swap tinymodels that have safetensors for some ci tests (#2641)

src/axolotl/train.py

@@ -2,6 +2,7 @@
 
 import importlib
 import inspect
+import logging
 import os
 import signal
 import sys
@@ -12,7 +13,6 @@ from typing import Any, Dict
 
 import torch
 import transformers.modelcard
-from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
@@ -42,7 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -63,7 +63,6 @@ def setup_model_and_tokenizer(
     # Load tokenizer
     LOG.debug(
         f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
     )
     tokenizer = load_tokenizer(cfg)
 

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -1,15 +1,36 @@
 """custom checkpointing utils"""
 
+import importlib
 from functools import partial
 
+from packaging import version
+
 from axolotl.utils.gradient_checkpointing.unsloth import (
     Unsloth_Offloaded_Gradient_Checkpointer,
 )
 
+transformers_version = version.parse(importlib.metadata.version("transformers"))
+if transformers_version > version.parse("4.51.3"):
+    from transformers.modeling_layers import GradientCheckpointingLayer
+
+    def uses_gc_layers(decoder_layer):
+        return isinstance(decoder_layer.func.__self__, GradientCheckpointingLayer)
+
+else:
+
+    def uses_gc_layers(_):
+        return False
+
 
 def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
     return Unsloth_Offloaded_Gradient_Checkpointer.apply(
         (
             decoder_layer.func.__self__