Enabling 8-bit quantization broke the model. Attempting to fix by removing 'pad_token_id' arg from AutoModelForCausalLM.from_pretrained

Gemma.py

@@ -12,7 +12,7 @@ class GemmaLLM:
         login(token=os.environ.get("GEMMA_TOKEN"))
         
         model_id = "google/gemma-3-4b-it"
-        model = AutoModelForCausalLM.from_pretrained(model_id, load_in_8bit=True, pad_token_id=0)
+        model = AutoModelForCausalLM.from_pretrained(model_id, load_in_8bit=True)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
         
         self.model = pipeline("text-generation", model=model, tokenizer=tokenizer, torch_dtype=torch.bfloat16)