Update info.sh

info.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
-echo "================= RUNTIME CAPABILITIES ================="
+set -euo pipefail
 
-echo "[GPU / Driver / CUDA]"
+echo "================= RUNTIME CAPABILITIES ================="
 nvidia-smi || true
 echo "CUDA_HOME: ${CUDA_HOME:-/usr/local/cuda}"
 echo "NVCC: $(nvcc --version 2>/dev/null | tail -n1 || echo 'N/A')"
@@ -9,19 +9,19 @@ echo
 
 echo "[PyTorch / CUDA backend]"
 python3 - <<'PY'
-import json, sys
+import json
 try:
     import torch
     info = {
-        "torch": torch.__version__,
-        "cuda_available": torch.cuda.is_available(),
-        "cuda_device_count": torch.cuda.device_count(),
-        "cuda_runtime_version": torch.version.cuda if hasattr(torch.version, "cuda") else None,
-        "cudnn_version": (torch.backends.cudnn.version() if torch.cuda.is_available() else None),
-        "tf32": (torch.backends.cuda.matmul.allow_tf32 if torch.cuda.is_available() else None),
-        "flash_sdp": (torch.backends.cuda.flash_sdp_enabled() if hasattr(torch.backends.cuda, "flash_sdp_enabled") else None),
-        "mem_efficient_sdp": (torch.backends.cuda.mem_efficient_sdp_enabled() if hasattr(torch.backends.cuda, "mem_efficient_sdp_enabled") else None),
-        "math_sdp": (torch.backends.cuda.math_sdp_enabled() if hasattr(torch.backends.cuda, "math_sdp_enabled") else None),
+      "torch": torch.__version__,
+      "cuda_available": torch.cuda.is_available(),
+      "cuda_device_count": torch.cuda.device_count(),
+      "cuda_runtime_version": getattr(torch.version, "cuda", None),
+      "cudnn_version": (torch.backends.cudnn.version() if torch.cuda.is_available() else None),
+      "tf32": (torch.backends.cuda.matmul.allow_tf32 if torch.cuda.is_available() else None),
+      "flash_sdp": (torch.backends.cuda.flash_sdp_enabled() if hasattr(torch.backends.cuda,"flash_sdp_enabled") else None),
+      "mem_efficient_sdp": (torch.backends.cuda.mem_efficient_sdp_enabled() if hasattr(torch.backends.cuda,"mem_efficient_sdp_enabled") else None),
+      "math_sdp": (torch.backends.cuda.math_sdp_enabled() if hasattr(torch.backends.cuda,"math_sdp_enabled") else None),
     }
     print(json.dumps(info, indent=2))
     if torch.cuda.is_available():
@@ -65,9 +65,11 @@ try:
     print(f"triton: OK (version={getattr(triton,'__version__','unknown')})")
     try:
         import triton.ops
-        print("triton.ops: OK")
+        print("triton.ops: legacy module present")
+    except ModuleNotFoundError:
+        print("triton.ops: not present (ok on Triton>=3.x)")
     except Exception as e:
-        print(f"triton.ops: ERR {type(e).__name__}: {e}")
+        print(f"triton.ops: WARN {type(e).__name__}: {e}")
 except Exception as e:
     print(f"triton: ERR {type(e).__name__}: {e}")
 PY
@@ -77,15 +79,15 @@ echo "[BitsAndBytes (Q8/Q4)]"
 python3 - <<'PY'
 try:
     import bitsandbytes as bnb
-    import importlib
     v = getattr(bnb, "__version__", "unknown")
     print(f"bitsandbytes: OK (version={v})")
-    # Teste de presença de kernels Triton/Q8
     try:
         import bitsandbytes.triton.int8_matmul_mixed_dequantize as q8
         print("bnb.triton.int8_matmul_mixed_dequantize: OK")
+    except ModuleNotFoundError:
+        print("bnb.q8.triton: not present (disabled or no GPU build)")
     except Exception as e:
-        print(f"bnb.triton.int8_matmul_mixed_dequantize: ERR {type(e).__name__}: {e}")
+        print(f"bnb.q8.triton: WARN {type(e).__name__}: {e}")
 except Exception as e:
     print(f"bitsandbytes: ERR {type(e).__name__}: {e}")
 PY