Add model filter param to benchmark endpoint

app.py

@@ -471,10 +471,14 @@ def _ensure_llama_cpp():
     return llama_cli
 
 
-def _run_benchmark():
-    """Run CPU benchmark on all models using llama.cpp binary."""
+def _run_benchmark(model_filter=None):
+    """Run CPU benchmark on models using llama.cpp binary.
+
+    Args:
+        model_filter: If provided, only run this model name (e.g. 'Gemma 4 E4B').
+    """
     global _BENCHMARK_RESULTS
-    if _BENCHMARK_RESULTS:
+    if _BENCHMARK_RESULTS and not model_filter:
         return _BENCHMARK_RESULTS
 
     import subprocess
@@ -486,7 +490,13 @@ def _run_benchmark():
     results = []
     prompt = "Explain recursion in programming in 3 sentences."
 
-    for model_info in _BENCHMARK_MODELS:
+    models = _BENCHMARK_MODELS
+    if model_filter:
+        models = [m for m in _BENCHMARK_MODELS if m["name"] == model_filter]
+        if not models:
+            raise ValueError(f"Unknown model '{model_filter}'. Available: {[m['name'] for m in _BENCHMARK_MODELS]}")
+
+    for model_info in models:
         result = {"name": model_info["name"], "status": "error"}
         try:
             _log(f"Benchmark: downloading {model_info['name']} (~{model_info['size_gb']} GB)...")
@@ -500,12 +510,15 @@ def _run_benchmark():
             env = os.environ.copy()
             bin_dir = os.path.dirname(llama_cli)
             env["LD_LIBRARY_PATH"] = bin_dir + ":" + env.get("LD_LIBRARY_PATH", "")
+            # 27B model needs more tokens limit and longer timeout
+            n_predict = 128 if model_info["size_gb"] <= 6 else 256
+            timeout_sec = 300 if model_info["size_gb"] <= 6 else 600
             cmd = [
                 llama_cli,
                 "-m", model_path,
                 "-t", str(n_threads),
                 "-c", "2048",
-                "-n", "128",
+                "-n", str(n_predict),
                 "--temp", "0.7",
                 "--top-p", "0.9",
                 "-p", prompt,
@@ -518,7 +531,7 @@ def _run_benchmark():
             t0 = time.time()
             try:
                 proc = subprocess.run(
-                    cmd, capture_output=True, text=True, timeout=300, env=env,
+                    cmd, capture_output=True, text=True, timeout=timeout_sec, env=env,
                 )
             except subprocess.TimeoutExpired:
                 result["error"] = "Timeout after 300s"
@@ -585,7 +598,8 @@ def _run_benchmark():
             except Exception:
                 pass
 
-    _BENCHMARK_RESULTS = results
+    if not model_filter:
+        _BENCHMARK_RESULTS = results
     return results
 
 
@@ -613,13 +627,14 @@ def _patched_create_app(blocks, **kwargs):
             return JSONResponse({"error": str(e)}, status_code=502)
 
     @fa_app.get("/respite/benchmark")
-    def _benchmark(force: bool = False):
-        """Run CPU LLM benchmark. Takes 5-15 minutes. Returns cached results on repeat calls."""
+    def _benchmark(force: bool = False, model: str = ""):
+        """Run CPU LLM benchmark. Pass ?model=Gemma 4 E4B to run a single model."""
         try:
             global _BENCHMARK_RESULTS
             if force:
                 _BENCHMARK_RESULTS = []
-            results = _run_benchmark()
+            model_filter = model.strip() if model.strip() else None
+            results = _run_benchmark(model_filter=model_filter)
             return JSONResponse({
                 "status": "complete",
                 "models": results,