upload_file(path_or_fileobj=TempDir/idx_n, path_in_repo=f"{sub}/{idx_n}" if sub else idx_n, repo_id=out, token=hf_token)
return "Done"
def task_extract(hf_token, org, tun, rank, out):
cleanup_temp()
if hf_token: login(hf_token.strip())
try:
p1 = download_lora_smart(org, hf_token)
p2 = download_lora_smart(tun, hf_token)
org_f, tun_f = MemoryEfficientSafeOpen(p1), MemoryEfficientSafeOpen(p2)
lora_sd = {}
common = set(org_f.keys()) & set(tun_f.keys())
for k in tqdm(common, desc="Extracting"):
if "num_batches_tracked" in k or "running_mean" in k or "running_var" in k: continue
m1, m2 = org_f.get_tensor(k).float(), tun_f.get_tensor(k).float()
if m1.shape != m2.shape: continue
diff = m2 - m1
if torch.max(torch.abs(diff)) < 1e-4: continue
out_d, in_d = diff.shape[0], diff.shape[1] if len(diff.shape) > 1 else 1
r = min(int(rank), in_d, out_d)
if len(diff.shape)==4: diff = diff.flatten(1)
elif len(diff.shape)==1: diff = diff.unsqueeze(1)
U, S, V = torch.svd_lowrank(diff, q=r+4, niter=4)
Vh = V.t()
U, S, Vh = U[:, :r], S[:r], Vh[:r, :]
U = U @ torch.diag(S)
dist = torch.cat([U.flatten(), Vh.flatten()])
hi_val = torch.quantile(torch.abs(dist), 0.99)
if hi_val > 0: U, Vh = U.clamp(-hi_val, hi_val), Vh.clamp(-hi_val, hi_val)
if len(m1.shape)==4:
U = U.reshape(out_d, r, 1, 1)
Vh = Vh.reshape(r, in_d, m1.shape[2], m1.shape[3])
else:
U, Vh = U.reshape(out_d, r), Vh.reshape(r, in_d)
stem = k.replace(".weight", "")
lora_sd[f"{stem}.lora_up.weight"] = U.contiguous()
lora_sd[f"{stem}.lora_down.weight"] = Vh.contiguous()
lora_sd[f"{stem}.alpha"] = torch.tensor(r).float()
out_f = TempDir/"extracted.safetensors"
save_file(lora_sd, out_f)
api.create_repo(repo_id=out, exist_ok=True, token=hf_token)
api.upload_file(path_or_fileobj=out_f, path_in_repo="extracted_lora.safetensors", repo_id=out, token=hf_token)
return "Done"
except Exception as e: return f"Error: {e}"
def load_full_state_dict(path):
raw = load_file(path, device="cpu")
cleaned = {}
for k, v in raw.items():
if "lora_A" in k: new_k = k.replace("lora_A", "lora_down")
elif "lora_B" in k: new_k = k.replace("lora_B", "lora_up")
else: new_k = k
cleaned[new_k] = v.float()
return cleaned
def task_merge_adapters_advanced(hf_token, inputs_text, method, weight_str, beta, sigma_rel, target_rank, out_repo, private):
cleanup_temp()
if hf_token: login(hf_token.strip())
urls = [line.strip() for line in inputs_text.replace(" ", "\n").split('\n') if line.strip()]
if len(urls) < 2: return "Error: Provide at least 2 adapters."
try: weights = [float(w.strip()) for w in weight_str.split(',')] if weight_str.strip() else [1.0] * len(urls)
except: return "Error parsing weights."
