fix test sdpa converter test cpp build issue (#4405)

Author: lanluo-nvidia

core/runtime/execute_engine.cpp

@@ -390,16 +390,6 @@ std::vector<at::Tensor> execute_engine(std::vector<at::Tensor> inputs, c10::intr
           compiled_engine->cudagraph.capture_begin();
           ctx->enqueueV3(recording_stream);
           compiled_engine->cudagraph.capture_end();
-          // Newer torch reworked at::cuda::CUDAGraph: capture_end() no longer
-          // instantiates the executable graph here, and the C++ replay() (unlike
-          // the Python wrapper) does not instantiate on demand -- it throws
-          // "replay before instantiated". Instantiate explicitly once the graph
-          // is captured. Guarded by has_graph_exec() so it stays correct if a
-          // future torch goes back to instantiating inside capture_end().
-          if (!compiled_engine->cudagraph.has_graph_exec()) {
-            compiled_engine->cudagraph.instantiate();
-          }
-
           if (compiled_engine->profile_execution) {
             cudaError_t debug_dump_err = cudaGraphDebugDotPrint(
                 compiled_engine->cudagraph.raw_cuda_graph(),

tests/core/conversion/converters/test_scaled_dot_product_attention.cpp

@@ -79,7 +79,8 @@ TEST(Converters, ATenScaledDotProductAttnMaskBoolDoesNotProduceNaN) {
   auto value = at::arange(16, {at::kCUDA}).to(at::kFloat).reshape({1, 1, 4, 4}) / 11.0;
   auto attn_mask =
       at::tensor(
-          {1, 0, 1, 0, 0, 1, 0, 1, 1, 1, 0, 0, 0, 0, 1, 1}, at::TensorOptions().dtype(at::kBool).device(at::kCUDA))
+          {true, false, true, false, false, true, false, true, true, true, false, false, false, false, true, true},
+          at::TensorOptions().device(at::kCUDA))
           .reshape({1, 1, 4, 4});
   auto params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
   auto jit_results = torch_tensorrt::tests::util::RunGraph(g, params, {query, key, value, attn_mask});
@@ -89,8 +90,8 @@ TEST(Converters, ATenScaledDotProductAttnMaskBoolDoesNotProduceNaN) {
   params = torch_tensorrt::core::ir::get_static_params(g->inputs(), {});
   auto trt_results = torch_tensorrt::tests::util::RunGraphEngine(g, params, {query, key, value, attn_mask});
 
-  ASSERT_FALSE(torch::isnan(jit_results[0]).any().item<bool>());
-  ASSERT_FALSE(torch::isnan(trt_results[0]).any().item<bool>());
+  ASSERT_FALSE(at::isnan(jit_results[0]).any().item<bool>());
+  ASSERT_FALSE(at::isnan(trt_results[0]).any().item<bool>());
   ASSERT_TRUE(torch_tensorrt::tests::util::almostEqual(jit_results[0], trt_results[0]));
 }
 

tests/py/dynamo/conversion/harness.py

@@ -311,11 +311,12 @@ def generate_graph(
     ):
         mod = mod.eval()
         if use_dynamo_tracer:
+            device = default_device()
+            mod = mod.to(device.to(torch.device))
             if torch_export_dynamic_shapes is None:
                 torch_export_dynamic_shapes = get_dynamic_shapes_args(
                     mod, original_inputs
                 )
-            device = default_device()
             torch_export_inputs = get_torch_inputs(original_inputs, device)
             exported_program = torch.export.export(
                 mod,