Repository navigation
Expand file tree
/
Copy pathlora.py
More file actions
376 lines (310 loc) · 14.9 KB
/
Copy pathlora.py
File metadata and controls
376 lines (310 loc) · 14.9 KB
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
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
"""PEFT-backed LoRA utilities for GLiNER2.
New code should use ``Extractor.apply_lora()`` and PEFT's own save/load
(``PeftModel.save_pretrained`` / ``PeftModel.from_pretrained``).
Every public symbol from the pre-PEFT era is preserved for backwards
compatibility. Each legacy entry point emits ``PendingDeprecationWarning``
on *call* (not on import) so existing user CI stays silent.
"""
from __future__ import annotations
import json
import logging
import warnings
from dataclasses import asdict, dataclass, field
from datetime import datetime, timezone
from pathlib import Path
from typing import Dict, List, Optional, Tuple, Union
import torch
import torch.nn as nn
from safetensors.torch import load_file, save_file
try:
from peft import LoraConfig as PeftLoraConfig, get_peft_model, PeftModel
from peft.tuners.lora.layer import LoraLayer as _PeftLoraLayer
except ModuleNotFoundError as exc:
if exc.name != "peft":
raise
raise ImportError(
"LoRA support requires the optional 'peft' package, which is not installed. "
'Install it with: pip install "gliner2[train]" for training, or '
'pip install "gliner2[local]" for inference.'
) from exc
from gliner2.training.lora_targets import ( # noqa: F401 - compatibility exports
ENCODER_PATTERNS,
TASK_MODULES,
_alias_targets,
_resolve_targets,
_task_module_names,
)
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Core (non-deprecated) — mirrors Gliner2Internal main
# ---------------------------------------------------------------------------
def _deprecation(name: str, replacement: str) -> None:
warnings.warn(
f"{name} is deprecated; use {replacement} instead.",
PendingDeprecationWarning,
stacklevel=3,
)
def _cast_lora_dtype(model: nn.Module) -> None:
"""Cast LoRA A/B weights to match their base-layer dtype (PR #4 fix)."""
for mod in model.modules():
if not isinstance(mod, _PeftLoraLayer):
continue
base = mod.get_base_layer()
if base is None:
continue
dtype = next(base.parameters()).dtype
for key in list(mod.lora_A):
mod.lora_A[key] = mod.lora_A[key].to(dtype=dtype)
for key in list(mod.lora_B):
mod.lora_B[key] = mod.lora_B[key].to(dtype=dtype)
# ---------------------------------------------------------------------------
# Preserved dataclasses (PendingDeprecationWarning on construction)
# ---------------------------------------------------------------------------
@dataclass
class LoRAConfig:
"""Configuration for LoRA parameter-efficient fine-tuning."""
enabled: bool = False
r: int = 8
alpha: float = 16.0
dropout: float = 0.0
target_modules: List[str] = field(default_factory=lambda: ["encoder"])
def __post_init__(self) -> None:
_deprecation("LoRAConfig", "peft.LoraConfig + Extractor.apply_lora()")
if self.r <= 0:
raise ValueError(f"LoRA rank must be > 0, got {self.r}")
if self.alpha <= 0:
raise ValueError(f"LoRA alpha must be > 0, got {self.alpha}")
if not 0 <= self.dropout < 1:
raise ValueError(f"LoRA dropout must be in [0, 1), got {self.dropout}")
if self.enabled and not self.target_modules:
raise ValueError("target_modules cannot be empty when LoRA is enabled")
@dataclass
class LoRAAdapterConfig:
"""Serializable metadata for a saved LoRA adapter."""
adapter_type: str = "lora"
adapter_version: str = "1.0"
lora_r: int = 8
lora_alpha: float = 16.0
lora_dropout: float = 0.0
target_modules: List[str] = field(default_factory=list)
created_at: str = ""
def save(self, path: Union[str, Path]) -> None:
path = Path(path)
path.mkdir(parents=True, exist_ok=True)
if not self.created_at:
self.created_at = datetime.now(timezone.utc).isoformat()
with open(path / "adapter_config.json", "w") as f:
json.dump(asdict(self), f, indent=2)
@classmethod
def load(cls, path: Union[str, Path]) -> LoRAAdapterConfig:
path = Path(path)
cfg_file = path / "adapter_config.json" if path.is_dir() else path
if not cfg_file.exists():
raise FileNotFoundError(f"Adapter config not found at {cfg_file}")
with open(cfg_file) as f:
data = json.load(f)
if "peft_type" in data:
return cls(
lora_r=data.get("r", 8),
lora_alpha=data.get("lora_alpha", 16.0),
lora_dropout=data.get("lora_dropout", 0.0),
target_modules=data.get("target_modules", []),
)
known = {f.name for f in cls.__dataclass_fields__.values()}
return cls(**{k: v for k, v in data.items() if k in known})
@classmethod
def is_adapter_path(cls, path: Union[str, Path]) -> bool:
path = Path(path)
if path.is_dir():
return (path / "adapter_config.json").exists()
return path.name == "adapter_config.json" and path.exists()
# isinstance-compat alias
LoRALayer = _PeftLoraLayer
# ---------------------------------------------------------------------------
# Legacy function shims (each emits PendingDeprecationWarning)
# ---------------------------------------------------------------------------
def apply_lora_to_model(
model: nn.Module,
config: LoRAConfig,
) -> Tuple[nn.Module, Dict[str, LoRALayer]]:
_deprecation("apply_lora_to_model", "Extractor.apply_lora()")
if not config.enabled:
return model, {}
peft_cfg = PeftLoraConfig(
r=config.r, lora_alpha=config.alpha, lora_dropout=config.dropout,
target_modules=_resolve_targets(model, config.target_modules),
bias="none",
)
peft_model = get_peft_model(model, peft_cfg)
_cast_lora_dtype(peft_model)
layers = {n: m for n, m in peft_model.named_modules() if isinstance(m, _PeftLoraLayer)}
return peft_model, layers
def get_lora_parameters(model: nn.Module) -> List[nn.Parameter]:
_deprecation("get_lora_parameters", "model.parameters() with requires_grad filter")
return [p for p in model.parameters() if p.requires_grad]
def get_lora_state_dict(model: nn.Module) -> Dict[str, torch.Tensor]:
_deprecation("get_lora_state_dict", "peft.get_peft_model_state_dict()")
from peft import get_peft_model_state_dict
if isinstance(model, PeftModel):
return dict(get_peft_model_state_dict(model))
return {}
def merge_lora_weights(model: nn.Module) -> int:
_deprecation("merge_lora_weights", "PeftModel.merge_adapter()")
if isinstance(model, PeftModel):
model.merge_adapter()
return sum(1 for m in model.modules() if isinstance(m, _PeftLoraLayer))
return 0
def unmerge_lora_weights(model: nn.Module) -> int:
_deprecation("unmerge_lora_weights", "PeftModel.unmerge_adapter()")
if isinstance(model, PeftModel):
model.unmerge_adapter()
return sum(1 for m in model.modules() if isinstance(m, _PeftLoraLayer))
return 0
def count_lora_parameters(model: nn.Module) -> Tuple[int, int, float]:
_deprecation("count_lora_parameters", "manual parameter counting")
lora_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
total_params = sum(p.numel() for p in model.parameters())
pct = (lora_params / total_params * 100) if total_params > 0 else 0.0
return lora_params, total_params, pct
def print_lora_info(model: nn.Module, config: LoRAConfig) -> None:
_deprecation("print_lora_info", "manual logging")
lp, tp, pct = count_lora_parameters.__wrapped__(model) if hasattr(count_lora_parameters, "__wrapped__") else _count_lora_params_raw(model)
n_layers = sum(1 for m in model.modules() if isinstance(m, _PeftLoraLayer))
print(f"LoRA: r={config.r}, alpha={config.alpha}, layers={n_layers}, "
f"trainable={lp:,}/{tp:,} ({pct:.2f}%)")
def _count_lora_params_raw(model: nn.Module) -> Tuple[int, int, float]:
lp = sum(p.numel() for p in model.parameters() if p.requires_grad)
tp = sum(p.numel() for p in model.parameters())
return lp, tp, (lp / tp * 100) if tp > 0 else 0.0
def remove_lora_from_model(model: nn.Module) -> nn.Module:
_deprecation("remove_lora_from_model", "PeftModel.merge_and_unload()")
if isinstance(model, PeftModel):
return model.merge_and_unload()
return model
# ---------------------------------------------------------------------------
# Dual-format save / load
# ---------------------------------------------------------------------------
def _is_peft_native_dir(path: Path) -> bool:
if (path / "adapter_model.safetensors").exists() or (path / "adapter_model.bin").exists():
return True
cfg_file = path / "adapter_config.json"
if cfg_file.exists():
with open(cfg_file) as f:
data = json.load(f)
if "peft_type" in data:
return True
return False
def save_lora_adapter(model: nn.Module, save_path: Union[str, Path]) -> None:
_deprecation("save_lora_adapter", "PeftModel.save_pretrained()")
save_path = Path(save_path)
save_path.mkdir(parents=True, exist_ok=True)
if isinstance(model, PeftModel):
peft_cfg = model.peft_config.get("default")
r = peft_cfg.r if peft_cfg else 8
alpha = peft_cfg.lora_alpha if peft_cfg else 16.0
dropout = peft_cfg.lora_dropout if peft_cfg else 0.0
target_mods = list(peft_cfg.target_modules) if peft_cfg and peft_cfg.target_modules else []
top_modules = sorted({n.split(".")[0] for n in target_mods if n})
legacy_state: Dict[str, torch.Tensor] = {}
for name, mod in model.named_modules():
if not isinstance(mod, _PeftLoraLayer):
continue
clean_name = name.replace("base_model.model.", "").replace("base_model.", "")
for adapter_key, mat in mod.lora_A.items():
legacy_state[f"{clean_name}.lora_A"] = mat.weight.data if hasattr(mat, "weight") else mat.data
for adapter_key, mat in mod.lora_B.items():
legacy_state[f"{clean_name}.lora_B"] = mat.weight.data if hasattr(mat, "weight") else mat.data
save_file(legacy_state, str(save_path / "adapter_weights.safetensors"))
cfg = LoRAAdapterConfig(
lora_r=r, lora_alpha=alpha, lora_dropout=dropout,
target_modules=top_modules,
)
cfg.save(save_path)
else:
raise ValueError("Model is not a PeftModel; nothing to save.")
