[upstream #3489] [release/2.12] PT2E INT8 accuracy regression on shufflenet_v2_x1_0 (XPU): Acc@1 66.5 → 0.09 between release/2.11 and 2.12.0+xpu

[Stonepia]

[upstream pytorch#3489] [release/2.12] PT2E INT8 accuracy regression on shufflenet_v2_x1_0 (XPU): Acc@1 66.5 → 0.09 between release/2.11 and 2.12.0+xpu

Fixes https://github.com/intel-sandbox/torch-xpu-ops-exp/issues/364

Root Cause: The regression is caused by commit b3bc797 which added qconv_pointwise.tensor variant support. When PT2E quantization produces tensor-typed x_scale/x_zp (x_scale_zp_are_tensors=True path), the new lowering handles them but does not ensure w_zp is properly converted to int32 for the qconv path on XPU (unlike the qlinear path which has explicit dtype conversion at L798-806). Additionally, the TensorMeta construction failure in select_algorithm.py for XPU extern kernels means benchmarking requests get empty metas, potentially causing incorrect algorithm/kernel selection for grouped/depthwise convolutions like those in shufflenet.

---

Diff stat:

torch/_inductor/mkldnn_lowerings.py | 100 +++++++++++++++++++++++++++++-------
 1 file changed, 81 insertions(+), 19 deletions(-)