if len(weights) < len(urls): weights += [1.0] * (len(urls) - len(weights))
paths = []
for url in tqdm(urls, desc="Downloading"): paths.append(download_lora_smart(url, hf_token))
merged = {}
if "Iterative EMA" in method:
base_sd = load_file(paths[0], device="cpu")
gamma = None
if sigma_rel > 0:
t_val = sigma_rel**-2
roots = np.roots([1, 7, 16 - t_val, 12 - t_val])
gamma = roots[np.isreal(roots) & (roots.real >= 0)].real.max()
for i, path in enumerate(paths[1:]):
current_beta = (1 - 1 / (i + 1)) ** (gamma + 1) if gamma is not None else beta
curr = load_file(path, device="cpu")
for k in base_sd:
if k in curr and "alpha" not in k:
base_sd[k] = base_sd[k].float() * current_beta + curr[k].float() * (1 - current_beta)
merged = base_sd
else:
states = [load_full_state_dict(p) for p in paths]
all_stems = set()
for s in states:
for k in s:
if "lora_" in k: all_stems.add(k.split(".lora_")[0])
for stem in tqdm(all_stems):
down_list, up_list = [], []
alpha_sum, total_delta = 0.0, None
for i, state in enumerate(states):
w = weights[i]
dk, uk, ak = f"{stem}.lora_down.weight", f"{stem}.lora_up.weight", f"{stem}.alpha"
if dk in state and uk in state:
d, u = state[dk], state[uk]
alpha_sum += state[ak].item() if ak in state else d.shape[0]
if "Concatenation" in method:
down_list.append(d); up_list.append(u * w)
elif "SVD" in method:
rank = d.shape[0]
alpha = state[ak].item() if ak in state else rank
scale = (alpha / rank) * w
delta = ((u.flatten(1) @ d.flatten(1)).reshape(u.shape[0], d.shape[1], d.shape[2], d.shape[3]) if len(d.shape)==4 else u @ d) * scale
total_delta = delta if total_delta is None else total_delta + delta
if "Concatenation" in method and down_list:
merged[f"{stem}.lora_down.weight"] = torch.cat(down_list, dim=0).contiguous()
merged[f"{stem}.lora_up.weight"] = torch.cat(up_list, dim=1).contiguous()
merged[f"{stem}.alpha"] = torch.tensor(alpha_sum)
elif "SVD" in method and total_delta is not None:
tr = int(target_rank)
flat = total_delta.flatten(1) if len(total_delta.shape)==4 else total_delta
try:
U, S, V = torch.svd_lowrank(flat, q=tr + 4, niter=4)
Vh = V.t()
U, S, Vh = U[:, :tr], S[:tr], Vh[:tr, :]
U = U @ torch.diag(S)
if len(total_delta.shape) == 4:
U = U.reshape(total_delta.shape[0], tr, 1, 1)
Vh = Vh.reshape(tr, total_delta.shape[1], total_delta.shape[2], total_delta.shape[3])
else:
U, Vh = U.reshape(total_delta.shape[0], tr), Vh.reshape(tr, total_delta.shape[1])
merged[f"{stem}.lora_down.weight"] = Vh.contiguous()
merged[f"{stem}.lora_up.weight"] = U.contiguous()
merged[f"{stem}.alpha"] = torch.tensor(tr).float()
except: pass
out = TempDir / "merged_adapters.safetensors"
if merged: save_file(merged, out)
api.create_repo(repo_id=out_repo, private=private, exist_ok=True, token=hf_token)
api.upload_file(path_or_fileobj=out, path_in_repo="merged_adapters.safetensors", repo_id=out_repo, token=hf_token)
return f"Success! Merged to {out_repo}"
def task_resize(hf_token, lora_input, new_rank, dynamic_method, dynamic_param, out_repo):
cleanup_temp()
if hf_token: login(hf_token.strip())
path = download_lora_smart(lora_input, hf_token)
state = load_file(path, device="cpu")
new_state = {}
groups = {}
for k in state:
simple = k.split(".lora_")[0]
if simple not in groups: groups[simple] = {}
if "lora_down" in k or "lora_A" in k: groups[simple]["down"] = state[k]
if "lora_up" in k or "lora_B" in k: groups[simple]["up"] = state[k]
if "alpha" in k: groups[simple]["alpha"] = state[k]
target_rank_limit = int(new_rank)
for stem, g in tqdm(groups.items()):
if "down" in g and "up" in g:
down, up = g["down"].float(), g["up"].float()
merged = (up.squeeze() @ down.squeeze()).reshape(up.shape[0], down.shape[1], down.shape[2], down.shape[3]) if len(down.shape)==4 else up @ down
flat = merged.flatten(1)
U, S, V = torch.svd_lowrank(flat, q=target_rank_limit + 32)
Vh = V.t()
calc_rank = target_rank_limit
if dynamic_method == "sv_ratio":
calc_rank = int(torch.sum(S > (S[0] / dynamic_param)).item())
elif dynamic_method == "sv_cumulative":
calc_rank = int(torch.searchsorted(torch.cumsum(S, 0) / torch.sum(S), dynamic_param)) + 1
elif dynamic_method == "sv_fro":
calc_rank = int(torch.searchsorted(torch.cumsum(S.pow(2), 0) / torch.sum(S.pow(2)), dynamic_param**2)) + 1
final_rank = max(1, min(calc_rank, target_rank_limit, S.shape[0]))
U = U[:, :final_rank] @ torch.diag(S[:final_rank])
Vh = Vh[:final_rank, :]
if len(down.shape) == 4:
U = U.reshape(up.shape[0], final_rank, 1, 1)
Vh = Vh.reshape(final_rank, down.shape[1], down.shape[2], down.shape[3])
new_state[f"{stem}.lora_down.weight"] = Vh.contiguous()
new_state[f"{stem}.lora_up.weight"] = U.contiguous()
new_state[f"{stem}.alpha"] = torch.tensor(final_rank).float()
out = TempDir / "shrunken.safetensors"
save_file(new_state, out)
api.create_repo(repo_id=out_repo, exist_ok=True, token=hf_token)
api.upload_file(path_or_fileobj=out, path_in_repo="shrunken.safetensors", repo_id=out_repo, token=hf_token)
return "Done"
# =================================================================================
# MERGEKIT & STREAMING LOGS (TABS 5-9)
# =================================================================================
def parse_weight(w_str):
if not w_str.strip(): return 1.0
try:
if "[" in w_str: return yaml.safe_load(w_str)
return float(w_str)
except: return 1.0
def run_mergekit_logic(config_dict, token, out_repo, private, shard_size, output_precision, tokenizer_source, chat_template, program="mergekit-yaml"):
# Using generator for streaming logs directly to a Textbox, bypassing component issues
logs = []
def log(msg):
logs.append(msg)
return "\n".join(logs)
yield log("Starting MergeKit Process...")
