fix: build against TensorRT older than 10.15

Building Torch-TensorRT against TensorRT 10.13 or older fails to compile:

  core/runtime/TRTEngine.cpp:298: error: 'class nvinfer1::ICudaEngine'
      has no member named 'getAliasedInputTensor'

That API was added in TensorRT 10.15, together with the IKVCacheUpdateLayer
that produces the aliasing it reports. The runtime calls it unconditionally,
while the package declares a bare tensorrt dependency with no minimum version,
so an older TensorRT produces a compile error rather than a clear message.

The call is only used to reconcile the build-time aliased I/O map against what
the engine reports. On older TensorRT there is nothing to reconcile against, so
the build-time map stands on its own and the reconciliation is skipped.

Guard it with a named feature macro, following the existing TRT_HAS_NATIVE_NCCL
convention in the same header.

Author: shoumikhin

core/runtime/TRTEngine.cpp

@@ -292,6 +292,7 @@ TRTEngine::TRTEngine(
   // engines built outside Torch-TensorRT). User-declared aliases (kind=kUser)
   // are preserved as-is since TRT doesn't know about them.
   this->aliased_io = aliased_io;
+#ifdef TRT_HAS_ENGINE_ALIASED_IO
   for (const auto& out_name : this->out_binding_names) {
     // TRT returns nullptr / empty string for non-aliased outputs; any thrown
     // exception is a real error in the engine state and propagates.
@@ -315,6 +316,10 @@ TRTEngine::TRTEngine(
       it->second = AliasedIOSpec{std::string(aliased_in), AliasKind::kKVCacheUpdate};
     }
   }
+#else
+  // Older TensorRT cannot report an engine's aliasing, so the build-time map is
+  // the only source of truth and there is nothing to reconcile against.
+#endif
 
   // Precompute the set of input binding names that are the alias source of some
   // output (for the O(1) per-call membership test) and validate every aliased

core/runtime/TRTEngine.h

@@ -27,6 +27,12 @@
 #define TRT_HAS_NATIVE_NCCL 1
 #endif
 
+// TensorRT 10.15+ can report an engine's aliased I/O via
+// ICudaEngine::getAliasedInputTensor, alongside IKVCacheUpdateLayer that produces it.
+#if NV_TENSORRT_MAJOR > 10 || (NV_TENSORRT_MAJOR == 10 && NV_TENSORRT_MINOR >= 15)
+#define TRT_HAS_ENGINE_ALIASED_IO 1
+#endif
+
 // Full TRT NCCL collectives support requires both:
 // 1. PyTorch built with NCCL (USE_C10D_NCCL defined via Bazel)
 // 2. TensorRT 10.16+ (TRT_HAS_NATIVE_NCCL defined above)