Skip loading triton.nvidia.cublas if not found

Summary:
We have an old triton internally that doesn't have the cublasLt
bindings

Reviewed By: adamomainz

Differential Revision: D63643619

fbshipit-source-id: 39aece74b52f7747fe2100d7bb905bad49ba1fa0

torchbenchmark/operators/fp8_gemm/persistent.py

@@ -5,13 +5,17 @@
 import triton.language as tl
 import triton.tools.experimental_descriptor
 
+cublas = None
 if torch.cuda.is_available():
-    from triton._C.libtriton import nvidia
+    try:
+        from triton._C.libtriton import nvidia
 
-    cublas_workspace = torch.empty(32 * 1024 * 1024, device="cuda", dtype=torch.uint8)
-    cublas = nvidia.cublas.CublasLt(cublas_workspace)
-else:
-    cublas = None
+        cublas_workspace = torch.empty(
+            32 * 1024 * 1024, device="cuda", dtype=torch.uint8
+        )
+        cublas = nvidia.cublas.CublasLt(cublas_workspace)
+    except (ImportError, IOError, AttributeError):
+        pass
 
 
 def is_cuda():