File size: 12,974 Bytes
d91766b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import json
import torch
import torch.nn as nn

from tqdm import tqdm
from glob import glob
from functools import partial
from safetensors import safe_open
from diffulex.config import Config
from diffulex.logger import get_logger
from diffulex.utils.checkpoint import LoadContext, ResolvedWeight

logger = get_logger(__name__)


def load_lora_config(lora_path: str) -> dict:
    """Load LoRA configuration from adapter_config.json."""
    config_path = os.path.join(lora_path, "adapter_config.json")
    if os.path.exists(config_path):
        with open(config_path, "r") as f:
            return json.load(f)
    return {}


def enable_lora_for_model(model: nn.Module, lora_config: dict):
    """Enable LoRA for existing linear layers in the model."""
    r = lora_config.get("r", 16)
    lora_alpha = lora_config.get("lora_alpha", 32.0)
    lora_dropout = lora_config.get("lora_dropout", 0.0)
    target_modules = lora_config.get("target_modules", [])

    for name, module in model.named_modules():
        if hasattr(module, "__init_lora__"):
            should_apply = True
            if target_modules:
                leaf = name.split(".")[-1] if name else name
                should_apply = any(target == leaf for target in target_modules)
            if should_apply:
                module.__init_lora__(r, lora_alpha, lora_dropout)
    return model


def default_weight_loader(param: nn.Parameter, loaded_weight: torch.Tensor):
    param.data.copy_(loaded_weight)


def resolve_weight_spec(
    model: nn.Module,
    weight_name: str,
    *,
    config: Config,
    named_modules: dict[str, nn.Module] | None = None,
) -> ResolvedWeight | None:
    if named_modules is None:
        named_modules = dict(model.named_modules())

    ctx = LoadContext(config=config, full_name=weight_name)
    parts = weight_name.split(".")
    for i in range(len(parts), 0, -1):
        prefix = ".".join(parts[:i])
        module = named_modules.get(prefix)
        if module is None:
            continue

        resolver = getattr(module, "resolve_checkpoint_weight", None)
        if resolver is None:
            continue

        suffix = ".".join(parts[i:])
        spec = resolver(suffix, ctx)
        if spec is not None:
            return spec

    root_resolver = getattr(model, "resolve_checkpoint_weight", None)
    if root_resolver is not None:
        return root_resolver(weight_name, ctx)
    return None


def apply_resolved_weight(spec: ResolvedWeight, loaded_weight: torch.Tensor):
    if spec.skip:
        return

    if spec.transform is not None:
        loaded_weight = spec.transform(loaded_weight)

    if spec.loader is not None:
        spec.loader(loaded_weight)
        return

    if spec.param is not None:
        weight_loader = getattr(spec.param, "weight_loader", default_weight_loader)
        if spec.shard_id is None:
            weight_loader(spec.param, loaded_weight)
        else:
            weight_loader(spec.param, loaded_weight, spec.shard_id)
        return

    if spec.buffer is not None:
        spec.buffer.copy_(loaded_weight)
        return

    raise ValueError("ResolvedWeight must specify loader, param, buffer, or skip.")


def try_load_direct(model: nn.Module, weight_name: str, loaded_weight: torch.Tensor) -> bool:
    try:
        param = model.get_parameter(weight_name)
        weight_loader = getattr(param, "weight_loader", default_weight_loader)
        weight_loader(param, loaded_weight)
        return True
    except (AttributeError, KeyError):
        pass

    try:
        buffer = model.get_buffer(weight_name)
        buffer.copy_(loaded_weight)
        return True
    except (AttributeError, KeyError):
        return False


def try_load_via_packed_mapping(
    model: nn.Module,
    packed_modules_mapping: dict,
    weight_name: str,
    loaded_weight: torch.Tensor,
    config: Config,
) -> bool:
    for k in packed_modules_mapping:
        if k not in weight_name:
            continue

        if config.model_name == "llada" and k == "ff_out" and "transformer.ff_out" in weight_name:
            continue
        elif config.model_name == "llada" and k == "transformer.ff_out":
            v, shard_id = packed_modules_mapping[k]
            assert v == "lm_head"
            param_name = "lm_head.weight"
        else:
            v, shard_id = packed_modules_mapping[k]
            param_name = weight_name.replace(k, v)

        if "layernorm" in param_name:
            try:
                param = model.get_parameter(param_name)
                weight_loader = getattr(param, "weight_loader", default_weight_loader)
                weight_loader(param, loaded_weight)
            except (AttributeError, KeyError):
                try:
                    buffer = model.get_buffer(param_name)
                    buffer.copy_(loaded_weight)
                except (AttributeError, KeyError):
                    pass
        else:
            try:
                param = model.get_parameter(param_name)
                weight_loader = partial(
                    getattr(param, "weight_loader"),
                    param,
                    loaded_weight,
                )
                if shard_id is None:
                    weight_loader()
                else:
                    weight_loader(shard_id)
            except (AttributeError, KeyError):
                pass
        return True

    return False


def load_model(model: nn.Module, config: Config):
    """Load model weights and optionally LoRA weights."""
    # Enable LoRA for linear layers if LoRA is enabled
    if config.use_lora and config.lora_path:
        lora_config = load_lora_config(config.lora_path)
        if lora_config:
            logger.info(f"LoRA Config Loaded: {lora_config}")
            model = enable_lora_for_model(model, lora_config)
        else:
            logger.info("No adapter_config.json found, using default LoRA parameters")
            default_config = {"r": 16, "lora_alpha": 32.0, "lora_dropout": 0.0}
            model = enable_lora_for_model(model, default_config)

