manually set roles to train on

custom_components/llama_conversation/__init__.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 
 import logging
 import os
+import shutil
 from typing import Final
 
 from homeassistant.config_entries import ConfigEntry, ConfigSubentry
@@ -96,9 +97,18 @@ async def _async_update_listener(hass: HomeAssistant, entry: LocalLLMConfigEntry
     await hass.config_entries.async_reload(entry.entry_id)
 
 async def async_unload_entry(hass: HomeAssistant, entry: LocalLLMConfigEntry) -> bool:
-    """Unload Ollama."""
+    """Unload the integration."""
     if not await hass.config_entries.async_unload_platforms(entry, PLATFORMS):
         return False
+
+    if entry.data[CONF_BACKEND_TYPE] == BACKEND_TYPE_LLAMA_CPP:
+        # clean up any disk cache resources
+        def cleanup_cache_dir():
+            cache_dir = entry.data[CONF_CHAT_MODEL].strip().replace(" ", "_").lower()
+            full_path = os.path.join(hass.config.media_dirs.get("local", hass.config.path("media")), "kv_cache", cache_dir)
+            shutil.rmtree(full_path, ignore_errors=True)
+
+        await hass.async_add_executor_job(cleanup_cache_dir)
     hass.data[DOMAIN].pop(entry.entry_id)
     return True
 

custom_components/llama_conversation/backends/llamacpp.py

@@ -180,11 +180,13 @@ class LlamaCppClient(LocalLLMClient):
         _LOGGER.debug("Model loaded")
 
         # create disk cache if enabled
+        # cache must be per-model to avoid conflicts with different hidden state sizes
         cache_size = model_settings.get(CONF_LLAMACPP_CACHE_SIZE_MB, DEFAULT_LLAMACPP_CACHE_SIZE_MB)
+        cache_dir = model_name.strip().replace(" ", "_").lower()
         if cache_size > 0:
             llm.set_cache(LlamaDiskCache(
-                capacity_bytes=int(cache_size * (1024 ** 3)),
-                cache_dir=os.path.join(self.hass.config.media_dirs.get("local", self.hass.config.path("media")), "kv_cache")
+                capacity_bytes=int(cache_size * (1024 ** 2)), # MB to bytes
+                cache_dir=os.path.join(self.hass.config.media_dirs.get("local", self.hass.config.path("media")), "kv_cache", cache_dir)
             ))
 
         if model_settings[CONF_PROMPT_CACHING_ENABLED]:
@@ -406,7 +408,7 @@ class LlamaCppClient(LocalLLMClient):
                     max_tokens=1,
                     grammar=grammar,
                     stream=False,
-                    stop=["<end_of_turn>", "<end_function_call>"]
+                    # stop=["<end_of_turn>", "<end_function_call>"]
                 )
 
                 self.last_cache_prime = time.time()
@@ -480,7 +482,7 @@ class LlamaCppClient(LocalLLMClient):
             grammar=grammar,
             stream=True,
             response_format=response_format,
-            stop=["<end_of_turn>", "<end_function_call>"] # FIXME: make configurable (pull from tool end token?)
+            # stop=["<end_of_turn>", "<end_function_call>"] # FIXME: make configurable (pull from tool end token?)
         )
 
         def next_token() -> Generator[tuple[Optional[str], Optional[List]]]:

data/generate_data.py

@@ -614,18 +614,21 @@ def format_example_sharegpt(example, persona, language, use_system_role, append_
         conversation = [
             { 
                 "role": "system", 
-                "content": [{"type": "text", "text": sys_prompt}]
+                "content": [{"type": "text", "text": sys_prompt}],
+                "train_on_turn": False,
             },
             { 
                 "role": "user", 
-                "content": [{ "type": "text", "text": question }]
+                "content": [{ "type": "text", "text": question }],
+                "train_on_turn": False,
             }
         ]
     else:
         conversation = [
             { 
                 "role": "user", 
-                "content": [{ "type": "text", "text": "\n".join([ sys_prompt, question ]) }]
+                "content": [{ "type": "text", "text": "\n".join([ sys_prompt, question ]) }],
+                "train_on_turn": False,
             }
         ]
     
@@ -646,7 +649,7 @@ def format_example_sharegpt(example, persona, language, use_system_role, append_
             call_names.append(call_name)
             formatted_calls.append({
                 "name": call_name,
-                "arguments": json.dumps(tool_call["tool_args"])
+                "arguments": json.dumps(tool_call["tool_args"]),
             })
 
         if formatted_calls:
@@ -679,12 +682,14 @@ def format_example_sharegpt(example, persona, language, use_system_role, append_
             if tool_response_format == "text":
                 conversation.append({
                     "role": "tool",
-                    "content": [{ "type": "text", "text": json.dumps(result) } for result in step_tool_results]
+                    "content": [{ "type": "text", "text": json.dumps(result) } for result in step_tool_results],
+                    "train_on_turn": False,
                 })
             elif tool_response_format == "functiongemma":
                 conversation.append({
                     "role": "tool",
-                    "content": [{ "name": result["tool_name"], "response": {"result": result["tool_result"]} } for result in step_tool_results]
+                    "content": [{ "name": result["tool_name"], "response": {"result": result["tool_result"]} } for result in step_tool_results],
+                    "train_on_turn": False,
                 })
     
     return { 

train/configs/functiongemma-270m.yml

@@ -308,7 +308,7 @@ datasets:
 - path: /workspace/data/datasets/sample.jsonl
   ds_type: json
   type: chat_template
-  # roles_to_train: [ "assistant" ]
+  roles_to_train: []
   message_field_training: train_on_turn
 
 dataset_prepared_path: /workspace/data/datasets/prepared/