allow bf16 flag but warn (#3563) [skip ci]

* allow bf16 flag but warn

Reason: when doing e.g. LoRA merges with CUDA_VISIBLE_DEVICES=, this will unnecessarily crash, even though the LoRA merge operation would have finished successfully. This seems to warrant changing it to a warning instead, as the code will most likely crash later if bf16 is unavailable and training begins anyway.

* don't use deprecated LOG.warn

* update tests to reflect validation change

src/axolotl/utils/schemas/config.py

@@ -1352,8 +1352,8 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                 and not self.is_preprocess
                 and (self.bf16 is True or self.bfloat16 is True)
             ):
-                raise ValueError(
+                LOG.warning(
-                    "bf16 requested, but AMP is not supported on this GPU. Requires Ampere series or above."
+                    "bf16 requested, but AMP is not supported on this GPU. Requires Ampere series or above. Training will fail, but other operations (such as merging) are still functional."
                 )
         return self
 

tests/patched/test_validation.py

@@ -726,8 +726,12 @@ class TestValidation(BaseValidation):
             | minimal_cfg
         )
 
-        with pytest.raises(ValueError, match=r".*AMP is not supported on this GPU*"):
+        with self._caplog.at_level("WARNING"):
             AxolotlConfigWCapabilities(**cfg.to_dict())
+            assert any(
+                "AMP is not supported" in record.message
+                for record in self._caplog.records
+            )
 
         cfg = (
             DictDefault(