Update info.sh

info.sh

@@ -20,17 +20,31 @@ fi
 echo
 echo "[PyTorch / CUDA backend]"
 python - <<'PY'
-import json, os, torch
+import json, os, torch, inspect
+def to_bool(x):
+    try:
+        if callable(x):
+            try:
+                sig = inspect.signature(x)
+                if len(sig.parameters)==0:
+                    return bool(x())
+            except Exception:
+                pass
+            return True
+        return bool(x)
+    except Exception:
+        return None
+
 info = {
   "torch": getattr(torch, "__version__", None),
   "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.backends.cudnn.is_available() else None,
-  "tf32": torch.backends.cuda.matmul.allow_tf32 if torch.cuda.is_available() else None,
-  "flash_sdp": torch.backends.cuda.enable_flash_sdp if torch.cuda.is_available() else None,
-  "mem_efficient_sdp": torch.backends.cuda.enable_mem_efficient_sdp if torch.cuda.is_available() else None,
-  "math_sdp": torch.backends.cuda.enable_math_sdp if torch.cuda.is_available() else None,
+  "tf32": (torch.backends.cuda.matmul.allow_tf32 if torch.cuda.is_available() else None),
+  "flash_sdp": (to_bool(getattr(torch.backends.cuda, "enable_flash_sdp", None)) if torch.cuda.is_available() else None),
+  "mem_efficient_sdp": (to_bool(getattr(torch.backends.cuda, "enable_mem_efficient_sdp", None)) if torch.cuda.is_available() else None),
+  "math_sdp": (to_bool(getattr(torch.backends.cuda, "enable_math_sdp", None)) if torch.cuda.is_available() else None),
 }
 print(json.dumps(info, indent=2))
 for i in range(min(torch.cuda.device_count(), 8)):
@@ -51,9 +65,8 @@ PY
 echo
 echo "[FlashAttention]"
 python - <<'PY'
-import importlib, sys
-mods = ["flash_attn", "flash_attn_2_cuda"]
-for m in mods:
+import importlib
+for m in ("flash_attn","flash_attn_2_cuda"):
   try:
     importlib.import_module(m); print(f"{m}: OK")
   except Exception as e:
@@ -95,11 +108,8 @@ names_to_try = [
 ]
 ok = ok_import(names_to_try)
 if not ok:
-    print("  Hint: faltam kernels de layer_norm/RMSNorm do FlashAttention.")
-    print("  Aceleração ficará reduzida; para instalar:")
-    print("   - Rodar builder para compilar e instalar flash_attn e salvar wheel para reuso;")
-    print("   - Ou instalar manualmente a tag compatível: Dao-AILab/flash-attention (csrc/layer_norm).")
-    print("  Doc: https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm")
+    print("  Hint: faltam kernels LN/RMSNorm do FlashAttention (performance reduzida).")
+    print("  Use builder.sh para compilar flash_attn e reutilizar a wheel.")
 PY
 
 echo
@@ -110,7 +120,7 @@ try:
   print("triton:", triton.__version__)
   try:
     import triton.ops as _; print("triton.ops: OK")
-  except Exception as e:
+  except Exception:
     print("triton.ops: not present (ok on Triton>=3.x)")
 except Exception as e:
   print("triton: FAIL ->", e)
@@ -149,7 +159,7 @@ env | grep -E '^(CUDA_VISIBLE_DEVICES|NCCL_|TORCH_|ENABLE_.*SDP|HF_HUB_.*|CUDA_|
 
 echo
 echo "[Caminhos e permissões de saída]"
-OUT="/app/output"
+OUT="/app/outputs"
 echo "OUT dir: $OUT"
 mkdir -p "$OUT"
 ls -la "$OUT" || true