cleanup_temp()
if chat_template and chat_template.strip():
config_dict["chat_template"] = chat_template.strip()
# Validation
try:
if program != "mergekit-moe":
MergeConfiguration.model_validate(config_dict)
yield log("Configuration Validated Successfully.")
except Exception as e:
yield log(f"Invalid Config: {e}")
return
if token:
login(token.strip())
os.environ["HF_TOKEN"] = token.strip()
if "dtype" not in config_dict: config_dict["dtype"] = output_precision
if "tokenizer_source" not in config_dict and tokenizer_source != "base":
config_dict["tokenizer_source"] = tokenizer_source
config_path = TempDir / "config.yaml"
with open(config_path, "w") as f: yaml.dump(config_dict, f, sort_keys=False)
yield log(f"Config saved to {config_path}")
yield log(f"YAML:\n{yaml.dump(config_dict, sort_keys=False)}")
try:
api.create_repo(repo_id=out_repo, private=private, exist_ok=True, token=token)
yield log(f"Repo {out_repo} ready.")
except Exception as e:
yield log(f"Repo Creation Error (might exist): {e}")
out_path = TempDir / "merge_output"
shard_arg = f"{int(float(shard_size) * 1024)}M"
cmd = [
program,
str(config_path),
str(out_path),
"--allow-crimes",
"--copy-tokenizer",
"--out-shard-size", shard_arg,
"--lazy-unpickle"
]
if torch.cuda.is_available():
cmd.extend(["--cuda", "--low-cpu-memory"])
yield log(f"Executing: {' '.join(cmd)}")
env = os.environ.copy()
env["HF_HOME"] = str(TempDir / ".cache")
# Run process
process = subprocess.Popen(cmd, env=env, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
for line in iter(process.stdout.readline, ""):
yield log(line.strip())
process.wait()
if process.returncode != 0:
yield log("Merge failed with exit code " + str(process.returncode))
return
yield log(f"Uploading to {out_repo}...")
try:
api.upload_folder(repo_id=out_repo, folder_path=out_path)
yield log("Upload Complete!")
except Exception as e:
yield log(f"Upload failed: {e}")
# --- UI Wrappers for Tabs 5-9 ---
def wrapper_amphinterpolative(token, method, base, t, norm, i8, flat, row, eps, m_iter, tol, m1, w1, m2, w2, m3, w3, m4, w4, m5, w5, out, priv, shard, prec, tok_src, chat_t):
params = {"normalize": norm, "int8_mask": i8}
if method in ["slerp", "nuslerp"]: params["t"] = float(t)
if method == "nuslerp": params.update({"flatten": flat, "row_wise": row})
if method == "multislerp": params["eps"] = float(eps)
if method == "karcher": params.update({"max_iter": int(m_iter), "tol": float(tol)})
config = {"merge_method": method}
if method in ["slerp", "nuslerp"]:
if not base.strip(): yield "Error: Base model required"; return
config["base_model"] = base.strip()
sources = [{"model": m, "parameters": {"weight": parse_weight(w)}} for m, w in [(m1,w1), (m2,w2)] if m.strip()]
config["slices"] = [{"sources": sources, "parameters": params}]
else:
if base.strip() and method == "multislerp": config["base_model"] = base.strip()
models = [{"model": m, "parameters": {"weight": parse_weight(w)}} for m, w in [(m1,w1), (m2,w2), (m3,w3), (m4,w4), (m5,w5)] if m.strip()]
config["models"] = models
config["parameters"] = params
yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-yaml")
def wrapper_stirtie(token, method, base, norm, i8, lamb, resc, topk, m1, w1, d1, g1, e1, m2, w2, d2, g2, e2, m3, w3, d3, g3, e3, m4, w4, d4, g4, e4, out, priv, shard, prec, tok_src, chat_t):
models = []
# Explicit loop over the 4 sets of model inputs
for m, w, d, g, e in [
(m1, w1, d1, g1, e1),
(m2, w2, d2, g2, e2),
(m3, w3, d3, g3, e3),
(m4, w4, d4, g4, e4)
]:
if not m.strip(): continue
p = {"weight": parse_weight(w)}
if method in ["ties", "dare_ties", "dare_linear", "breadcrumbs_ties"]: p["density"] = parse_weight(d)
if "breadcrumbs" in method: p["gamma"] = float(g)
if "della" in method: p["epsilon"] = float(e)
models.append({"model": m, "parameters": p})
g_params = {"normalize": norm, "int8_mask": i8}
