Allow linear to be consumed by nvFuser by default

thunder/executors/nvfuserex_impl.py

@@ -2240,6 +2240,10 @@ def map_inside_replacement(x: Any) -> None:
 
 def _linear_check(a: TensorProxy, b: TensorProxy, bias: TensorProxy | None) -> bool:
     enable_linear: None | bool = get_compile_option("nv_enable_linear", "Enable nvFuser linear.")
+    # Enable linear by default if nvFuser version is 0.2.23 or later
+    # Because nvFuser 0.2.23 has a fix for https://github.com/NVIDIA/Fuser/issues/3369
+    if enable_linear is None and nvfuser_version() >= LooseVersion("0.2.23"):
+        enable_linear = True
     if not enable_linear:
         return False
     # Verify linear inputs and bias (optional) are supported tensors.

thunder/tests/test_nvfuser.py

@@ -347,6 +347,7 @@ def test_cse_rematerialization(executor, device, _):
         disable_torch_autograd=True,
         executors=executor.executors_list(),
         nv_enable_bookend=False,
+        nv_enable_linear=False,
     )
     compiled_func(x, y)