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

extensions/openai/completions.py

@@ -343,7 +343,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
             if len(tool_call) > 0:
                 for tc in tool_call:
                     tc["id"] = getToolCallId()
-                    tc["index"] = str(len(tool_calls))
+                    tc["index"] = len(tool_calls)
                     tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
                     tool_calls.append(tc)
                 end_last_tool_call = len(answer)
@@ -391,7 +391,7 @@ def chat_completions_common(body: dict, is_legacy: bool = False, stream=False, p
             resp_list: [{
                 "index": 0,
                 "finish_reason": stop_reason,
-                "message": {"role": "assistant", "content": answer, "tool_calls": tool_calls},
+                "message": {"role": "assistant", "content": answer, **({"tool_calls": tool_calls} if tool_calls else {})},
             }],
             "usage": {
                 "prompt_tokens": token_count,

modules/exllamav3.py

@@ -339,11 +339,16 @@ class Exllamav3Model:
 
             # 3. Get the priority list and handle temperature_last
             default_priority = ['repetition_penalty', 'presence_frequency_penalty', 'top_k', 'top_p', 'min_p', 'temperature']
-            sampler_priority = state.get('sampler_priority') or default_priority
+            sampler_priority = list(state.get('sampler_priority') or default_priority)
 
             if state['temperature_last'] and 'temperature' in sampler_priority:
                 sampler_priority.append(sampler_priority.pop(sampler_priority.index('temperature')))
 
+            # The preset system uses separate 'presence_penalty' and
+            # 'frequency_penalty', but ExLlamaV3 has a single combined
+            # SS_PresFreqP sampler. Normalize to the combined name.
+            sampler_priority = ['presence_frequency_penalty' if x in ('presence_penalty', 'frequency_penalty') else x for x in sampler_priority]
+
             # 4. Sort the unordered list based on the priority list
             def custom_sort_key(sampler_obj):
                 class_name = sampler_obj.__class__.__name__

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 = []

modules/training.py

@@ -333,7 +333,8 @@ def do_train(lora_name: str, always_override: bool, all_linear: bool, q_proj_en:
         yield "Cannot input zeroes."
         return
 
-    gradient_accumulation_steps = batch_size // micro_batch_size
+    gradient_accumulation_steps = max(1, batch_size // micro_batch_size)
+    original_chat_template = getattr(shared.tokenizer, 'chat_template', None)
     if shared.tokenizer.pad_token_id is None:
         shared.tokenizer.pad_token_id = shared.tokenizer.eos_token_id
     shared.tokenizer.padding_side = "right"
@@ -820,6 +821,10 @@ def do_train(lora_name: str, always_override: bool, all_linear: bool, q_proj_en:
         logger.info("Training complete, saving")
         lora_model.save_pretrained(lora_file_path)
 
+    # Restore the original chat_template if we changed it for training
+    if shared.tokenizer is not None and hasattr(shared.tokenizer, 'chat_template'):
+        shared.tokenizer.chat_template = original_chat_template
+
     if WANT_INTERRUPT:
         logger.info("Training interrupted.")
         yield f"Interrupted. Incomplete LoRA saved to `{lora_file_path}`."