8bit and deepspeed changes

ds_config.json

@@ -20,10 +20,12 @@
     }
   },
   "scheduler": {
-    "type": "OneCycle",
+    "type": "WarmupDecayLR",
     "params": {
-      "cycle_min_lr": 1e-7,
-      "cycle_max_lr": 1e-4
+      "warmup_min_lr": "auto",
+      "warmup_max_lr": "auto",
+      "warmup_num_steps": "auto",
+      "total_num_steps": "auto"
     }
   },
   "zero_optimization": {

src/axolotl/utils/models.py

@@ -101,19 +101,12 @@ def load_model(
             )
             load_in_8bit = False
         elif is_llama_derived_model and "LlamaForCausalLM" in globals():
-            if not cfg.load_in_8bit:
-                model = LlamaForCausalLM.from_pretrained(
-                    base_model,
-                    device_map=cfg.device_map,
-                )
-            else:
-                model = LlamaForCausalLM.from_pretrained(
-                    base_model,
-                    load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-                    torch_dtype=torch_dtype,
-                    device_map=cfg.device_map,
-                )
-
+            model = LlamaForCausalLM.from_pretrained(
+                base_model,
+                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
+                torch_dtype=torch_dtype,
+                device_map=cfg.device_map,
+            )
         elif model_type:
             model = getattr(transformers, model_type).from_pretrained(
                 base_model,