PoC: add chat template heuristics (#1283)

* PoC: add chat template heuristics

The fallback chat template adapter of Vicuna is not ideal in some cases (e.g. a test against a sub-portion of the BBC news classification task on Kaggle gave an 82% accuracy with Vicuna and 88% with the official ChatML format for a q4_k_m Qwen 2.5 3B-Instruct gguf).

This PR adds a proof of concept simple heuristic which looks at the chat template and upgrades the adapter when it is able to.

* gemma 2 heuristic

* Phi 4, Llama 3.x heuristics

* better qwen vs generic heuristic

* cleanup

* mistral (generic) heuristic

* fix sys msg for mistral

* phi 3.5

* mistral v3

* cohere (aya expanse 32b based)

* only derive from chat template if AutoGuess

* add notes about alpaca fallbacks

* added AutoGuess.json dummy

* add mistral v7

* switch to using a json list with search strings

kcpp_adapters/AutoGuess.json

@@ -0,0 +1,113 @@
+[
+{
+    "search": ["<|im_start|>assistant", "<|im_end|>", "<|im_sep|>"],
+    "name": "ChatML (Phi 4)",
+    "adapter": {
+        "system_start": "<|im_start|>system<|im_sep|>",
+        "system_end": "<|im_end|>",
+        "user_start": "<|im_start|>user<|im_sep|>",
+        "user_end": "<|im_end|>",
+        "assistant_start": "<|im_start|>assistant<|im_sep|>",
+        "assistant_end": "<|im_end|>"
+    }
+}, {
+    "search": ["<|im_start|>assistant", "<|im_end|>", "You are provided with function signatures within <tools>"],
+    "name": "ChatML (Qwen 2.5 based).",
+    "adapter": {
+        "system_start": "<|im_start|>system\n\n",
+        "system_end": "<|im_end|>\n\n",
+        "user_start": "<|im_start|>user\n\n",
+        "user_end": "<|im_end|>\n\n",
+        "assistant_start": "<|im_start|>assistant\n\n",
+        "assistant_end": "<|im_end|>\n\n",
+        "tools_start": "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n\n<tools>\n",
+        "tools_end": "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n"
+    }
+}, {
+    "search": ["<|im_start|>assistant", "<|im_end|>"],
+    "name": "ChatML (Generic).",
+    "adapter": {
+        "system_start": "<|im_start|>system\n\n",
+        "system_end": "<|im_end|>\n\n",
+        "user_start": "<|im_start|>user\n\n",
+        "user_end": "<|im_end|>\n\n",
+        "assistant_start": "<|im_start|>assistant\n\n",
+        "assistant_end": "<|im_end|>\n\n"
+    }
+}, {
+    "search": ["System role not supported", "<start_of_turn>"],
+    "name": "Google Gemma 2.",
+    "adapter": {
+        "user_start": "<start_of_turn>user\n",
+        "user_end": "<end_of_turn>\n",
+        "assistant_start": "<start_of_turn>model\n",
+        "assistant_end": "<end_of_turn>\n"
+    }
+}, {
+    "search": ["<|start_header_id|>system"],
+    "name": "Llama 3.x.",
+    "adapter": {
+        "system_start": "<|start_header_id|>system<|end_header_id|>\n\n",
+        "system_end": "<|eot_id|>\n\n",
+        "user_start": "<|start_header_id|>user<|end_header_id|>\n\n",
+        "user_end": "<|eot_id|>\n\n",
+        "assistant_start": "<|start_header_id|>assistant<|end_header_id|>\n\n",
+        "assistant_end": "<|eot_id|>\n\n"
+    }
+}, {
+    "search": ["[/INST]", "[SYSTEM_PROMPT]"],
+    "name": "Mistral V7 (with system prompt)",
+    "adapter": {
+        "system_start": "[SYSTEM_PROMPT] ",
+        "system_end": "[/SYSTEM_PROMPT]",
+        "user_start": "[INST] ",
+        "user_end": "[/INST]",
+        "assistant_start": " ",
+        "assistant_end": "</s>"
+    }
+}, {
+    "search": ["[/INST]", "\"[INST] \" + system_message"],
+    "name": "Mistral V3",
+    "adapter": {
+        "system_start": "[INST] ",
+        "system_end": "[/INST] ",
+        "user_start": "[INST] ",
+        "user_end": "[/INST] ",
+        "assistant_start": "",
+        "assistant_end": "</s>"
+    }
+}, {
+    "search": ["[/INST]"],
+    "name": "Mistral (Generic)",
+    "adapter": {
+        "system_start": "[INST]",
+        "system_end": "[/INST]\n",
+        "user_start": "[INST]",
+        "user_end": "[/INST]\n",
+        "assistant_start": "",
+        "assistant_end": "</s>"
+    }
+}, {
+    "search": ["<|system|>", "<|user|>"],
+    "name": "Phi 3.5",
+    "adapter": {
+        "system_start": "<|system|>\n",
+        "system_end": "<|end|>\n",
+        "user_start": "<|user|>\n",
+        "user_end": "<|end|>\n",
+        "assistant_start": "<|assistant|>\n",
+        "assistant_end": "<|end|>\n"
+    }
+}, {
+    "search": ["<|START_OF_TURN_TOKEN|>"],
+    "name": "Cohere (Aya Expanse 32B based)",
+    "adapter": {
+        "system_start": "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>",
+        "system_end": "<|END_OF_TURN_TOKEN|>",
+        "user_start": "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>",
+        "user_end": "<|END_OF_TURN_TOKEN|>",
+        "assistant_start": "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
+        "assistant_end": "<|END_OF_TURN_TOKEN|>"
+    }
+}
+]

koboldcpp.py

@@ -4648,6 +4648,29 @@ def main(launch_args,start_server=True):
             exitcounter = 999
             exit_with_error(3,"Could not load text model: " + modelname)
 
+    if (
+        chatcompl_adapter is not None
+        and isinstance(chatcompl_adapter, list)
+    ):
+        # The chat completions adapter is a list that needs derivation from chat templates
+        # Try to derive chat completions adapter from chat template, now that we have the model loaded
+        ctbytes = handle.get_chat_template()
+        chat_template = ctypes.string_at(ctbytes).decode("UTF-8","ignore")
+        candidates = chatcompl_adapter
+        chatcompl_adapter = None
+        if chat_template != "":
+            for entry in candidates:
+                if all(s in chat_template for s in entry['search']):
+                    print(f"Chat completion heuristic: {entry['name']}")
+                    chatcompl_adapter = entry['adapter']
+                    break
+        if chatcompl_adapter is None:
+            print("Chat template heuristics failed to identify chat completions format. Alpaca will be used.")
+
+    if chatcompl_adapter is None and not args.chatcompletionsadapter:
+        print("Note: Alpaca format will be used for OpenAI Compatible API chat completions. Use --chatcompletionsadapter=AutoGuess to use chat template heuristics.")
+
+
     #handle loading image model
     if args.sdmodel and args.sdmodel!="":
         imgmodel = args.sdmodel