Use offline chat prompt config to set context window of loaded chat model

Previously you couldn't configure the n_ctx of the loaded offline chat
model. This made it hard to use good offline chat model (which these
days also have larger context) on machines with lower VRAM

pyproject.toml

@@ -78,6 +78,7 @@ dependencies = [
     "phonenumbers == 8.13.27",
     "markdownify ~= 0.11.6",
     "websockets == 12.0",
+    "psutil >= 5.8.0",
 ]
 dynamic = ["version"]
 
@@ -105,7 +106,6 @@ dev = [
     "pytest-asyncio == 0.21.1",
     "freezegun >= 1.2.0",
     "factory-boy >= 3.2.1",
-    "psutil >= 5.8.0",
     "mypy >= 1.0.1",
     "black >= 23.1.0",
     "pre-commit >= 3.0.4",

src/khoj/processor/conversation/offline/chat_model.py

@@ -30,6 +30,7 @@ def extract_questions_offline(
     use_history: bool = True,
     should_extract_questions: bool = True,
     location_data: LocationData = None,
+    max_prompt_size: int = None,
 ) -> List[str]:
     """
     Infer search queries to retrieve relevant notes to answer user query
@@ -41,7 +42,7 @@ def extract_questions_offline(
         return all_questions
 
     assert loaded_model is None or isinstance(loaded_model, Llama), "loaded_model must be of type Llama, if configured"
-    offline_chat_model = loaded_model or download_model(model)
+    offline_chat_model = loaded_model or download_model(model, max_tokens=max_prompt_size)
 
     location = f"{location_data.city}, {location_data.region}, {location_data.country}" if location_data else "Unknown"
 
@@ -67,12 +68,14 @@ def extract_questions_offline(
         location=location,
     )
     messages = generate_chatml_messages_with_context(
-        example_questions, model_name=model, loaded_model=offline_chat_model
+        example_questions, model_name=model, loaded_model=offline_chat_model, max_prompt_size=max_prompt_size
     )
 
     state.chat_lock.acquire()
     try:
-        response = send_message_to_model_offline(messages, loaded_model=offline_chat_model)
+        response = send_message_to_model_offline(
+            messages, loaded_model=offline_chat_model, max_prompt_size=max_prompt_size
+        )
     finally:
         state.chat_lock.release()
 
@@ -138,7 +141,7 @@ def converse_offline(
     """
     # Initialize Variables
     assert loaded_model is None or isinstance(loaded_model, Llama), "loaded_model must be of type Llama, if configured"
-    offline_chat_model = loaded_model or download_model(model)
+    offline_chat_model = loaded_model or download_model(model, max_tokens=max_prompt_size)
     compiled_references_message = "\n\n".join({f"{item}" for item in references})
 
     current_date = datetime.now().strftime("%Y-%m-%d")
@@ -190,18 +193,18 @@ def converse_offline(
     )
 
     g = ThreadedGenerator(references, online_results, completion_func=completion_func)
-    t = Thread(target=llm_thread, args=(g, messages, offline_chat_model))
+    t = Thread(target=llm_thread, args=(g, messages, offline_chat_model, max_prompt_size))
     t.start()
     return g
 
 
-def llm_thread(g, messages: List[ChatMessage], model: Any):
+def llm_thread(g, messages: List[ChatMessage], model: Any, max_prompt_size: int = None):
     stop_phrases = ["<s>", "INST]", "Notes:"]
 
     state.chat_lock.acquire()
     try:
         response_iterator = send_message_to_model_offline(
-            messages, loaded_model=model, stop=stop_phrases, streaming=True
+            messages, loaded_model=model, stop=stop_phrases, max_prompt_size=max_prompt_size, streaming=True
         )
         for response in response_iterator:
             g.send(response["choices"][0]["delta"].get("content", ""))
@@ -216,9 +219,10 @@ def send_message_to_model_offline(
     model="NousResearch/Hermes-2-Pro-Mistral-7B-GGUF",
     streaming=False,
     stop=[],
+    max_prompt_size: int = None,
 ):
     assert loaded_model is None or isinstance(loaded_model, Llama), "loaded_model must be of type Llama, if configured"
-    offline_chat_model = loaded_model or download_model(model)
+    offline_chat_model = loaded_model or download_model(model, max_tokens=max_prompt_size)
     messages_dict = [{"role": message.role, "content": message.content} for message in messages]
     response = offline_chat_model.create_chat_completion(messages_dict, stop=stop, stream=streaming)
     if streaming:

src/khoj/processor/conversation/offline/utils.py

@@ -1,18 +1,19 @@
 import glob
 import logging
+import math
 import os
 
 from huggingface_hub.constants import HF_HUB_CACHE
 
 from khoj.utils import state
+from khoj.utils.helpers import get_device_memory
 
 logger = logging.getLogger(__name__)
 
