fixes for larger models

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

@@ -533,6 +533,7 @@ class Config:
         Dict[Union[int, Literal["cpu", "disk"]], Union[int, str]]
     ] = None
     gpu_memory_limit: Optional[Union[int, str]] = None
+    low_cpu_mem_usage: Optional[bool] = None
 
     chat_template: Optional[ChatTemplate] = None
     default_system_message: Optional[str] = None

src/axolotl/utils/models.py

@@ -402,7 +402,9 @@ def load_model(
         from accelerate import infer_auto_device_map
 
         with init_empty_weights():
-            model_canvas = AutoModelForCausalLM.from_config(model_config)
+            model_canvas = AutoModelForCausalLM.from_config(
+                model_config, trust_remote_code=cfg.trust_remote_code or False
+            )
         model_canvas.tie_weights()
         device_map = infer_auto_device_map(
             model_canvas,
@@ -502,6 +504,9 @@ def load_model(
         model_kwargs["attn_implementation"] = "eager"
         model_config._attn_implementation = "eager"  # pylint: disable=protected-access
 
+    if cfg.low_cpu_mem_usage:
+        model_kwargs["low_cpu_mem_usage"] = True
+
     qlora_fsdp = cfg.fsdp and cfg.adapter == "qlora"
 
     try:

src/axolotl/utils/trainer.py

@@ -312,6 +312,8 @@ def setup_fsdp_envs(cfg):
         os.environ["FSDP_USE_ORIG_PARAMS"] = "true"
     if cfg.fsdp_config.fsdp_state_dict_type:
         os.environ["FSDP_STATE_DICT_TYPE"] = cfg.fsdp_config.fsdp_state_dict_type
+    if cfg.fsdp_config.fsdp_auto_wrap_policy:
+        os.environ["FSDP_AUTO_WRAP_POLICY"] = cfg.fsdp_config.fsdp_auto_wrap_policy
     if cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap:
         os.environ[
             "FSDP_TRANSFORMER_CLS_TO_WRAP"