Switch to the Composer integration of LoRA (works with FSDP) (#886)

Author: dakinggg

llmfoundry/callbacks/hf_checkpointer.py

@@ -9,7 +9,7 @@
 import re
 import tempfile
 from pathlib import Path
-from typing import Optional, Sequence, Union
+from typing import Any, Dict, Optional, Sequence, Union
 
 import torch
 from composer.core import Callback, Event, State, Time, TimeUnit
@@ -203,14 +203,17 @@ def _save_checkpoint(self, state: State, logger: Logger):
             from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
 
             if state.is_model_ddp:
+                composer_model = state.model.module
                 original_model: PreTrainedModel = state.model.module.model
                 state_dict_model = state.model.module.model
                 original_tokenizer = state.model.module.tokenizer
             elif isinstance(state.model.model, FSDP):
+                composer_model = state.model
                 original_model: PreTrainedModel = state.model.model.module
                 state_dict_model = state.model.model
                 original_tokenizer = state.model.tokenizer
             else:
+                composer_model = state.model
                 original_model: PreTrainedModel = state.model.model
                 state_dict_model = state.model.model
                 original_tokenizer = state.model.tokenizer
@@ -237,10 +240,25 @@ def _save_checkpoint(self, state: State, logger: Logger):
                     copied_config.init_device = 'cpu'
 
                 log.debug(f'Creating new model instance')
-                # First create the model instance on meta device to avoid the
-                # initialization cost.
-                with init_empty_weights():
-                    new_model_instance = type(original_model)(copied_config)
+
+                if composer_model.using_peft:
+                    # We don't use meta here because the state dict does not contain the full
+                    # model, only the adapter weights.
+                    active_adapter = original_model.active_adapter
+                    base_model = original_model.get_base_model()
+                    new_base_model_instance = type(base_model)(copied_config)
+
+                    new_model_instance = type(original_model)(
+                        new_base_model_instance,
+                        original_model.peft_config[active_adapter])
+                else:
+                    # First create the model instance on meta device to avoid the
+                    # initialization cost.
+                    with init_empty_weights():
+                        new_model_instance = type(original_model)(copied_config)
+
+                new_model_instance.to(dtype=self.dtype)
+                new_model_instance.load_state_dict(state_dict)
 
                 # Then load the state dict in with "assign" so that the state dict
                 # is loaded properly even though the model is initially on meta device.
@@ -295,12 +313,24 @@ def _save_checkpoint(self, state: State, logger: Logger):
                         # TODO: Remove after mlflow fixes the bug that makes this necessary
                         import mlflow
                         mlflow.store._unity_catalog.registry.rest_store.get_feature_dependencies = lambda *args, **kwargs: ''
-                        mlflow_logger.save_model(
-                            flavor='transformers',
-                            transformers_model=components,
-                            path=local_save_path,
-                            **self.mlflow_logging_config,
-                        )
+                        model_saving_kwargs: Dict[str, Any] = {
+                            'path': local_save_path
+                        }
+                        if composer_model.using_peft:
+                            model_saving_kwargs['flavor'] = 'peft'
+                            model_saving_kwargs[
+                                'save_pretrained_dir'] = temp_save_dir
+                            model_saving_kwargs[
+                                'metadata'] = self.mlflow_logging_config[
+                                    'metadata']
+                        else:
+                            model_saving_kwargs['flavor'] = 'transformers'
+                            model_saving_kwargs[
+                                'transformers_model'] = components
+                            model_saving_kwargs.update(
+                                self.mlflow_logging_config)
+
+                        mlflow_logger.save_model(**model_saving_kwargs)
 
                         license_filename = _maybe_get_license_filename(
                             local_save_path)