if method != "sce": g_params["lambda"] = float(lamb)
if method == "dare_linear": g_params["rescale"] = resc
if method == "sce": g_params["select_topk"] = float(topk)
config = {
"merge_method": method,
"base_model": base.strip() if base.strip() else models[0]["model"],
"parameters": g_params,
"models": models
}
yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-yaml")
def wrapper_specious(token, method, base, norm, i8, t, filt_w, m1, w1, f1, m2, w2, m3, w3, m4, w4, m5, w5, out, priv, shard, prec, tok_src, chat_t):
models = []
if method == "passthrough":
p = {"weight": parse_weight(w1)}
if f1.strip(): p["filter"] = f1.strip()
models.append({"model": m1, "parameters": p})
else:
models = [{"model": m, "parameters": {"weight": parse_weight(w)}} for m, w in [(m1,w1), (m2,w2), (m3,w3), (m4,w4), (m5,w5)] if m.strip()]
config = {"merge_method": method, "parameters": {"normalize": norm, "int8_mask": i8}}
if base.strip(): config["base_model"] = base.strip()
if method == "nearswap": config["parameters"]["t"] = float(t)
if method == "model_stock": config["parameters"]["filter_wise"] = filt_w
config["models"] = models
yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-yaml")
def wrapper_moer(token, base, expert1, prompt1, expert2, prompt2, expert3, prompt3, expert4, prompt4, expert5, prompt5, gate, dtype, out, priv, shard, prec, tok_src, chat_t):
experts = []
for exp, pmt in [
(expert1, prompt1), (expert2, prompt2), (expert3, prompt3),
(expert4, prompt4), (expert5, prompt5)
]:
if exp.strip():
expert_entry = {"source_model": exp.strip()}
# Parse prompts (comma-separated)
if pmt.strip():
prompts = [p.strip() for p in pmt.split(',') if p.strip()]
expert_entry["positive_prompts"] = prompts
else:
expert_entry["positive_prompts"] = [""]
experts.append(expert_entry)
if len(experts) < 2:
return "Error: At least 2 experts required"
# Build config for MoE
config = {
"base_model": base.strip(),
"gate_mode": gate,
"dtype": dtype,
"experts": experts
}
# Uses mergekit-moe CLI
yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-moe")
def wrapper_rawer(token, models, method, dtype, out, priv, shard, prec, tok_src, chat_t):
models_list = [{"model": m.strip(), "parameters": {"weight": 1.0}} for m in models.split('\n') if m.strip()]
config = {
"models": models_list,
"merge_method": method,
"dtype": dtype
}
yield from run_mergekit_logic(config, token, out, priv, shard, prec, tok_src, chat_t, program="mergekit-yaml")
# --- TAB 10 (Custom DARE) Logic ---
def task_dare_custom(token, base, ft, ratio, mask, out, priv):
cleanup_temp()
if token: login(token.strip())
try:
b_path = download_lora_smart(base, token)
f_path = download_lora_smart(ft, token)
b_sd = load_file(b_path, device="cpu")
f_sd = load_file(f_path, device="cpu")
merged = {}
common = set(b_sd.keys()) & set(f_sd.keys())
for k in tqdm(common, desc="Merging"):
tb, tf = b_sd[k], f_sd[k]
if tb.shape != tf.shape:
merged[k] = tf
continue
delta = tf.float() - tb.float()
if mask > 0:
m = torch.bernoulli(torch.full_like(delta, 1.0 - mask))
delta = (delta * m) / (1.0 - mask)
merged[k] = (tb.float() + ratio * delta).to(tb.dtype)
out_f = TempDir / "model.safetensors"
save_file(merged, out_f)
api.create_repo(repo_id=out, private=priv, exist_ok=True, token=token)
api.upload_file(path_or_fileobj=out_f, path_in_repo="model.safetensors", repo_id=out, token=token)
return f"Done! {out}"
except Exception as e: return str(e)
# =================================================================================
# UI GENERATION
# =================================================================================
css = ".container { max-width: 1100px; margin: auto; }"
with gr.Blocks() as demo:
gr.HTML("""<h1><img src="https://huggingface.co/spaces/AlekseyCalvin/Soon_Merger/resolve/main/SMerger3.png" alt="SOONmerge®"> Transform Transformers for FREE!</h1>""")
