Comparison https://github.com/intel/torch-xpu-ops/actions/runs/14900177123

Distributed test timeout is expected, UT failures are not related with this change

Please work with @RUIJIEZHONG66166 to skip the temp ut failures. And for the e2e performance, I noticed that there are some performance regression as below. Does it cause by machine difference?

Please work with @RUIJIEZHONG66166 to skip the temp ut failures. And for the e2e performance, I noticed that there are some performance regression as below. Does it cause by machine difference?

Category Model Target eager Target inductor Inductor vs. Eager [Target] Baseline eager Baseline inductor Inductor vs. Eager [Baseline] Target vs. Baseline [Eager] Target vs. Baseline [Inductor]
torchbench_bfloat16_training vision_maskrcnn 41.289908 35.392219 1.166638 38.862280 30.328529 1.281377 0.941205 0.856926
torchbench_bfloat16_training dcgan 1.914976 2.312614 0.828057 1.761685 1.995414 0.882867 0.919951 0.862839
torchbench_bfloat16_training mnasnet1_0 25.409208 21.975949 1.156228 27.014366 19.572720 1.380205 1.063172 0.890643
torchbench_bfloat16_training mobilenet_v3_large 27.558122 25.776550 1.069116 27.264004 23.698865 1.150435 0.989327 0.919396
torchbench_bfloat16_training fastNLP_Bert 57.760067 44.603815 1.294958 59.994632 41.839274 1.433931 1.038687 0.938020
huggingface_bfloat16_training BlenderbotForCausalLM 80.269073 78.515200 1.022338 77.331836 73.737783 1.048741 0.963408 0.939153
torchbench_bfloat16_training BERT_pytorch 44.388574 29.732542 1.492929 44.020327 27.952662 1.574817 0.991704 0.940137

Should be caused by machine, need time to confirm. I will collect performance this weekend on all CI machine to check it

Distributed test failed is expected

LGTM, do we have any method to reset the best?

Launch a test manually and upload the best.