preserve quantization config

app.py

@@ -44,9 +44,20 @@ def load_model():
             )
         else:
             logger.info("CUDA not available, loading with CPU optimizations")
-            if getattr(config, "quantization_config", None) is not None:
-                logger.info("Disabling quantization settings for CPU execution")
-                config.quantization_config = None
+            quant_config = getattr(config, "quantization_config", None)
+            if quant_config is not None:
+                logger.info("Adjusting quantization settings for CPU execution")
+                if isinstance(quant_config, dict):
+                    quant_config.pop("load_in_4bit", None)
+                    quant_config.pop("load_in_8bit", None)
+                else:
+                    if hasattr(quant_config, "load_in_4bit"):
+                        quant_config.load_in_4bit = False
+                    if hasattr(quant_config, "load_in_8bit"):
+                        quant_config.load_in_8bit = False
+                    if hasattr(quant_config, "llm_int8_enable_fp32_cpu_offload"):
+                        quant_config.llm_int8_enable_fp32_cpu_offload = False
+                config.quantization_config = quant_config
             model = AutoModelForCausalLM.from_pretrained(
                 MODEL_NAME,
                 config=config,
@@ -264,7 +275,7 @@ def create_interface():
                 num_beams = gr.Slider(
                     minimum=1,
                     maximum=8,
-                    value=3,
+                    value=1,
                     step=1,
                     label="Number of Beams",
                     info="Higher values = better quality but slower generation (beam search)",