[ET-VK][ez] Support exporting of custom operator calls via higher_order_auto_functionalized

Pull Request resolved: #5884

As title. This diff adds the ability to partition custom op calls to the Vulkan delegate.
ghstack-source-id: 246400774
@exported-using-ghexport

Differential Revision: [D63913434](https://our.internmc.facebook.com/intern/diff/D63913434/)

Author: SS-JIA

backends/vulkan/partitioner/supported_ops.py

@@ -84,6 +84,8 @@ def __contains__(self, op):
     # Convolution
     exir_ops.edge.aten.convolution.default,
     exir_ops.edge.et_vk.conv_with_clamp.default,
+    # Custom ops
+    "llama::sdpa_with_kv_cache",
 ]
 
 NO_DYNAMIC_SHAPE = [

backends/vulkan/partitioner/vulkan_partitioner.py

@@ -143,16 +143,22 @@ def is_node_supported(
     def _is_node_supported(
         self, submodules: Mapping[str, torch.nn.Module], node: torch.fx.Node
     ) -> bool:
+        target = node.target
+        if node.target == torch.ops.higher_order.auto_functionalized:
+            first_arg = node.args[0]
+            assert isinstance(first_arg, torch._ops.OpOverload)
+            target = first_arg.name()
+
         if self.is_linear_permute(node):
             return True
 
         if self.is_in_local_scalar_dense_chain(node):
             return True
 
-        if node.target not in VulkanSupportedOperators._ops:
+        if target not in VulkanSupportedOperators._ops:
             return False
 
-        features = VulkanSupportedOperators._ops[node.target]
+        features = VulkanSupportedOperators._ops[target]
 
         if self.require_dynamic_shapes and not features.supports_dynamic_shape:
             return False

backends/vulkan/vulkan_preprocess.py

@@ -40,6 +40,10 @@
 
 from executorch.exir.program._program import _copy_module
 
+from torch.export._remove_auto_functionalized_pass import (
+    unsafe_remove_auto_functionalized_pass,
+)
+
 DEFAULT_DEBUG_HANDLE = 65535
 
 
@@ -52,6 +56,8 @@ def preprocess(  # noqa: C901
         program: ExportedProgram,
         module_compile_spec: List[CompileSpec],
     ) -> PreprocessResult:
+        program = unsafe_remove_auto_functionalized_pass(program)
+
         passes = [
             RemoveCloneOpsTransform(),
             AddmmToLinearTransform(),