Remove patch for QNN to use the legacy export flow.

[lucylq]

🐛 Describe the bug

QNN test failure on pin-update: #7273

Test: pull / test-llama-runner-qnn-linux (fp32, qnn_8a8w, qnn) / linux-job (pull_request)
Error logs:

Error:  [Qnn ExecuTorch]: graph_prepare.cc:204:ERROR:could not create op: q::ScatterNd.tcm

Error:  [Qnn ExecuTorch]: graph_prepare.cc:1377:ERROR:Op 0x12cb0000000b0 preparation failed with err:-1

Error:  [Qnn ExecuTorch]: QnnDsp <E> "aten_index_put_default_1" generated: could not create op

Error:  [Qnn ExecuTorch]: QnnDsp <E> RouterX86 graph prepare failed 12

Error:  [Qnn ExecuTorch]: QnnDsp <E> Failed to finalize graph (id: 1) with err 1002

Error:  [Qnn ExecuTorch]: Failed to finalize Qnn Graph with error: 1002
Error:  [Qnn ExecuTorch]: Fail to compile QNN graph

Potentially related to export changes in the pin-update from functional to non-functional IR. As a workaround, there is a patch to use the legacy export flow for QNN.

cc @cccclai @winskuo-quic @shewu-quic @tugsbayasgalan

Versions

CI env.

[Jack-Khuu]

@mcr229 Are we still relying on the patch?

[mcr229]

@Jack-Khuu sorry, i might need a bit more context for this. What is the patch? what is the workaround?

[Jack-Khuu]

As a workaround, there is a patch to use the legacy export flow for QNN.

I'm not too familiar with this one either. Maybe @lucylq remembers
If not we can probably close for now

[lucylq]

Ah the patch is here:

executorch/extension/llm/export/builder.py

Lines 193 to 200 in 57891b7

	# TODO: this is temporary, as qnn flow does not work with new, non-functional export IR.
	# See issue: https://github.com/pytorch/executorch/issues/7373
	with patch.object(
	torch._utils_internal,
	"export_training_ir_rollout_check",
	return_value=False,
	):

Export has migrated to non-functional flow, which breaks QNN. The patch uses the old export flow (functional) for QNN to unblock the migration.

I don't think anyone is currently looking at the issue. @tugsbayasgalan, @cccclai and I did some investigation earlier but couldn't figure out how functionalization was changing the graph. May need some more investigation and help from Qualcomm side.

[lucylq]

Potentially resolved with: #12218

[cccclai]

It's fixed