[ExecuTorch] Preserve undelegated Linear ops in Llama demo export

extension/llm/export/builder.py

@@ -16,6 +16,7 @@
 from executorch.backends.transforms.duplicate_dynamic_quant_chain import (
     DuplicateDynamicQuantChainPass,
 )
+from executorch.backends.xnnpack.passes.convert_to_linear import ConvertToLinearPass
 from executorch.exir import EdgeProgramManager
 from executorch.exir.backend.partitioner import Partitioner
 
@@ -382,6 +383,10 @@ def to_executorch(self) -> "LLMEdgeManager":
             ExecutorchBackendConfig(
                 extract_delegate_segments=True,
                 passes=[
+                    # If there are Linear operations left in the graph, let's execute
+                    # them with the optimized op_linear rather than materializing a
+                    # transpose followed by a regular op_mm.
+                    ConvertToLinearPass(),
                     QuantFusionPass(),
                 ],
                 memory_planning_pass=MemoryPlanningPass(

kernels/optimized/optimized-oss.yaml

@@ -45,6 +45,11 @@
     - arg_meta: null
       kernel_name: torch::executor::opt_le_tensor_out
 
+- op: linear.out
+  kernels:
+    - arg_meta: null
+      kernel_name: torch::executor::opt_linear_out
+
 - op: mul.out
   kernels:
     - arg_meta: null