Add CPU model smoke diagnostics

app.py

@@ -286,9 +286,68 @@ def _run_bench_locked(model_filter=None):
     return results
 
 
-# ---------------------------------------------------------------------------
-# Audio generation
-# ---------------------------------------------------------------------------
+def _run_cpu_smoke(model_filter):
+    """Run a short CPU-only loader/inference diagnostic for one model."""
+    import subprocess
+    from huggingface_hub import hf_hub_download
+
+    model = next((m for m in _BENCH_MODELS if m["name"] == model_filter), None)
+    if model is None:
+        raise ValueError(f"Unknown model: {model_filter}")
+    cli = _ensure_llama()
+    env = os.environ.copy()
+    env["LD_LIBRARY_PATH"] = os.path.dirname(cli)
+    env["CUDA_VISIBLE_DEVICES"] = ""
+    env["GGML_CUDA"] = "0"
+    result = {
+        "name": model["name"],
+        "engine": "llama-cli",
+        "gpu_layers": 0,
+        "threads": _parse_cpu_count() or 8,
+        "phases": {},
+    }
+
+    t0 = time.time()
+    version = subprocess.run([cli, "--version"], capture_output=True, text=True, timeout=15, env=env)
+    result["phases"]["binary"] = {
+        "ok": version.returncode == 0,
+        "elapsed_sec": round(time.time() - t0, 2),
+        "returncode": version.returncode,
+        "output": (version.stdout + version.stderr)[-1000:],
+    }
+    if version.returncode != 0:
+        return result
+
+    t0 = time.time()
+    try:
+        path = hf_hub_download(repo_id=model["repo"], filename=model["file"], cache_dir=_BENCH_CACHE)
+        result["model_path"] = path
+        result["phases"]["download"] = {"ok": True, "elapsed_sec": round(time.time() - t0, 2)}
+    except Exception as exc:
+        result["phases"]["download"] = {"ok": False, "elapsed_sec": round(time.time() - t0, 2), "error": str(exc)}
+        return result
+
+    command = [cli, "-m", path, "-t", str(result["threads"]), "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.", "--no-display-prompt"]
+    t0 = time.time()
+    try:
+        run = subprocess.run(command, capture_output=True, text=True, timeout=45, env=env)
+        result["phases"]["load_and_one_token"] = {
+            "ok": run.returncode == 0,
+            "elapsed_sec": round(time.time() - t0, 2),
+            "returncode": run.returncode,
+            "stdout": (run.stdout or "")[-1000:],
+            "stderr": (run.stderr or "")[-3000:],
+        }
+    except subprocess.TimeoutExpired as exc:
+        result["phases"]["load_and_one_token"] = {
+            "ok": False,
+            "elapsed_sec": round(time.time() - t0, 2),
+            "timeout": True,
+            "stdout": str(exc.stdout or "")[-1000:],
+            "stderr": str(exc.stderr or "")[-3000:],
+        }
+    return result
+
 
 MODEL_CACHE = {}
 
@@ -351,6 +410,14 @@ def _patched_create_app(blocks, **kwargs):
         try: return JSONResponse(_search_web(q, backend=backend, max_results=max_results, extract_top=extract_top, region=region))
         except Exception as e: return JSONResponse({"error": str(e)}, status_code=502)
 
+    @fa_app.get("/respite/benchmark/smoke")
+    def _smoke(model: str="Gemma 4 E4B"):
+        try:
+            with _BENCH_LOCK:
+                return JSONResponse(_run_cpu_smoke(model.strip()))
+        except Exception as e:
+            return JSONResponse({"error": str(e)}, status_code=500)
+
     @fa_app.get("/respite/benchmark")
     def _b(force: bool=False, model: str=""):
         try: