| """Read historical BioMod serialized names without rewriting checkpoint files. |
| |
| Current interfaces use BioMod names. Legacy spellings are kept only here so |
| existing configuration dictionaries and tensor state dictionaries still load. |
| This module intentionally has no model or tensor-library imports. |
| """ |
|
|
| from typing import Dict |
|
|
|
|
| def normalize_biomod_decoder_type(decoder_type: str) -> str: |
| """Normalize the legacy decoder option without changing the attention backend.""" |
| if decoder_type == "performer": |
| return "biomod" |
| return decoder_type |
|
|
|
|
| def _migrate_legacy_biomod_name(value: str) -> str: |
| """Translate serialized pre-rename names without changing their semantics.""" |
| replacements = ( |
| ("PATHCH", "MODULE"), |
| ("Pathch", "Module"), |
| ("pathch", "module"), |
| ("PATCHES", "MODULES"), |
| ("Patches", "Modules"), |
| ("patches", "modules"), |
| ("PATCH", "MODULE"), |
| ("Patch", "Module"), |
| ("patch", "module"), |
| ) |
| for old, new in replacements: |
| value = value.replace(old, new) |
| return value |
|
|
|
|
| def migrate_legacy_biomod_config(config: Dict[str, object]) -> Dict[str, object]: |
| """Return a BioMod-compatible copy of an older serialized configuration.""" |
| def translate(value): |
| if isinstance(value, dict): |
| migrated = {} |
| for key, item in value.items(): |
| new_key = _migrate_legacy_biomod_name(key) if isinstance(key, str) else key |
| if new_key in migrated: |
| raise ValueError( |
| f"Legacy configuration migration would duplicate key {new_key!r}" |
| ) |
| migrated[new_key] = translate(item) |
| return migrated |
| if isinstance(value, list): |
| return [translate(item) for item in value] |
| if isinstance(value, tuple): |
| return tuple(translate(item) for item in value) |
| if isinstance(value, str): |
| return _migrate_legacy_biomod_name(value) |
| return value |
|
|
| migrated = translate(dict(config)) |
| if "decoder_type" in migrated: |
| migrated["decoder_type"] = normalize_biomod_decoder_type(migrated["decoder_type"]) |
| return migrated |
|
|
|
|
| def migrate_legacy_biomod_state_dict(state_dict): |
| """Translate old state-dict key names before PyTorch performs strict loading.""" |
| try: |
| migrated = state_dict.__class__() |
| except TypeError: |
| migrated = {} |
|
|
| for key, value in state_dict.items(): |
| new_key = _migrate_legacy_biomod_name(key) if isinstance(key, str) else key |
| if new_key in migrated: |
| |
| if key == new_key: |
| migrated[new_key] = value |
| continue |
| migrated[new_key] = value |
|
|
| if hasattr(state_dict, "_metadata"): |
| migrated._metadata = { |
| _migrate_legacy_biomod_name(key): value |
| for key, value in state_dict._metadata.items() |
| } |
| return migrated |
|
|
|
|
| def migrate_legacy_encoder_prefix(state_dict, prefix: str) -> None: |
| """Translate one legacy language-model subtree in place before loading. |
| |
| Explicit current keys take precedence if both spellings are present. |
| The caller supplies the full parent prefix, including its trailing dot. |
| """ |
| old_root = prefix + "performer" |
| new_root = prefix + "biomod" |
| old_prefix = old_root + "." |
| new_prefix = new_root + "." |
| for key in list(state_dict): |
| if isinstance(key, str) and key.startswith(old_prefix): |
| new_key = new_prefix + key[len(old_prefix):] |
| if new_key not in state_dict: |
| state_dict[new_key] = state_dict[key] |
| del state_dict[key] |
|
|
| metadata = getattr(state_dict, "_metadata", None) |
| if metadata is not None: |
| for key in list(metadata): |
| if key == old_root or key.startswith(old_prefix): |
| new_key = new_root + key[len(old_root):] |
| if new_key not in metadata: |
| metadata[new_key] = metadata[key] |
| del metadata[key] |
|
|
