Model: Check for unsupported cache mode in exllamav2

backends/exllamav2/model.py

@@ -185,9 +185,17 @@ class ExllamaV2Container(BaseModelContainer):
         # MARK: User configuration
 
         # Get cache mode
-        # TODO: Separate validation for Exl2 and Exl3 q-cache options
         self.cache_mode = unwrap(kwargs.get("cache_mode"), "FP16")
 
+        # Catch exllamav3 cache_mode
+        if not self.cache_mode.startswith("Q"):
+            logger.warning(
+                f"Provided cache mode '{self.cache_mode}' is not a "
+                "valid choice for exllamav2, please check your settings. "
+                "Defaulting to FP16."
+            )
+            self.cache_mode = "FP16"
+
         # Turn off GPU split if the user is using 1 GPU
         gpu_count = torch.cuda.device_count()
         gpu_split_auto = unwrap(kwargs.get("gpu_split_auto"), True)
@@ -393,6 +401,15 @@ class ExllamaV2Container(BaseModelContainer):
             # Set draft cache mode
             self.draft_cache_mode = unwrap(draft_args.get("draft_cache_mode"), "FP16")
 
+            # Catch exllamav3 draft_cache_mode
+            if not self.draft_cache_mode.startswith("Q"):
+                logger.warning(
+                    f"Provided draft cache mode '{self.draft_cache_mode}' is not a "
+                    "valid choice for exllamav2, please check your settings. "
+                    "Defaulting to FP16."
+                )
+                self.draft_cache_mode = "FP16"
+
             # Edit the draft config size
             if chunk_size:
                 self.draft_config.max_input_len = chunk_size

common/config_models.py

@@ -229,7 +229,6 @@ class ModelConfig(BaseConfigModel):
             "or auto-calculate."
         ),
     )
-    # TODO: Separate validation for Exl2 and Exl3 q-cache options
     cache_mode: Optional[CACHE_TYPE] = Field(
         "FP16",
         description=(