Raise explicit error when FP8 is requested via from_config

src/transformers/modeling_utils.py

@@ -1312,6 +1312,12 @@ def __init__(self, config: PreTrainedConfig, *inputs, **kwargs):
                 f"`model = {self.__class__.__name__}.from_pretrained(PRETRAINED_MODEL_NAME)`"
             )
         self.config = config
+        quant_config = getattr(config, "quantization_config", None)
+        if quant_config is not None:
+            raise NotImplementedError(
+                "Quantization via `from_config()` is not supported. "
+                "Quantized models must be created via `from_pretrained()` with an appropriate backend."
+            )
 
         # Check the attention implementation is supported, or set it if not yet set (on the internal attr, to avoid
         # setting it recursively)

tests/quantization/config/test_from_config.py

@@ -0,0 +1,14 @@
+import pytest
+
+from transformers import AutoConfig, AutoModel
+
+
+def test_quantization_from_config_raises():
+    config = AutoConfig.from_pretrained("gpt2")
+    config.quantization_config = {"quant_method": "fp8"}
+
+    with pytest.raises(
+        NotImplementedError,
+        match="Quantization via",
+    ):
+        AutoModel.from_config(config)