 
-def download_model(repo_id: str, filename: str = "*Q4_K_M.gguf"):
-    from llama_cpp.llama import Llama
-
-    # Initialize Model Parameters. Use n_ctx=0 to get context size from the model
+def download_model(repo_id: str, filename: str = "*Q4_K_M.gguf", max_tokens: int = None):
+    # Initialize Model Parameters
+    # Use n_ctx=0 to get context size from the model
     kwargs = {"n_threads": 4, "n_ctx": 0, "verbose": False}
 
     # Decide whether to load model to GPU or CPU
@@ -23,23 +24,33 @@ def download_model(repo_id: str, filename: str = "*Q4_K_M.gguf"):
     model_path = load_model_from_cache(repo_id, filename)
     chat_model = None
     try:
-        if model_path:
-            chat_model = Llama(model_path, **kwargs)
-        else:
-            Llama.from_pretrained(repo_id=repo_id, filename=filename, **kwargs)
+        chat_model = load_model(model_path, repo_id, filename, kwargs)
     except:
         # Load model on CPU if GPU is not available
         kwargs["n_gpu_layers"], device = 0, "cpu"
-        if model_path:
-            chat_model = Llama(model_path, **kwargs)
-        else:
-            chat_model = Llama.from_pretrained(repo_id=repo_id, filename=filename, **kwargs)
+        chat_model = load_model(model_path, repo_id, filename, kwargs)
 
-    logger.debug(f"{'Loaded' if model_path else 'Downloaded'} chat model to {device.upper()}")
+    # Now load the model with context size set based on:
+    # 1. context size supported by model and
+    # 2. configured size or machine (V)RAM
+    kwargs["n_ctx"] = infer_max_tokens(chat_model.n_ctx(), max_tokens)
+    chat_model = load_model(model_path, repo_id, filename, kwargs)
 
+    logger.debug(
+        f"{'Loaded' if model_path else 'Downloaded'} chat model to {device.upper()} with {kwargs['n_ctx']} token context window."
+    )
     return chat_model
 
 
+def load_model(model_path: str, repo_id: str, filename: str = "*Q4_K_M.gguf", kwargs: dict = {}):
+    from llama_cpp.llama import Llama
+
+    if model_path:
+        return Llama(model_path, **kwargs)
+    else:
+        return Llama.from_pretrained(repo_id=repo_id, filename=filename, **kwargs)
+
+
 def load_model_from_cache(repo_id: str, filename: str, repo_type="models"):
     # Construct the path to the model file in the cache directory
     repo_org, repo_name = repo_id.split("/")
@@ -52,3 +63,12 @@ def load_model_from_cache(repo_id: str, filename: str, repo_type="models"):
         return paths[0]
     else:
         return None
+
+
+def infer_max_tokens(model_context_window: int, configured_max_tokens=math.inf) -> int:
+    """Infer max prompt size based on device memory and max context window supported by the model"""
+    vram_based_n_ctx = int(get_device_memory() / 1e6)  # based on heuristic
+    if configured_max_tokens:
+        return min(configured_max_tokens, model_context_window)
+    else:
+        return min(vram_based_n_ctx, model_context_window)

src/khoj/processor/conversation/utils.py

@@ -1,5 +1,6 @@
 import json
 import logging
+import math
 import queue
 from datetime import datetime
 from time import perf_counter
@@ -141,14 +142,12 @@ def generate_chatml_messages_with_context(
     tokenizer_name=None,
 ):
     """Generate messages for ChatGPT with context from previous conversation"""
-    # Set max prompt size from user config, pre-configured for model or to default prompt size
-    try:
-        max_prompt_size = max_prompt_size or model_to_prompt_size[model_name]
-    except:
-        max_prompt_size = 2000
-        logger.warning(
-            f"Fallback to default prompt size: {max_prompt_size}.\nConfigure max_prompt_size for unsupported model: {model_name} in Khoj settings to longer context window."
-        )
+    # Set max prompt size from user config or based on pre-configured for model and machine specs
+    if not max_prompt_size:
+        if loaded_model:
+            max_prompt_size = min(loaded_model.n_ctx(), model_to_prompt_size.get(model_name, math.inf))
+        else:
+            max_prompt_size = model_to_prompt_size.get(model_name, 2000)
 
     # Scale lookback turns proportional to max prompt size supported by model
     lookback_turns = max_prompt_size // 750
@@ -187,7 +186,7 @@ def truncate_messages(
     max_prompt_size,
     model_name: str,
     loaded_model: Optional[Llama] = None,
-    tokenizer_name=None,
+    tokenizer_name="hf-internal-testing/llama-tokenizer",
 ) -> list[ChatMessage]:
     """Truncate messages to fit within max prompt size supported by model"""
 
