Spaces:
Running
Running
File size: 8,695 Bytes
ec5ad0d | 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 | # Copyright (C) 2025 Arcee AI
# SPDX-License-Identifier: LGPL-3.0-only
from typing import Any, Dict, Iterable, List, Optional, Tuple, Union
import yaml
from pydantic import BaseModel, model_validator
from typing_extensions import Literal, TypeAlias
from mergekit.common import ModelReference
from mergekit.tokenizer.config import TokenizerConfig
ScalarOrGradient: TypeAlias = Union[float, List[float], str, bool] # ScalarOrGradient: TypeAlias = Union[float, List[float]]
class ConditionalParameter(BaseModel):
value: ScalarOrGradient
filter: Optional[str] = None
ParameterSetting: TypeAlias = Union[
ConditionalParameter, List[ConditionalParameter], ScalarOrGradient, str
]
def evaluate_setting(
tensor_name: str, setting: ParameterSetting, t: float = 0
) -> Optional[float]:
if isinstance(setting, (float, int, bool, str)):
return setting
elif isinstance(setting, list):
if all(isinstance(e, (int, float)) for e in setting):
scaled = t * (len(setting) - 1)
i0 = int(scaled)
i1 = min(len(setting) - 1, i0 + 1)
frac = scaled - i0
return (1 - frac) * setting[i0] + frac * setting[i1]
elif all(isinstance(e, (float, int, bool, str)) for e in setting):
return setting[int(t * (len(setting) - 1))]
else:
for cond in setting:
if (
(cond.filter is None)
or (cond.filter == "*")
or (tensor_name and cond.filter in tensor_name)
):
res = evaluate_setting(tensor_name, cond.value, t)
return res
else:
raise RuntimeError(f"Unexpected setting value: {setting}")
return None
class InputSliceDefinition(BaseModel):
model: ModelReference
layer_range: Tuple[int, int]
parameters: Optional[Dict[str, ParameterSetting]] = None
class InputModelDefinition(BaseModel):
model: ModelReference
parameters: Optional[Dict[str, ParameterSetting]] = None
class OutputSliceDefinition(BaseModel):
sources: List[InputSliceDefinition]
base_model: Optional[ModelReference] = None
residual_weight: Optional[float] = None
parameters: Optional[Dict[str, ParameterSetting]] = None
class OutputModuleDefinition(BaseModel):
slices: Optional[List[OutputSliceDefinition]] = None
models: Optional[List[InputModelDefinition]] = None
parameters: Optional[Dict[str, ParameterSetting]] = None
@model_validator(mode="after")
def validate_inputs(self):
if ((not self.slices) and (not self.models)) or (self.slices and self.models):
raise RuntimeError("Must specify either output slices or models to merge")
return self
class MergeConfiguration(BaseModel):
modules: Optional[Dict[str, OutputModuleDefinition]] = None
slices: Optional[List[OutputSliceDefinition]] = None
models: Optional[List[InputModelDefinition]] = None
merge_method: str
base_model: Optional[ModelReference] = None
dtype: Optional[str] = None
tokenizer_source: Union[Literal["union"], Literal["base"], ModelReference, None] = (
None
)
tokenizer: Optional[TokenizerConfig] = None
chat_template: Optional[str] = None
out_dtype: Optional[str] = None
parameters: Optional[Dict[str, ParameterSetting]] = None
def referenced_models(self) -> List[ModelReference]:
models = set()
if self.base_model:
models.add(self.base_model)
if self.models:
for model_in in self.models:
models.add(model_in.model)
if self.slices:
for s in self.slices:
for src in s.sources:
models.add(src.model)
if self.modules:
for m in self.modules.values():
if m.models:
for model_in in m.models:
models.add(model_in.model)
if m.slices:
for s in m.slices:
for src in s.sources:
models.add(src.model)
return list(models)
@model_validator(mode="after")
def validate_inputs(self):
set_ct = 0
if self.modules:
set_ct += 1
if self.slices:
set_ct += 1
if self.models:
set_ct += 1
if set_ct != 1:
raise RuntimeError(
"Exactly one of 'models', 'slices', or 'modules' must be present"
)
return self
@model_validator(mode="after")
def validate_tokenizer(self):
if self.tokenizer_source and self.tokenizer:
raise RuntimeError("Cannot specify both tokenizer_source and tokenizer")
return self
def to_yaml(self) -> str:
return yaml.dump(
self.model_dump(exclude_defaults=True, mode="json"),
Dumper=ConfigYamlDumper,
).rstrip()
class ConfigReader(BaseModel):
config: MergeConfiguration
t: float
tensor_name: Optional[str] = None
slice_out: Optional[OutputSliceDefinition] = None
module: Optional[OutputModuleDefinition] = None
@property
def base_model(self) -> Optional[ModelReference]:
if self.slice_out and self.slice_out.base_model:
res = self.slice_out.base_model
else:
res = self.config.base_model
return res
def for_out_slice(self, slice: OutputSliceDefinition) -> "ConfigReader":
return ConfigReader(
config=self.config,
t=self.t,
tensor_name=self.tensor_name,
slice_out=slice,
module=self.module,
)
def for_tensor(self, tensor_name: str) -> "ConfigReader":
return ConfigReader(
config=self.config,
t=self.t,
tensor_name=tensor_name,
slice_out=self.slice_out,
module=self.module,
)
def with_t(self, t: float) -> "ConfigReader":
return ConfigReader(
config=self.config,
t=t,
tensor_name=self.tensor_name,
slice_out=self.slice_out,
module=self.module,
)
def for_module(self, module: OutputModuleDefinition) -> "ConfigReader":
return ConfigReader(
config=self.config,
t=self.t,
tensor_name=self.tensor_name,
slice_out=self.slice_out,
module=module,
)
def parameter(
self,
name: str,
model: Optional[ModelReference] = None,
default: Any = None,
required: bool = False,
) -> Any:
if self.slice_out:
if model:
for s in self.slice_out.sources:
if s.model == model and s.parameters and name in s.parameters:
value = evaluate_setting(
self.tensor_name, s.parameters[name], self.t
)
if value is not None:
return value
if self.slice_out.parameters and name in self.slice_out.parameters:
value = evaluate_setting(
self.tensor_name, self.slice_out.parameters[name], self.t
)
if value is not None:
return value
if self.module and self.module.parameters and name in self.module.parameters:
value = evaluate_setting(
self.tensor_name,
self.module.parameters[name],
self.t,
)
if value is not None:
return value
if self.config.parameters and name in self.config.parameters:
value = evaluate_setting(
self.tensor_name,
self.config.parameters[name],
self.t,
)
if value is not None:
return value
if required:
path_paths = [str(s) for s in [model, self.tensor_name] if s]
p = ".".join(path_paths)
suffix = f" for {p}" if p else ""
raise RuntimeError(f"Missing required parameter {name}{suffix}")
return default
class ConfigYamlDumper(yaml.Dumper):
"""Custom YAML dumper to format lists of numbers in flow style."""
def represent_list(self, data: Iterable[Any]) -> yaml.SequenceNode:
flow_style = all(isinstance(e, (int, float)) for e in data)
return self.represent_sequence(
"tag:yaml.org,2002:seq", data, flow_style=flow_style
)
ConfigYamlDumper.add_representer(list, ConfigYamlDumper.represent_list)
|