def load_lora_adapter(
model: nn.Module,
adapter_path: Union[str, Path],
auto_unload: bool = True,
) -> Dict[str, LoRALayer]:
_deprecation("load_lora_adapter", "PeftModel.from_pretrained()")
adapter_path = Path(adapter_path)
if auto_unload and isinstance(model, PeftModel):
model = model.merge_and_unload()
if _is_peft_native_dir(adapter_path):
peft_model = PeftModel.from_pretrained(model, str(adapter_path))
_cast_lora_dtype(peft_model)
else:
adapter_cfg = LoRAAdapterConfig.load(adapter_path)
weights = load_file(str(adapter_path / "adapter_weights.safetensors"))
resolved = _resolve_targets(model, adapter_cfg.target_modules)
peft_cfg = PeftLoraConfig(
r=adapter_cfg.lora_r, lora_alpha=adapter_cfg.lora_alpha,
lora_dropout=adapter_cfg.lora_dropout,
target_modules=resolved, bias="none",
)
peft_model = get_peft_model(model, peft_cfg)
for name, mod in peft_model.named_modules():
if not isinstance(mod, _PeftLoraLayer):
continue
clean = name.replace("base_model.model.", "").replace("base_model.&q
76B1
uot;, "")
a_key, b_key = f"{clean}.lora_A", f"{clean}.lora_B"
if a_key in weights and b_key in weights:
base = mod.get_base_layer()
dtype = next(base.parameters()).dtype if base is not None else torch.float32
device = next(base.parameters()).device if base is not None else torch.device("cpu")
for adapter_name, mat in mod.lora_A.items():
if hasattr(mat, "weight"):
mat.weight.data.copy_(weights[a_key].to(device=device, dtype=dtype))
else:
mat.data.copy_(weights[a_key].to(device=device, dtype=dtype))
for adapter_name, mat in mod.lora_B.items():
if hasattr(mat, "weight"):
mat.weight.data.copy_(weights[b_key].to(device=device, dtype=dtype))
else:
mat.data.copy_(weights[b_key].to(device=device, dtype=dtype))
_cast_lora_dtype(peft_model)
layers = {n: m for n, m in peft_model.named_modules() if isinstance(m, _PeftLoraLayer)}
return layers
def unload_lora_adapter(model: nn.Module) -> int:
_deprecation("unload_lora_adapter", "PeftModel.merge_and_unload() or PeftModel.unload()")
if isinstance(model, PeftModel):
count = sum(1 for m in model.modules() if isinstance(m, _PeftLoraLayer))
model.unload()
return count
# `model` is not itself a `PeftModel` instance whenever it went through
# load_lora_adapter()'s legacy (non-peft-native) path: that path calls
# get_peft_model(model, ...), which injects LoraLayer modules directly
# into `model`'s own submodule tree in place, then discards the
# PeftModel wrapper it returns. `model` stays a plain nn.Module even
# though it now contains live LoraLayer submodules, so the
# isinstance(model, PeftModel) check above is always False for that
# path and this function used to silently no-op — has_adapter/
# _lora_layers bookkeeping would reset, but the injected LoraLayer
# modules kept running in every subsequent forward pass. Detect and
# unwrap that case directly: replace each injected LoraLayer with its
# original frozen base layer (LoRA never modifies base-layer weights,
# so this exactly restores pre-adapter behavior).
count = 0
for parent in list(model.modules()):
for child_name, child in list(parent.named_children()):
if isinstance(child, _PeftLoraLayer):
setattr(parent, child_name, child.get_base_layer())
count += 1
return count
def has_lora_adapter(model: nn.Module) -> bool:
_deprecation("has_lora_adapter", "isinstance(model, PeftModel)")
return isinstance(model, PeftModel) or any(isinstance(m, _PeftLoraLayer) for m in model.modules())
def get_adapter_config(model: nn.Module) -> Optional[LoRAAdapterConfig]:
_deprecation("get_adapter_config", "model.peft_config")
if not isinstance(model, PeftModel):
return None
peft_cfg = model.peft_config.get("default")
if peft_cfg is None:
return None
return LoRAAdapterConfig(
lora_r=peft_cfg.r,
lora_alpha=peft_cfg.lora_alpha,
lora_dropout=peft_cfg.lora_dropout,
target_modules=sorted(peft_cfg.target_modules) if peft_cfg.target_modules else [],
)