[`from_pretrained`] Simpler code for peft (#25726)

* refactor complicated from pretrained for peft

* nits

* more nits

* Update src/transformers/modeling_utils.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* make tests happy

* fixup after merge

---------

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
This commit is contained in:
Arthur 2023-08-24 16:18:39 +02:00 committed by GitHub
parent 0a365c3e6a
commit fecf08560c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 19 additions and 29 deletions

View File

@ -2391,33 +2391,23 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
else:
commit_hash = getattr(config, "_commit_hash", None)
if is_peft_available() and _adapter_model_path is None:
maybe_adapter_model_path = find_adapter_config_file(
pretrained_model_name_or_path,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
_commit_hash=commit_hash,
)
elif is_peft_available() and _adapter_model_path is not None:
maybe_adapter_model_path = _adapter_model_path
else:
maybe_adapter_model_path = None
has_adapter_config = maybe_adapter_model_path is not None
if has_adapter_config:
if _adapter_model_path is not None:
adapter_model_id = _adapter_model_path
else:
with open(maybe_adapter_model_path, "r", encoding="utf-8") as f:
adapter_model_id = pretrained_model_name_or_path
pretrained_model_name_or_path = json.load(f)["base_model_name_or_path"]
if is_peft_available():
if _adapter_model_path is None:
_adapter_model_path = find_adapter_config_file(
pretrained_model_name_or_path,
cache_dir=cache_dir,
force_download=force_download,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
revision=revision,
subfolder=subfolder,
_commit_hash=commit_hash,
)
if _adapter_model_path is not None and os.path.isfile(_adapter_model_path):
with open(_adapter_model_path, "r", encoding="utf-8"):
_adapter_model_path = pretrained_model_name_or_path
# change device_map into a map if we passed an int, a str or a torch.device
if isinstance(device_map, torch.device):
@ -3246,9 +3236,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
if quantization_method_from_config == QuantizationMethod.GPTQ:
model = quantizer.post_init_model(model)
if has_adapter_config:
if _adapter_model_path is not None:
model.load_adapter(
adapter_model_id,
_adapter_model_path,
adapter_name=adapter_name,
revision=revision,
token=token,