Record CUDA benchmark diagnostics

packages/node/meshnet_node/hardware.py

@@ -192,7 +192,7 @@ def detect_hardware() -> dict:
     }
 
 
-def benchmark_throughput(device_str: str = "cpu") -> float:
+def benchmark_throughput_checked(device_str: str = "cpu") -> tuple[float, bool, str | None]:
     """
     Estimate compute throughput via a synthetic transformer GEMM benchmark.
 
@@ -201,7 +201,8 @@ def benchmark_throughput(device_str: str = "cpu") -> float:
     The value is used as benchmark_tokens_per_sec in tracker registration for
     routing tiebreaks; it is not an absolute token rate.
 
-    Falls back to 1.0 if torch is unavailable.
+    Returns (score, ok, error). Score falls back to 1.0 when the requested
+    device cannot run the benchmark.
     """
     try:
         import torch  # type: ignore[import]
@@ -233,6 +234,12 @@ def benchmark_throughput(device_str: str = "cpu") -> float:
         _sync()
         elapsed = time.perf_counter() - t0
 
-        return round(n_iters / max(elapsed, 1e-9), 2)
-    except Exception:
-        return 1.0
+        return round(n_iters / max(elapsed, 1e-9), 2), True, None
+    except Exception as exc:
+        return 1.0, False, f"{type(exc).__name__}: {exc}"
+
+
+def benchmark_throughput(device_str: str = "cpu") -> float:
+    """Return only the numeric throughput index, preserving the legacy API."""
+    score, _ok, _error = benchmark_throughput_checked(device_str)
+    return score

packages/node/meshnet_node/startup.py

@@ -14,7 +14,7 @@ from pathlib import Path
 from typing import Any
 
 from .downloader import compute_shard_checksum, download_shard
-from .hardware import detect_hardware, benchmark_throughput
+from .hardware import detect_hardware, benchmark_throughput_checked
 from .relay_bridge import RelayHttpBridge, peer_id_from_wallet
 from .server import StubNodeServer
 from .torch_server import TorchNodeServer
@@ -386,7 +386,18 @@ def run_startup(
     print(f"  Memory budget: {memory_budget_mb / 1024:.1f} GB {memory_budget_source}", flush=True)
 
     print("Benchmarking compute...", flush=True)
-    bench_tps = benchmark_throughput(device)
+    if device != "cuda" and gpu_name:
+        _cuda_score, cuda_ok, cuda_error = benchmark_throughput_checked("cuda")
+        hw["cuda_benchmark_ok"] = cuda_ok
+        if cuda_error:
+            hw["cuda_benchmark_error"] = cuda_error
+        if not cuda_ok:
+            print(f"  CUDA benchmark unavailable: {cuda_error}; using CPU benchmark", flush=True)
+    bench_tps, bench_ok, bench_error = benchmark_throughput_checked(device)
+    hw["benchmark_device"] = device
+    hw["benchmark_ok"] = bench_ok
+    if bench_error:
+        hw["benchmark_error"] = bench_error
     device_label = "GPU" if device == "cuda" else "CPU"
     print(f"  {device_label} throughput index: {bench_tps:,.0f}", flush=True)