gr.Markdown("# 🧰Training-Free CPU-run Model Creation Toolkit | **MergeKit** implementation in Tabs 5-9 & MORE")
with gr.Tabs():
# --- TAB 1: RESTORED ---
with gr.Tab("Merge to Base Model + Reshard Output"):
t1_token = gr.Textbox(label="Token", type="password")
t1_base = gr.Textbox(label="Base Repo", value="name/repo")
t1_sub = gr.Textbox(label="Subfolder (Optional)", value="")
t1_lora = gr.Textbox(label="LoRA Direct Link or Repo", value="https://huggingface.co/GuangyuanSD/Z-Image-Re-Turbo-LoRA/resolve/main/Z-image_re_turbo_lora_8steps_rank_32_v1_fp16.safetensors")
with gr.Row():
t1_scale = gr.Slider(label="Scale", value=1.0, minimum=0, maximum=3.0, step=0.1)
t1_prec = gr.Radio(["bf16", "fp16", "float32"], value="bf16", label="Precision")
t1_shard = gr.Slider(label="Max Shard Size (GB)", value=2.0, minimum=0.1, maximum=10.0, step=0.1)
t1_out = gr.Textbox(label="Output Repo")
t1_struct = gr.Textbox(label="Extras Source (copies configs/components/etc)", value="name/repo")
t1_priv = gr.Checkbox(label="Private", value=True)
gr.Button("Merge").click(task_merge_legacy, [t1_token, t1_base, t1_sub, t1_lora, t1_scale, t1_prec, t1_shard, t1_out, t1_struct, t1_priv], gr.Textbox(label="Result"))
# --- TAB 2: RESTORED ---
with gr.Tab("Extract Adapter"):
t2_token = gr.Textbox(label="Token", type="password")
t2_org = gr.Textbox(label="Original Model")
t2_tun = gr.Textbox(label="Tuned or Homologous Model")
t2_rank = gr.Number(label="Extract At Rank", value=32, minimum=1, maximum=1024, step=1)
t2_out = gr.Textbox(label="Output Repo")
gr.Button("Extract").click(task_extract, [t2_token, t2_org, t2_tun, t2_rank, t2_out], gr.Textbox(label="Result"))
# --- TAB 3: RESTORED ---
with gr.Tab("Merge Adapters"):
gr.Markdown("### Batch Adapter Merging")
t3_token = gr.Textbox(label="Token", type="password")
t3_urls = gr.TextArea(label="Adapter URLs/Repos (one per line, or space-separated)")
t3_method = gr.Dropdown(["Iterative EMA (Linear w/ Beta/Sigma coefficient)", "Concatenation (MOE-like weights-stack)", "SVD Fusion (Task Arithmetic/Compressed)"], value="Iterative EMA (Linear w/ Beta/Sigma coefficient)", label="Merge Method")
with gr.Row():
t3_weights = gr.Textbox(label="Weights (comma-separated) – for Concat/SVD")
t3_rank = gr.Number(label="Target Rank – For SVD only", value=128)
with gr.Row():
t3_beta = gr.Slider(label="Beta – for linear/post-hoc EMA", value=0.95, minimum=0.01, maximum=1.00)
t3_sigma = gr.Slider(label="Sigma Rel – for linear/post-hoc EMA", value=0.21, minimum=0.01, maximum=1.00)
t3_out = gr.Textbox(label="Output Repo")
t3_priv = gr.Checkbox(label="Private Output", value=True)
gr.Button("Merge").click(task_merge_adapters_advanced, [t3_token, t3_urls, t3_method, t3_weights, t3_beta, t3_sigma, t3_rank, t3_out, t3_priv], gr.Textbox(label="Result"))
# --- TAB 4: RESTORED ---
with gr.Tab("Resize Adapter"):
t4_token = gr.Textbox(label="Token", type="password")
t4_in = gr.Textbox(label="LoRA")
with gr.Row():
t4_rank = gr.Number(label="To Rank (Safety Ceiling)", value=8)
t4_method = gr.Dropdown(["None", "sv_ratio", "sv_fro", "sv_cumulative"], value="None", label="Dynamic Method")
t4_param = gr.Number(label="Dynamic Param", value=0.9)
gr.Markdown("### 📉 Dynamic Resizing Guide\nThese methods intelligently determine the best rank per layer.\n- **sv_ratio (Relative Strength):** Keeps features that are at least `1/Param` as strong as the main feature. **Param must be >= 2**.\n- **sv_fro (Visual Information Density):** Preserves `Param%` of total information content. **Param between 0.0 and 1.0**.\n- **sv_cumulative (Cumulative Sum):** Preserves weights that sum up to `Param%` of total strength. **Param between 0.0 and 1.0**.\n- **⚠️ Safety Ceiling:** The **'To Rank'** slider acts as a hard limit.")
