From 39d67bab94156656f07f06de1b15b92eb3910608 Mon Sep 17 00:00:00 2001 From: Siddartha Pothapragada Date: Wed, 16 Sep 2026 01:19:10 -0700 Subject: [PATCH 1/2] [Qualcomm] Re-enable FP MHA test skipped since QNN 2.26 Summary: Remove the stale @unittest.skip decorator on TestQNNFloatingPointOperator.test_qnn_backend_mha. The skip dates to Sep 2024 ('failed to lower in QNN 2.26'); the quantized twin has no skip, MHA lowering support has since landed, and the rework framework covers MHA. This un-skips mha across test_qnn_delegate_x86, the fb simulator, and the recipe-session simulator. Test Plan: ast.parse passes on the edited file. Needs CI run of test_qnn_delegate_x86 and fb:test_qnn_delegate_simulator mha tests to confirm pass. --- backends/qualcomm/tests/test_qnn_delegate.py | 1 - 1 file changed, 1 deletion(-) diff --git a/backends/qualcomm/tests/test_qnn_delegate.py b/backends/qualcomm/tests/test_qnn_delegate.py index 20930534809..0d5694742d8 100644 --- a/backends/qualcomm/tests/test_qnn_delegate.py +++ b/backends/qualcomm/tests/test_qnn_delegate.py @@ -2045,7 +2045,6 @@ def test_qnn_backend_mean(self): test[QCOM_MODULE], test[QCOM_SAMPLE_INPUTS] ) - @unittest.skip("failed to lower in QNN 2.26") def test_qnn_backend_mha(self): module = MultiheadAttention() # noqa: F405 sample_input = (torch.randn(1, 197, 96),) From c133cb6a3df1aaf9d610cac0a8e59d7ba83cddba Mon Sep 17 00:00:00 2001 From: Siddartha Pothapragada Date: Wed, 16 Sep 2026 10:15:22 -0700 Subject: [PATCH 2/2] Revert: restore FP MHA skip - still fails to lower on QNN 2.37 CI on #22872 showed test_qnn_backend_mha fails with 'Failed to generate Qnn context binary' - the FP (fp16) SDPA decomposition (bmm + _safe_softmax + linear) is still rejected by the QNN compiler. Quantized twin passes. Restoring the skip; real fix needs MHA lowering work (e.g. mha2sha path). --- backends/qualcomm/tests/test_qnn_delegate.py | 1 + 1 file changed, 1 insertion(+) diff --git a/backends/qualcomm/tests/test_qnn_delegate.py b/backends/qualcomm/tests/test_qnn_delegate.py index 0d5694742d8..20930534809 100644 --- a/backends/qualcomm/tests/test_qnn_delegate.py +++ b/backends/qualcomm/tests/test_qnn_delegate.py @@ -2045,6 +2045,7 @@ def test_qnn_backend_mean(self): test[QCOM_MODULE], test[QCOM_SAMPLE_INPUTS] ) + @unittest.skip("failed to lower in QNN 2.26") def test_qnn_backend_mha(self): module = MultiheadAttention() # noqa: F405 sample_input = (torch.randn(1, 197, 96),)