Update expert mode to use diffusiongemma model and include read_b64 helper

agents/nvidia_llm.py

@@ -5,6 +5,7 @@ import queue as _queue_module
 import time
 import requests
 import keyring
+import base64
 from flask import Flask, request, jsonify
 
 # ── Logging ───────────────────────────────────────────────────────────────────
@@ -29,7 +30,7 @@ _QUEUE_MAX_SIZE    = int(os.getenv("NVIDIA_QUEUE_MAX", "8"))
 _REQUEST_TIMEOUT_S = int(os.getenv("NVIDIA_LLM_TIMEOUT", "600"))
 
 # DEFAULT_MODEL = "google/diffusiongemma-26b-a4b-it"
-DEFAULT_MODEL = "mistralai/mistral-medium-3.5-128b"
+DEFAULT_MODEL = "google/diffusiongemma-26b-a4b-it"
 
 # ── Flask app ─────────────────────────────────────────────────────────────────
 app = Flask(__name__)
@@ -47,6 +48,10 @@ _inference_queue: _queue_module.Queue = _queue_module.Queue(maxsize=_QUEUE_MAX_S
 if not NVIDIA_API_KEY:
     log.warning("NVIDIA_API_KEY is not set. API calls might fail if the token is required.")
 
+def read_b64(path):
+    with open(path, "rb") as f:
+        return base64.b64encode(f.read()).decode()
+
 def _run_inference(data: dict) -> dict:
     raw_prompt = data.get("prompt", "")
     if not raw_prompt:
@@ -64,11 +69,11 @@ def _run_inference(data: dict) -> dict:
         reasoning_effort = None
     else:
         model_name = data.get("model", DEFAULT_MODEL)
-        max_tokens = int(data.get("max_tokens", 16384))
-        chat_template_kwargs = None
-        temperature = float(data.get("temperature", 0.70))
-        top_p = float(data.get("top_p", 1.00))
-        reasoning_effort = "high"
+        max_tokens = int(data.get("max_tokens", 4096))
+        chat_template_kwargs = {"enable_thinking": True}
+        temperature = float(data.get("temperature", 1.00))
+        top_p = float(data.get("top_p", 0.95))
+        reasoning_effort = None
 
     invoke_url = "https://integrate.api.nvidia.com/v1/chat/completions"
     headers = {