fix the llm test failing issue + add back cu126/cu128 (#4118)

Author: lanluo-nvidia

.github/scripts/filter-matrix.py

@@ -16,7 +16,7 @@
 # rtx 1.2 currently only supports cu129 and cu130
 rtx_cuda_versions: List[str] = ["cu129", "cu130"]
 # trt 10.14.1 currently only supports cu129 and cu130
-trt_cuda_versions: List[str] = ["cu129", "cu130"]
+trt_cuda_versions: List[str] = ["cu126", "cu128", "cu129", "cu130"]
 
 jetpack_container_image: str = "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
 sbsa_container_image: str = "quay.io/pypa/manylinux_2_39_aarch64"

py/torch_tensorrt/dynamo/lowering/passes/_FakeTensorUpdater.py

@@ -17,6 +17,15 @@
 )
 from torch.utils._ordered_set import OrderedSet
 
+# Try to import reinplace module - may not be available in all PyTorch builds
+try:
+    import importlib
+
+    reinplace_module = importlib.import_module("torch._inductor.fx_passes.reinplace")
+    _generalized_scatter = getattr(reinplace_module, "_generalized_scatter", None)
+except Exception:
+    _generalized_scatter = None
+
 
 # Adapted from torch._inductor.fx_utils.FakeTensorUpdater
 class FakeTensorUpdater:
@@ -159,8 +168,10 @@ def should_process_node(node: torch.fx.Node) -> bool:
             return node.op == "call_function" and (
                 isinstance(node.target, torch._ops.OpOverload)
                 or node.target is operator.getitem
-                or node.target
-                is torch._inductor.fx_passes.reinplace._generalized_scatter
+                or (
+                    _generalized_scatter is not None
+                    and node.target is _generalized_scatter
+                )
             )
 
         to_process = OrderedSet[int]()