t4_out = gr.Textbox(label="Output")
gr.Button("Resize").click(task_resize, [t4_token, t4_in, t4_rank, t4_method, t4_param, t4_out], gr.Textbox(label="Result"))
# --- TAB 5 ---
with gr.Tab("Amphinterpolative"):
gr.Markdown("### Spherical Interpolation Methods Family: slerp, nuslerp, multislerp, karcher")
t5_token = gr.Textbox(label="HF Token", type="password")
t5_method = gr.Dropdown(["slerp", "nuslerp", "multislerp", "karcher"], value="slerp", label="Merge Method")
gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
with gr.Row():
t5_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0)
t5_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision")
t5_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source")
t5_chat = gr.Textbox(label="Chat Template (default: auto)", placeholder="auto")
gr.Markdown("Built-in Chat Templates: alpaca, chatml, llama3, mistral, exaone, auto")
with gr.Row():
t5_base = gr.Textbox(label="Base Model")
t5_t = gr.Slider(0, 1, 0.5, label="t")
with gr.Row():
t5_norm = gr.Checkbox(label="Normalize Weights", value=True); t5_i8 = gr.Checkbox(label="Int8 Mask", value=False); t5_flat = gr.Checkbox(label="Flatten Tensors (NuSlerp)", value=False); t5_row = gr.Checkbox(label="Row Wise (NuSlerp)", value=False)
with gr.Row():
t5_eps = gr.Textbox(label="eps (Stabilization Constant) (MultiSlerp)", value="1e-8"); t5_iter = gr.Number(label="Max Iterations (Karcher)", value=10); t5_tol = gr.Textbox(label="tol (Convergence Tolerance) (Karcher)", value="1e-5")
gr.Markdown("**MODELS**: **slerp:** 2 models exactly, one of the 2 also listed as *Base* | **nuslerp:** 2 models exactly; *Base*: optional | **multislerp:** 2+ models; *Base*: optional | **karcher:** 2+ models; *Base*: none")
m1, w1 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"); m2, w2 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0")
with gr.Accordion("More", open=False):
m3, w3 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); m4, w4 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); m5, w5 = gr.Textbox(label="Model 5"), gr.Textbox(label="Weight 5", value="1.0")
t5_out = gr.Textbox(label="Output Repo"); t5_priv = gr.Checkbox(label="Private", value=True)
t5_btn = gr.Button("Execute")
t5_res = gr.Textbox(label="Result", lines=10)
t5_btn.click(wrapper_amphinterpolative, [t5_token, t5_method, t5_base, t5_t, t5_norm, t5_i8, t5_flat, t5_row, t5_eps, t5_iter, t5_tol, m1, w1, m2, w2, m3, w3, m4, w4, m5, w5, t5_out, t5_priv, t5_shard, t5_prec, t5_tok, t5_chat], t5_res)
# --- TAB 6 ---
with gr.Tab("Stir/Tie Bases"):
gr.Markdown("### Task Vector Methods Family: task_arithmetic, ties, dare_ties, dare_linear, della, della_linear, breadcrumbs, breadcrumbs_ties, sce")
t6_token = gr.Textbox(label="Token", type="password")
t6_method = gr.Dropdown(["task_arithmetic", "ties", "dare_ties", "dare_linear", "della", "della_linear", "breadcrumbs", "breadcrumbs_ties", "sce"], value="ties", label="Merge Method")
gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
with gr.Row():
t6_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t6_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t6_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t6_chat = gr.Textbox(label="Chat Template", placeholder="auto")
gr.Markdown("Built-in **Chat Templates**: alpaca, chatml, llama3, mistral, exaone, auto (default)")
t6_base = gr.Textbox(label="Base Model (required)")
gr.Markdown("**MODELS**: These methods all accept **2 or more models**, and require one of these designated as *Base*")
with gr.Row():
t6_norm = gr.Checkbox(label="Normalize Weights", value=True); t6_i8 = gr.Checkbox(label="Int8 Mask", value=False); t6_resc = gr.Checkbox(label="Rescale (Dare_Linear)", value=True); t6_lamb = gr.Number(label="Lambda", value=1.0); t6_topk = gr.Slider(0, 1, 1.0, label="Select TopK (SCE)")
