Add CPU model loader diagnostics

app.py

@@ -360,6 +360,60 @@ def _run_cpu_smoke(model_filter):
     return result
 
 
+def _run_cpu_diagnose(model_filter):
+    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()
+    path = hf_hub_download(repo_id=model["repo"], filename=model["file"], cache_dir=_BENCH_CACHE)
+    env = os.environ.copy()
+    env["LD_LIBRARY_PATH"] = os.path.dirname(cli)
+    env["CUDA_VISIBLE_DEVICES"] = ""
+    env["GGML_CUDA"] = "0"
+
+    def command_output(command, timeout=10):
+        try:
+            p = subprocess.run(command, capture_output=True, text=True, timeout=timeout, env=env)
+            return {"returncode": p.returncode, "output": (p.stdout + p.stderr)[-5000:]}
+        except Exception as exc:
+            return {"error": str(exc)}
+
+    result = {
+        "name": model["name"],
+        "model_path": path,
+        "model_size_bytes": os.path.getsize(path),
+        "model_magic": open(path, "rb").read(4).decode("ascii", "replace"),
+        "cpu_cores": _parse_cpu_count(),
+        "cpu_flags": command_output(["bash", "-lc", "grep -m1 '^flags' /proc/cpuinfo"], 5),
+        "filesystem": command_output(["df", "-T", path], 5),
+        "binary": command_output(["file", cli], 5),
+        "libraries": command_output(["ldd", cli], 5),
+    }
+    trace_path = os.path.join(_BENCH_CACHE, "llama-load.trace")
+    if shutil.which("strace"):
+        try:
+            p = subprocess.run(
+                ["strace", "-f", "-tt", "-o", trace_path, cli, "-m", path,
+                 "-t", "1", "-ngl", "0", "-c", "256", "-n", "1", "-p", "Say hi.",
+                 "--no-display-prompt", "--no-warmup", "-fa", "0"],
+                capture_output=True, text=True, timeout=25, env=env,
+            )
+            result["load_trace"] = {"returncode": p.returncode, "stdout": (p.stdout or "")[-1000:], "stderr": (p.stderr or "")[-2000:]}
+        except subprocess.TimeoutExpired:
+            result["load_trace"] = {"timeout": True}
+        try:
+            with open(trace_path, "r", encoding="utf-8", errors="replace") as f:
+                result["trace_tail"] = f.read()[-12000:]
+        except OSError as exc:
+            result["trace_tail_error"] = str(exc)
+    else:
+        result["load_trace"] = {"available": False}
+    return result
+
+
 def _run_benchmark_job(job_id, kind, model_filter=None, force=False):
     try:
         if kind == "smoke":
@@ -464,6 +518,14 @@ 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/diagnose")
+    def _diagnose(model: str="TinyLlama 1.1B (control)"):
+        try:
+            with _BENCH_LOCK:
+                return JSONResponse(_run_cpu_diagnose(model.strip()))
+        except Exception as e:
+            return JSONResponse({"error": str(e)}, status_code=500)
+
     @fa_app.get("/respite/benchmark/smoke")
     def _smoke(model: str="Gemma 4 E4B", job_id: str=""):
         if job_id: