File size: 11,110 Bytes
0249933 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 | """Merge configuration YAML generator with presets and validation."""
from dataclasses import dataclass, field
from typing import Optional
import yaml
# ===== MERGE METHOD DEFINITIONS =====
MERGE_METHODS = {
"dare_ties": {
"name": "DARE-TIES",
"description": "Drop And REscale with TIES — trims low-magnitude parameters and resolves sign conflicts. Best for combining 2+ specialist models.",
"min_models": 2,
"max_models": 10,
"needs_base": True,
"params": ["weight", "density"],
"global_params": ["int8_mask", "normalize"],
"supports_slices": True,
},
"ties": {
"name": "TIES",
"description": "Trim, Elect Sign, Merge — resolves parameter interference between models. Similar to DARE-TIES but without the drop step.",
"min_models": 2,
"max_models": 10,
"needs_base": True,
"params": ["weight", "density"],
"global_params": ["int8_mask", "normalize"],
"supports_slices": True,
},
"slerp": {
"name": "SLERP",
"description": "Spherical Linear Interpolation — smoothly blends two models along a curved path in weight space. Best for two-model merges.",
"min_models": 2,
"max_models": 2,
"needs_base": False,
"params": [],
"global_params": ["t"],
"supports_slices": True,
},
"linear": {
"name": "Linear",
"description": "Simple weighted average of model parameters. Fast and predictable baseline.",
"min_models": 2,
"max_models": 10,
"needs_base": False,
"params": ["weight"],
"global_params": ["normalize"],
"supports_slices": True,
},
"task_arithmetic": {
"name": "Task Arithmetic",
"description": "Add or subtract task vectors from a base model. Use negative weights to remove capabilities.",
"min_models": 1,
"max_models": 10,
"needs_base": True,
"params": ["weight"],
"global_params": [],
"supports_slices": False,
},
"passthrough": {
"name": "Passthrough (Frankenmerge)",
"description": "Stack layers from different models. Can create larger models from smaller ones. Supports different layer counts.",
"min_models": 1,
"max_models": 10,
"needs_base": False,
"params": [],
"global_params": [],
"supports_slices": True,
"requires_slices": True,
},
}
# ===== PRESETS =====
@dataclass
class MergePreset:
name: str
description: str
method: str
weight_strategy: str # "equal", "first_dominant", "last_dominant", "auto_detect"
def apply(self, model_ids: list[str]) -> tuple[list[float], list[float]]:
"""Generate weights and densities for given models."""
n = len(model_ids)
if n == 0:
return [], []
if self.weight_strategy == "equal":
weights = [round(1.0 / n, 3)] * n
densities = [0.6] * n
elif self.weight_strategy == "first_dominant":
weights = [0.6] + [round(0.4 / (n - 1), 3)] * (n - 1) if n > 1 else [1.0]
densities = [0.7] + [0.5] * (n - 1)
elif self.weight_strategy == "last_dominant":
weights = [round(0.4 / (n - 1), 3)] * (n - 1) + [0.6] if n > 1 else [1.0]
densities = [0.5] * (n - 1) + [0.7]
elif self.weight_strategy == "auto_detect":
weights, densities = _auto_detect_weights(model_ids)
else:
weights = [round(1.0 / n, 3)] * n
densities = [0.6] * n
return weights, densities
def _auto_detect_weights(model_ids: list[str]) -> tuple[list[float], list[float]]:
"""Auto-detect optimal weights based on model names/tags."""
n = len(model_ids)
weights = []
densities = []
for mid in model_ids:
name = mid.lower()
if "code" in name or "coder" in name:
weights.append(0.5)
densities.append(0.7)
elif "math" in name:
weights.append(0.4)
densities.append(0.6)
elif "instruct" in name and "code" not in name:
weights.append(0.3)
densities.append(0.5)
else:
weights.append(0.3)
densities.append(0.5)
# Normalize weights to sum to 1
total = sum(weights)
if total > 0:
weights = [round(w / total, 3) for w in weights]
return weights, densities
PRESETS = {
"equal": MergePreset("Equal", "Equal weights for all models", "dare_ties", "equal"),
"first_dominant": MergePreset("First Model Dominant", "Prioritize the first model", "dare_ties", "first_dominant"),
"last_dominant": MergePreset("Last Model Dominant", "Prioritize the last model", "dare_ties", "last_dominant"),
"coding_focus": MergePreset("Coding Focus", "Higher weight for code-related models", "dare_ties", "auto_detect"),
"balanced_slerp": MergePreset("Balanced SLERP", "50/50 interpolation between two models", "slerp", "equal"),
}
# ===== CONFIG GENERATION =====
@dataclass
class MergeConfig:
"""Complete merge configuration."""