m1_6, w1_6 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"); d1_6, g1_6, e1_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
m2_6, w2_6 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0"); d2_6, g2_6, e2_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
with gr.Accordion("More", open=False):
m3_6, w3_6 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); d3_6, g3_6, e3_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
m4_6, w4_6 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); d4_6, g4_6, e4_6 = gr.Textbox(label="Density (DARE/TIES)", value="1.0"), gr.Number(label="Gamma (breadcrumbs)", value=0.01), gr.Number(label="Epsilon (DELLA)", value=0.15)
t6_out = gr.Textbox(label="Output Repo"); t6_priv = gr.Checkbox(label="Private", value=True)
t6_btn = gr.Button("Execute")
t6_res = gr.Textbox(label="Result", lines=10)
t6_btn.click(wrapper_stirtie, [t6_token, t6_method, t6_base, t6_norm, t6_i8, t6_lamb, t6_resc, t6_topk, m1_6, w1_6, d1_6, g1_6, e1_6, m2_6, w2_6, d2_6, g2_6, e2_6, m3_6, w3_6, d3_6, g3_6, e3_6, m4_6, w4_6, d4_6, g4_6, e4_6, t6_out, t6_priv, t6_shard, t6_prec, t6_tok, t6_chat], t6_res)
# --- TAB 7 ---
with gr.Tab("Specious"):
gr.Markdown("### Specialized Methods: model_stock, nearswap, arcee_fusion, passthrough")
t7_token = gr.Textbox(label="Token", type="password")
t7_method = gr.Dropdown(["model_stock", "nearswap", "arcee_fusion", "passthrough", "linear"], value="model_stock", label="Merge Method")
gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
with gr.Row():
t7_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t7_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t7_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t7_chat = gr.Textbox(label="Chat Template", placeholder="auto")
gr.Markdown("Built-in **Chat Templates**: alpaca, chatml, llama3, mistral, exaone, auto (default)")
t7_base = gr.Textbox(label="Base Model (required for nearswap/arcee_fusion/model_stock)", placeholder="org/base-model")
gr.Markdown("**MODELS**: **passthrough:** 1 model acc. to Docs, but [Examples](https://github.com/arcee-ai/mergekit/tree/main/examples) shows 2+ | **nearswap/arcee_fusion:** 2 models, one also listed as *Base* | **model_stock:** 3+ models, one also listed as *Base*")
with gr.Row():
t7_norm = gr.Checkbox(label="Normalize", value=True); t7_i8 = gr.Checkbox(label="Int8 Mask", value=False); t7_t = gr.Slider(0, 1, 0.5, label="t (Interpolation Ratio, for Nearswap)"); t7_filt_w = gr.Checkbox(label="Filter Wise (for Model_Stock)", value=False)
m1_7, w1_7, f1_7 = gr.Textbox(label="Model 1"), gr.Textbox(label="Weight 1", value="1.0"), gr.Textbox(label="Filter Model Component")
m2_7, w2_7 = gr.Textbox(label="Model 2"), gr.Textbox(label="Weight 2", value="1.0")
with gr.Accordion("More", open=False):
m3_7, w3_7 = gr.Textbox(label="Model 3"), gr.Textbox(label="Weight 3", value="1.0"); m4_7, w4_7 = gr.Textbox(label="Model 4"), gr.Textbox(label="Weight 4", value="1.0"); m5_7, w5_7 = gr.Textbox(label="Model 5"), gr.Textbox(label="Weight 5", value="1.0")
t7_out = gr.Textbox(label="Output Repo"); t7_priv = gr.Checkbox(label="Private", value=True)
t7_btn = gr.Button("Execute")
t7_res = gr.Textbox(label="Result", lines=10)
t7_btn.click(wrapper_specious, [t7_token, t7_method, t7_base, t7_norm, t7_i8, t7_t, t7_filt_w, m1_7, w1_7, f1_7, m2_7, w2_7, m3_7, w3_7, m4_7, w4_7, m5_7, w5_7, t7_out, t7_priv, t7_shard, t7_prec, t7_tok, t7_chat], t7_res)
# --- TAB 8 (MoEr) ---
with gr.Tab("MoEr"):
gr.Markdown("### Mixture of Experts: fuses self-attention & normalization layers from *Base* w/MLP layers from *Experts*")
gr.Markdown("See [MergeKit MoE doc](https://github.com/arcee-ai/mergekit/blob/main/docs/moe.md) for more info.")
