[MLX] Small fixes to support HF runner (#23173)

As titled, to support
huggingface/transformers#48986

It has two small fixes:
1. Accept empty tensors as input to support starting state for dynamic
cache
2. Normalize symbolic ops

Author: metascroy

backends/mlx/partitioner.py

@@ -117,7 +117,28 @@ def __init__(self, compile_specs: List[CompileSpec] | None = None) -> None:
     def transform_for_pre_decomposition(
         self, exported_program: ExportedProgram
     ) -> ExportedProgram:
-        """Normalize SDPA ranks before deciding which ops to preserve."""
+        """Normalize symbolic ops and SDPA ranks before deciding which ops to preserve."""
+        from executorch.exir.dialects.edge._ops import EdgeOpOverload
+        from executorch.exir.passes.executorch_prim_ops_registry import (
+            _EXECUTORCH_SYM_OPS,
+        )
+
+        # Exporters may wrap symbolic primitives before this ATen-stage validation.
+        for module in exported_program.graph_module.modules():
+            if not isinstance(module, torch.fx.GraphModule):
+                continue
+            modified = False
+            for node in module.graph.nodes:
+                if (
+                    node.op == "call_function"
+                    and isinstance(node.target, EdgeOpOverload)
+                    and node.target in _EXECUTORCH_SYM_OPS
+                ):
+                    node.target = node.target._op
+                    modified = True
+            if modified:
+                module.recompile()
+
         return ExportedProgramPassManager([NormalizeSDPAInputRankPass()])(
             exported_program
         ).exported_program

backends/mlx/runtime/MLXBackend.cpp

@@ -68,7 +68,7 @@ namespace {
 array tensor_to_mlx(
     const ETTensor& t,
     const std::optional<TensorMeta>& expected_meta = std::nullopt) {
-  if (!executorch::runtime::tensor_is_contiguous(t)) {
+  if (t.numel() != 0 && !executorch::runtime::tensor_is_contiguous(t)) {
     throw std::runtime_error("tensor_to_mlx: input tensor is not contiguous");
   }
 
@@ -97,6 +97,11 @@ array tensor_to_mlx(
     shape.push_back(static_cast<int>(dim_size));
   }
 
+  // Empty inputs have no storage to wrap, but retain their shape and dtype.
+  if (t.numel() == 0) {
+    return ::mlx::core::zeros(shape, dtype);
+  }
+
   // SAFETY: MLX reads this data during async_eval() Metal command encoding,
   // which completes before the lock is released. The ET tensor must remain
   // valid until async_eval returns.