method: str = "dare_ties"
models: list[str] = field(default_factory=list)
base_model: str = ""
weights: list[float] = field(default_factory=list)
densities: list[float] = field(default_factory=list)
tokenizer_source: str = ""
dtype: str = "bfloat16"
# Method-specific params
slerp_t: float = 0.5
int8_mask: bool = True
normalize: bool = True
# Passthrough/slice params
slices: list[dict] = field(default_factory=list)
# Output
output_name: str = ""
def validate(self) -> list[str]:
"""Validate the configuration. Returns list of error messages."""
errors = []
method_info = MERGE_METHODS.get(self.method)
if not method_info:
errors.append(f"Unknown merge method: {self.method}")
return errors
n = len(self.models)
if n < method_info["min_models"]:
errors.append(f"{method_info['name']} requires at least {method_info['min_models']} models")
if n > method_info["max_models"]:
errors.append(f"{method_info['name']} supports at most {method_info['max_models']} models")
if method_info["needs_base"] and not self.base_model:
errors.append(f"{method_info['name']} requires a base_model")
if "weight" in method_info["params"]:
if self.weights and len(self.weights) != n:
errors.append(f"Expected {n} weights, got {len(self.weights)}")
if self.weights and any(w < -1 or w > 2 for w in self.weights):
errors.append("Weights should be between -1 and 2")
if "density" in method_info["params"]:
if self.densities and len(self.densities) != n:
errors.append(f"Expected {n} densities, got {len(self.densities)}")
if self.densities and any(d < 0 or d > 1 for d in self.densities):
errors.append("Densities must be between 0 and 1")
if self.method == "slerp" and (self.slerp_t < 0 or self.slerp_t > 1):
errors.append("SLERP t parameter must be between 0 and 1")
if method_info.get("requires_slices") and not self.slices:
errors.append(f"{method_info['name']} requires slice definitions")
return errors
def generate_yaml(config: MergeConfig) -> str:
"""Generate mergekit-compatible YAML configuration.
Args:
config: MergeConfig with all parameters
Returns:
YAML string ready for mergekit
"""
errors = config.validate()
if errors:
return f"# VALIDATION ERRORS:\n" + "\n".join(f"# - {e}" for e in errors)
method_info = MERGE_METHODS[config.method]
doc = {}
# Passthrough uses slices format
if config.method == "passthrough":
doc["slices"] = config.slices or _default_slices(config)
doc["merge_method"] = config.method
doc["dtype"] = config.dtype
return yaml.dump(doc, default_flow_style=False, sort_keys=False)
# Standard methods
doc["merge_method"] = config.method
if method_info["needs_base"]:
doc["base_model"] = config.base_model
# Models with parameters
if config.method == "slerp":
doc["models"] = [{"model": m} for m in config.models]
doc["parameters"] = {"t": config.slerp_t}
else:
models_list = []
for i, model_id in enumerate(config.models):
entry = {"model": model_id}
params = {}
if "weight" in method_info["params"] and config.weights:
params["weight"] = config.weights[i]
if "density" in method_info["params"] and config.densities:
params["density"] = config.densities[i]
if params:
entry["parameters"] = params
models_list.append(entry)
doc["models"] = models_list
# Global parameters
global_params = {}
if "int8_mask" in method_info.get("global_params", []):
global_params["int8_mask"] = config.int8_mask
if "normalize" in method_info.get("global_params", []):
global_params["normalize"] = config.normalize
if global_params:
doc["parameters"] = global_params
doc["dtype"] = config.dtype
if config.tokenizer_source:
doc["tokenizer_source"] = config.tokenizer_source
return yaml.dump(doc, default_flow_style=False, sort_keys=False)
def _default_slices(config: MergeConfig) -> list[dict]:
"""Generate default slice config for passthrough merges."""
slices = []
for model_id in config.models:
slices.append({
"sources": [{"model": model_id, "layer_range": [0, 32]}]
})
return slices
def generate_from_preset(
preset_name: str,
model_ids: list[str],
base_model: str = "",
tokenizer_source: str = "",
dtype: str = "bfloat16",
) -> str:
"""Quick config generation from a preset name.
Args:
preset_name: Key from PRESETS dict
model_ids: List of model IDs to merge
base_model: Base model for methods that need one
tokenizer_source: Which model's tokenizer to use
dtype: Data type for merge
Returns:
YAML string
"""
preset = PRESETS.get(preset_name)
if not preset:
return f"# Unknown preset: {preset_name}\n# Available: {', '.join(PRESETS.keys())}"
weights, densities = preset.apply(model_ids)
config = MergeConfig(
method=preset.method,
models=model_ids,
base_model=base_model or (model_ids[0] if model_ids else ""),
weights=weights,
densities=densities,
tokenizer_source=tokenizer_source or base_model or (model_ids[0] if model_ids else ""),
dtype=dtype,
)
return generate_yaml(config)
def get_method_info(method: str) -> dict:
"""Get human-readable info about a merge method."""
return MERGE_METHODS.get(method, {"name": "Unknown", "description": "Unknown method"})
|