Merge pull request #130 from chichun-charlie-liu/fix4PR102

fix: disable granite in custom gptq as gptqmodel already supports it, fix …

Author: chichun-charlie-liu

fms_mo/run_quant.py

@@ -144,6 +144,12 @@ def run_gptq(model_args, data_args, opt_args, gptq_args):
 
     # Add custom model_type mapping to gptqmodel LUT so GPTQModel can recognize them.
     for mtype, cls in custom_gptq_classes.items():
+        if mtype in MODEL_MAP:
+            logger.info(
+                f"Custom GPTQ model type {mtype} already exists in default MODEL_MAP.\n"
+                "The mapping for this type is overwritten by user defined type now."
+                "Please make sure this is the behavior you'd like to have."
+            )
         SUPPORTED_MODELS.append(mtype)
         MODEL_MAP[mtype] = cls
 

fms_mo/utils/custom_gptq_models.py

@@ -49,6 +49,6 @@ class GraniteMoeGPTQForCausalLM(BaseGPTQModel):
 #       config.json). Make sure you cover the ones in the model family you want to use, as they may
 #       not be under the same model_type. See Granite as an example.
 custom_gptq_classes = {
-    "granite": GraniteGPTQForCausalLM,
+    # "granite": GraniteGPTQForCausalLM,
     "granitemoe": GraniteMoeGPTQForCausalLM,
 }