Update inductor_xpu_test.sh: don't add --freezing in reference mode; add new mode: inference-with-freezing

[anmyachev]

I noticed because (although I'm not sure if that's root cause):

# works
python benchmarks/dynamo/torchbench.py --accuracy --float32 -d xpu -n10 --inference --only detectron2_maskrcnn_r_50_c4  --backend=inductor --cold-start-latency

# doesn't work
# run it from .scripts_cache/pytorch
../../scripts/inductor_xpu_test.sh torchbench float32 inference accuracy xpu 0 static 1 0 detectron2_maskrcnn_r_50_c4

torch-xpu-ops sources in main: https://github.com/intel/torch-xpu-ops/blob/5b4d7444484576f721d2295761cf8fafa924ef36/.github/scripts/inductor_xpu_test.sh#L30

in custom branch (on which they made a reproducer for us): https://github.com/intel/torch-xpu-ops/blob/6e05eb3f3cb84bceb818a9f0461da773c56ac3ef/.github/scripts/inductor_xpu_test.sh#L31

[pbchekin]

There is mode "inference-no-freezing" in the script. You can set this mode via workflow input

[anmyachev]

There is mode "inference-no-freezing" in the script. You can set this mode via workflow input

yes, it could, but it's not so obvious. For those who don't know (like me at the beginning :) ), it will seem like a separate special mod. And since the regression issue indicates a different mode, it's quite hard to think of using this one. Can we remove it as well?

[anmyachev]

• detectron2_maskrcnn_r_50_c4 works: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15580323890/job/43873837147.
• detectron2_fasterrcnn_r_50_fpn eager_1st_run_OOM (due to another torch+torchvision): https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15582351847/job/43880432340
• doctr_reco_predictor works: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/15582396516/job/43880576117

[pbchekin]

yes, it could, but it's not so obvious. For those who don't know (like me at the beginning :) ), it will seem like a separate special mod. And since the regression issue indicates a different mode, it's quite hard to think of using this one. Can we remove it as well?

I am not against it) @vlad-penkin suggested some time ago to keep both modes (inference with freezing, inference without freezing). Potentially we can make the latter default?

[anmyachev]

yes, it could, but it's not so obvious. For those who don't know (like me at the beginning :) ), it will seem like a separate special mod. And since the regression issue indicates a different mode, it's quite hard to think of using this one. Can we remove it as well?

I am not against it) @vlad-penkin suggested some time ago to keep both modes (inference with freezing, inference without freezing). Potentially we can make the latter default?

Added inference-with-freezing. Is it better now?

[pbchekin]

Yep, works for me. Thanks!