Several small fixes

- Stop llama-server subprocess on model unload instead of relying on GC
- Fix tool_calls[].index being string instead of int in API responses
- Omit tool_calls key from API response when empty per OpenAI spec
- Prevent division by zero when micro_batch_size > batch_size in training
- Copy sampler_priority list before mutating in ExLlamaV3
- Normalize presence/frequency_penalty names for ExLlamaV3 sampler sorting
- Restore original chat_template after training instead of leaving it mutated

modules/models.py

@@ -126,6 +126,8 @@ def unload_model(keep_model_name=False):
 
     if model_class_name in ['Exllamav3Model', 'Exllamav3HF', 'TensorRTLLMModel']:
         shared.model.unload()
+    elif model_class_name == 'LlamaServer':
+        shared.model.stop()
 
     shared.model = shared.tokenizer = None
     shared.lora_names = []