    # Load base model weights
    packed_modules_mapping = getattr(model, "packed_modules_mapping", {})
    named_modules = dict(model.named_modules())
    for file in tqdm(glob(os.path.join(config.model, "*.safetensors")), desc="Loading base model"):
        with safe_open(file, "pt", "cpu") as f:
            for weight_name in f.keys():
                loaded_weight = f.get_tensor(weight_name)

                spec = resolve_weight_spec(
                    model,
                    weight_name,
                    config=config,
                    named_modules=named_modules,
                )
                if spec is not None:
                    apply_resolved_weight(spec, loaded_weight)
                    continue

                if try_load_via_packed_mapping(model, packed_modules_mapping, weight_name, loaded_weight, config):
                    continue

                try_load_direct(model, weight_name, loaded_weight)

    # Load LoRA weights if enabled
    if config.use_lora and config.lora_path:
        if os.path.exists(config.lora_path):
            logger.info(f"Loading LoRA weights from {config.lora_path}")
            model = load_lora_weights(
                model,
                config.lora_path,
                packed_modules_mapping=packed_modules_mapping if config.model_name == "llada" else None,
                pre_merge_lora=getattr(config, "pre_merge_lora", False),
            )
        else:
            logger.warning(f"LoRA path {config.lora_path} does not exist, skipping LoRA loading")

    return model


def load_lora_weights(
    model: nn.Module,
    lora_path: str,
    packed_modules_mapping: dict | None = None,
    pre_merge_lora: bool = False,
):
    """Load LoRA weights into LoRA-enabled layers.

    Args:
        model: The model with LoRA-enabled linear layers.
        lora_path: Path to LoRA checkpoint.
        packed_modules_mapping: Optional mapping for packed modules (e.g. llada lm_head).
        pre_merge_lora: If True, merge LoRA into base weights after loading so that
            forward does not need to run LoRA computation each time. If False, keep
            LoRA separate and apply it in lora_forward during each forward pass.
    """
    try:
        lora_config = load_lora_config(lora_path)
        target_modules = lora_config.get("target_modules", [])

        lora_weights = {}

        for file in tqdm(glob(os.path.join(lora_path, "*.safetensors")), desc="Loading LoRA"):
            with safe_open(file, "pt", "cpu") as f:
                for weight_name in f.keys():
                    lora_weights[weight_name] = f.get_tensor(weight_name)

        applied_count = 0

        modified_modules = None
        if packed_modules_mapping is not None:
            modified_modules = [v for k, (v, _) in packed_modules_mapping.items() if k in target_modules]
            rev_mapping = {v: k for k, (v, _) in packed_modules_mapping.items()}

        for name, module in model.named_modules():
            if hasattr(module, "lora_A") and hasattr(module, "lora_B"):
                should_apply = True

                if modified_modules is not None:
                    modified_module_type = ".".join(name.split(".")[-2:])
                    org_module_type = rev_mapping[modified_module_type]
                    org_name = name.replace(modified_module_type, org_module_type)
                    should_apply = any(target in modified_module_type for target in modified_modules)
                elif target_modules:
                    module_type = name.split(".")[-1] if "." in name else name
                    should_apply = any(target in module_type for target in target_modules)

                if not should_apply:
                    continue

                base_patterns = (
                    [
                        name,
                        f"base_model.model.{name}",
                        f"model.{name}",
                    ]
                    if modified_modules is None
                    else [
                        org_name,
                        f"base_model.model.{org_name}",
                        f"model.{org_name}",
                    ]
                )

                found_a = found_b = None
                for base_name in base_patterns:
                    lora_a_keys = [
                        f"{base_name}.lora_A.weight",
                        f"{base_name}.lora_A.default.weight",
                        f"{base_name}.lora_A",
                    ]
                    lora_b_keys = [
                        f"{base_name}.lora_B.weight",
                        f"{base_name}.lora_B.default.weight",
                        f"{base_name}.lora_B",
                    ]

                    for key in lora_a_keys:
                        if key in lora_weights:
                            found_a = lora_weights[key]
                            break
                    for key in lora_b_keys:
                        if key in lora_weights:
                            found_b = lora_weights[key]
                            break

                    if found_a is not None and found_b is not None:
                        break

                if found_a is not None and found_b is not None:
                    if hasattr(module, "tp_size") and module.tp_size > 1:
                        if hasattr(module, "tp_dim") and module.tp_dim == 0:
                            shard_size = found_b.size(0) // module.tp_size
                            start_idx = module.tp_rank * shard_size
                            found_b = found_b[start_idx : start_idx + shard_size]
                        elif hasattr(module, "tp_dim") and module.tp_dim == 1:
                            shard_size = found_a.size(1) // module.tp_size
                            start_idx = module.tp_rank * shard_size
                            found_a = found_a[:, start_idx : start_idx + shard_size]

                    try:
                        module.lora_A.data.copy_(found_a)
                        module.lora_B.data.copy_(found_b)
                        applied_count += 1
                    except Exception as e:
                        logger.warning(f"Failed to load LoRA weights for {name}: {e}")

        if pre_merge_lora:
            mergeable_modules = [module for module in model.modules() if hasattr(module, "merge_lora")]
            for module in tqdm(mergeable_modules, desc="Merging LoRA"):
                module.merge_lora()
            logger.info(f"LoRA weights applied to {applied_count} layers and merged into base")
        else:
            logger.info(f"LoRA weights applied to {applied_count} layers (unmerged, applied per forward)")

    except Exception as e:
        logger.error(f"Error loading LoRA weights: {e}")
        logger.warning("Continuing with base model only")

    return model