resume from optimizer checkpoint only

* fix: allow trainer builder to use custom jinja chat template

* chore: use get_chat_template_from_config

Co-authored-by: Chirag Jain <jain.chirag925@gmail.com>

* fix: swap imports

---------

Co-authored-by: Chirag Jain <jain.chirag925@gmail.com>

requirements.txt

@@ -61,4 +61,4 @@ antlr4-python3-runtime==4.13.2
 torchao==0.7.0
 schedulefree==1.3.0
 
-axolotl-contribs-lgpl==0.0.1b2
+axolotl-contribs-lgpl==0.0.2

src/axolotl/cli/main.py

@@ -93,7 +93,7 @@ def evaluate(config: str, accelerate: bool, **kwargs):
 @click.argument("config", type=click.Path(exists=True, path_type=str))
 @click.option(
     "--accelerate/--no-accelerate",
-    default=True,
+    default=False,
     help="Use accelerate launch for multi-GPU inference",
 )
 @click.option(
@@ -124,7 +124,7 @@ def inference(
     if lora_model_dir:
         kwargs["lora_model_dir"] = lora_model_dir
     if base_model:
-        kwargs["output_dir"] = base_model
+        kwargs["base_model"] = base_model
 
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.inference"]

src/axolotl/core/trainer_builder.py

@@ -68,7 +68,7 @@ from axolotl.utils.callbacks import (
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
 from axolotl.utils.callbacks.profiler import PytorchProfilerCallback
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
@@ -424,6 +424,11 @@ class SchedulerMixin(Trainer):
 
         return self.lr_scheduler
 
+    def _load_optimizer_and_scheduler(self, checkpoint):
+        if not checkpoint and self.args.optimizer_checkpoint is not None:
+            checkpoint = self.args.optimizer_checkpoint
+        return super()._load_optimizer_and_scheduler(checkpoint)
+
 
 class AxolotlTrainer(SchedulerMixin, Trainer):
     """
@@ -1764,6 +1769,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
         training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
+        if self.cfg.optimizer_checkpoint:
+            training_arguments_kwargs[
+                "optimizer_checkpoint"
+            ] = self.cfg.optimizer_checkpoint
 
         if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
@@ -1834,8 +1843,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
         if self.cfg.chat_template:
-            training_arguments_kwargs["chat_template"] = get_chat_template(
-                self.cfg.chat_template,
+            training_arguments_kwargs["chat_template"] = get_chat_template_from_config(
+                cfg=self.cfg,
                 tokenizer=self.tokenizer,
             )
 

src/axolotl/train.py

@@ -1,5 +1,6 @@
 """Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
 
+import inspect
 import os
 import signal
 import sys
@@ -126,7 +127,20 @@ def train(
     )
 
     if cfg.fix_untrained_tokens:
-        fix_untrained_tokens(model, tokenizer, train_dataset)
+        # check if the `token_ids_to_fix` kwarg exists in the fix_untrained_tokens args
+        sig = inspect.signature(fix_untrained_tokens)
+        # if the function has the `token_ids_to_fix` arg, and fix_untrained_tokens is a list
+        if "token_ids_to_fix" in sig.parameters and isinstance(
+            cfg.fix_untrained_tokens, list
+        ):
+            fix_untrained_tokens(
+                model,
+                tokenizer,
+                train_dataset,
+                token_ids_to_fix=cfg.fix_untrained_tokens,
+            )
+        else:
+            fix_untrained_tokens(model, tokenizer, train_dataset)
         if cfg.local_rank == 0:
             model.save_pretrained(
                 str(Path(cfg.output_dir)), safe_serialization=safe_serialization

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -603,6 +603,8 @@ class AxolotlInputConfig(
     strict: Optional[bool] = Field(default=False)
     resume_from_checkpoint: Optional[str] = None
     auto_resume_from_checkpoints: Optional[bool] = None
+    optimizer_checkpoint: Optional[str] = None
+
     resize_token_embeddings_to_32x: Optional[bool] = None
     mean_resizing_embeddings: Optional[bool] = False
 
@@ -794,7 +796,7 @@ class AxolotlInputConfig(
     chat_template_jinja: Optional[str] = None
     default_system_message: Optional[str] = None
 
-    fix_untrained_tokens: Optional[bool] = None
+    fix_untrained_tokens: Optional[Union[int, List[int]]] = None
 
     # INTERNALS - document for now, generally not set externally
     is_preprocess: Optional[bool] = None