convert exponential notation lr to floats (#771)

src/axolotl/utils/config.py

@@ -119,6 +119,9 @@ def normalize_config(cfg):
         or (cfg.model_type and "mistral" in cfg.model_type.lower())
     )
 
+    if isinstance(cfg.learning_rate, str):
+        cfg.learning_rate = float(cfg.learning_rate)
+
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
 

tests/test_normalize_config.py

@@ -0,0 +1,39 @@
+"""
+Test classes for checking functionality of the cfg normalization
+"""
+import unittest
+
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+
+class NormalizeConfigTestCase(unittest.TestCase):
+    """
+    test class for normalize_config checks
+    """
+
+    def _get_base_cfg(self):
+        return DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "num_epochs": 1,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+            }
+        )
+
+    def test_lr_as_float(self):
+        cfg = (
+            self._get_base_cfg()
+            | DictDefault(  # pylint: disable=unsupported-binary-operation
+                {
+                    "learning_rate": "5e-5",
+                }
+            )
+        )
+
+        normalize_config(cfg)
+
+        assert cfg.learning_rate == 0.00005