@@ -197,15 +196,11 @@ def truncate_messages(
         elif model_name.startswith("gpt-"):
             encoder = tiktoken.encoding_for_model(model_name)
         else:
-            try:
-                encoder = download_model(model_name).tokenizer()
-            except:
-                encoder = AutoTokenizer.from_pretrained(tokenizer_name or model_to_tokenizer[model_name])
+            encoder = download_model(model_name).tokenizer()
     except:
-        default_tokenizer = "hf-internal-testing/llama-tokenizer"
-        encoder = AutoTokenizer.from_pretrained(default_tokenizer)
+        encoder = AutoTokenizer.from_pretrained(tokenizer_name)
         logger.warning(
-            f"Fallback to default chat model tokenizer: {default_tokenizer}.\nConfigure tokenizer for unsupported model: {model_name} in Khoj settings to improve context stuffing."
+            f"Fallback to default chat model tokenizer: {tokenizer_name}.\nConfigure tokenizer for unsupported model: {model_name} in Khoj settings to improve context stuffing."
         )
 
     # Extract system message from messages

src/khoj/routers/api.py

@@ -315,8 +315,9 @@ async def extract_references_and_questions(
             using_offline_chat = True
             default_offline_llm = await ConversationAdapters.get_default_offline_llm()
             chat_model = default_offline_llm.chat_model
+            max_tokens = default_offline_llm.max_prompt_size
             if state.offline_chat_processor_config is None:
-                state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model=chat_model)
+                state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model, max_tokens)
 
             loaded_model = state.offline_chat_processor_config.loaded_model
 
@@ -326,6 +327,7 @@ async def extract_references_and_questions(
                 conversation_log=meta_log,
                 should_extract_questions=True,
                 location_data=location_data,
+                max_prompt_size=conversation_config.max_prompt_size,
             )
         elif conversation_config and conversation_config.model_type == ChatModelOptions.ModelType.OPENAI:
             openai_chat_config = await ConversationAdapters.get_openai_chat_config()

src/khoj/routers/helpers.py

@@ -82,9 +82,10 @@ async def is_ready_to_chat(user: KhojUser):
 
     if has_offline_config and user_conversation_config and user_conversation_config.model_type == "offline":
         chat_model = user_conversation_config.chat_model
+        max_tokens = user_conversation_config.max_prompt_size
         if state.offline_chat_processor_config is None:
             logger.info("Loading Offline Chat Model...")
-            state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model=chat_model)
+            state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model, max_tokens)
         return True
 
     ready = has_openai_config or has_offline_config
@@ -385,10 +386,11 @@ async def send_message_to_model_wrapper(
         raise HTTPException(status_code=500, detail="Contact the server administrator to set a default chat model.")
 
     chat_model = conversation_config.chat_model
+    max_tokens = conversation_config.max_prompt_size
 
     if conversation_config.model_type == "offline":
         if state.offline_chat_processor_config is None or state.offline_chat_processor_config.loaded_model is None:
-            state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model)
+            state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model, max_tokens)
 
         loaded_model = state.offline_chat_processor_config.loaded_model
         truncated_messages = generate_chatml_messages_with_context(
@@ -455,7 +457,9 @@ def generate_chat_response(
         conversation_config = ConversationAdapters.get_valid_conversation_config(user, conversation)
         if conversation_config.model_type == "offline":
             if state.offline_chat_processor_config is None or state.offline_chat_processor_config.loaded_model is None:
-                state.offline_chat_processor_config = OfflineChatProcessorModel(conversation_config.chat_model)
+                chat_model = conversation_config.chat_model
+                max_tokens = conversation_config.max_prompt_size
+                state.offline_chat_processor_config = OfflineChatProcessorModel(chat_model, max_tokens)
 
             loaded_model = state.offline_chat_processor_config.loaded_model
             chat_response = converse_offline(

src/khoj/utils/config.py

@@ -69,11 +69,11 @@ class OfflineChatProcessorConfig:
 
 
 class OfflineChatProcessorModel:
-    def __init__(self, chat_model: str = "NousResearch/Hermes-2-Pro-Mistral-7B-GGUF"):
+    def __init__(self, chat_model: str = "NousResearch/Hermes-2-Pro-Mistral-7B-GGUF", max_tokens: int = None):
         self.chat_model = chat_model
         self.loaded_model = None
         try:
-            self.loaded_model = download_model(self.chat_model)
+            self.loaded_model = download_model(self.chat_model, max_tokens=max_tokens)
         except ValueError as e:
             self.loaded_model = None
             logger.error(f"Error while loading offline chat model: {e}", exc_info=True)

src/khoj/utils/helpers.py

@@ -17,6 +17,7 @@ from time import perf_counter
 from typing import TYPE_CHECKING, Optional, Union
 from urllib.parse import urlparse
 
+import psutil
 import torch
 from asgiref.sync import sync_to_async
 from magika import Magika
@@ -271,6 +272,17 @@ def log_telemetry(
     return request_body
 
 
+def get_device_memory() -> int:
+    """Get device memory in GB"""
+    device = get_device()
+    if device.type == "cuda":
+        return torch.cuda.get_device_properties(device).total_memory
+    elif device.type == "mps":
+        return torch.mps.driver_allocated_memory()
+    else:
+        return psutil.virtual_memory().total
+
+
 def get_device() -> torch.device:
     """Get device to run model on"""
     if torch.cuda.is_available():