t8_token = gr.Textbox(label="Token", type="password")
with gr.Row():
t8_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t8_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t8_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t8_chat = gr.Textbox(label="Chat Template", placeholder="auto")
t8_base = gr.Textbox(label="Base Model (Required)"); t8_gate = gr.Dropdown(["cheap_embed", "random", "hidden"], value="cheap_embed", label="Gate Mode"); t8_dtype = gr.Dropdown(["float16", "bfloat16"], value="bfloat16", label="Internal Dtype")
gr.Markdown("#### Experts (at least 2 required)")
gr.Markdown("Prompts are comma-separated descriptors for each expert")
with gr.Row():
t8_expert1 = gr.Textbox(label="Expert 1", placeholder="org/expert1")
t8_prompt1 = gr.Textbox(label="Positive Prompts", placeholder="math, reasoning, logic")
with gr.Row():
t8_expert2 = gr.Textbox(label="Expert 2", placeholder="org/expert2")
t8_prompt2 = gr.Textbox(label="Positive Prompts", placeholder="creative, writing, storytelling")
with gr.Row():
t8_expert3 = gr.Textbox(label="Expert 3 (optional)", placeholder="org/expert3")
t8_prompt3 = gr.Textbox(label="Positive Prompts", placeholder="code, programming")
with gr.Row():
t8_expert4 = gr.Textbox(label="Expert 4 (optional)", placeholder="org/expert4")
t8_prompt4 = gr.Textbox(label="Positive Prompts", placeholder="")
with gr.Row():
t8_expert5 = gr.Textbox(label="Expert 5 (optional)", placeholder="org/expert5")
t8_prompt5 = gr.Textbox(label="Positive Prompts", placeholder="")
t8_out = gr.Textbox(label="Output Repo"); t8_priv = gr.Checkbox(label="Private", value=True)
t8_btn = gr.Button("Build MoE")
t8_res = gr.Textbox(label="Result", lines=10)
t8_btn.click(wrapper_moer, [t8_token, t8_base, t8_expert1, t8_prompt1, t8_expert2, t8_prompt2, t8_expert3, t8_prompt3, t8_expert4, t8_prompt4, t8_expert5, t8_prompt5, t8_gate, t8_dtype, t8_out, t8_priv, t8_shard, t8_prec, t8_tok, t8_chat], t8_res)
# --- TAB 9 (Rawer) ---
with gr.Tab("Rawer"):
gr.Markdown("### Raw PyTorch MergeKit / Non-pipeline-classed Models")
t9_token = gr.Textbox(label="Token", type="password"); t9_models = gr.TextArea(label="Models (one per line)")
with gr.Row():
t9_shard = gr.Slider(label="Max Shard Size (GB)", value=5.0, minimum=0.5, maximum=20.0); t9_prec = gr.Dropdown(["float16", "bfloat16", "float32"], value="bfloat16", label="Output Precision"); t9_tok = gr.Dropdown(["base", "union", "model:path"], value="base", label="Tokenizer Source"); t9_chat = gr.Textbox(label="Chat Template", placeholder="auto")
gr.Markdown("Built-in Chat Templates: alpaca, chatml, llama3, mistral, exaone, auto")
gr.Markdown("See [MergeKit Merge Method Docs](https://github.com/arcee-ai/mergekit/blob/main/docs/merge_methods.md) for more info.")
t9_method = gr.Dropdown(["linear", "passthrough"], value="linear", label="Merge Method"); t9_dtype = gr.Dropdown(["float32", "float16", "bfloat16"], value="float32", label="Config dtype")
t9_out = gr.Textbox(label="Output Repo"); t9_priv = gr.Checkbox(label="Private", value=True)
t9_btn = gr.Button("Merge Raw")
t9_res = gr.Textbox(label="Result", lines=10)
t9_btn.click(wrapper_rawer, [t9_token, t9_models, t9_method, t9_dtype, t9_out, t9_priv, t9_shard, t9_prec, t9_tok, t9_chat], t9_res)
# --- TAB 10 ---
with gr.Tab("Mario,DARE!"):
gr.Markdown("### Model-Agnostic DARE Implementation (Drop And REscale)")
gr.Markdown("From [sft-merger by Martyn Garcia](https://github.com/martyn)")
t10_token = gr.Textbox(label="Token", type="password")
gr.Markdown(
"""
### How DARE Works:
1. **Compute Delta**: Difference between fine-tuned and base weights
2. **Drop Elements**: Randomly mask out delta values based on mask rate
3. **Rescale**: Compensate for dropped elements by rescaling remaining values
4. **Apply**: Add scaled delta back to base model
**Mask Rate**: 0.5 = drop 50% of delta values, 0.9 = drop 90% (more aggressive sparsification)
"""
)
with gr.Row():
t10_base = gr.Textbox(label="Base Model", placeholder="org/base-model"); t10_ft = gr.Textbox(label="Fine-Tuned Model", placeholder="org/fine-tuned-model")
with gr.Row():
t10_ratio = gr.Slider(value=1.0, minimum=0.0, maximum=2.0, step=0.1, label="Merge Ratio (delta weight)"); t10_mask = gr.Slider(value=0.5, minimum=0.0, maximum=0.99, step=0.01, label="Mask Rate (drop probability)")
t10_out = gr.Textbox(label="Output Repo"); t10_priv = gr.Checkbox(label="Private", value=True)
gr.Button("Run").click(task_dare_custom, [t10_token, t10_base, t10_ft, t10_ratio, t10_mask, t10_out, t10_priv], gr.Textbox(label="Result"))
if __name__ == "__main__":
demo.queue().launch(css=css, ssr_mode=False, mcp_server=True)