Update PyTorch pin

[anmyachev]

Patch conflict due to pytorch/pytorch#134666

Benchmark CI:

• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11855565969 (elapsed_time failed)
• https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11856129996 (legacy profiler works)

Inductor CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11856999355 (passed)

[alexbaden]

Looks good - can we run inductor tests and also verify that the timing is reasonable for the benchmark results?

[anmyachev]

Looks good - can we run inductor tests and also verify that the timing is reasonable for the benchmark results?

Sure, here is https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11856999355 CI for inductor.

verify that the timing is reasonable for the benchmark results?

Do you mean for tutorials? The separate benchmark workflow doesn't work with the new compiler yet (to be precise elapsed_time doesn't work).

[alexbaden]

Do you mean for tutorials?

Yes, whatever reports out to the dashboards w/ the previous elapsed time hack (probably just tutorials).

[anmyachev]

Do you mean for tutorials?

Yes, whatever reports out to the dashboards w/ the previous elapsed time hack (probably just tutorials).

Overall the results look reasonable, but there are places where the numbers change dramatically and this is noticeable, for example for one of the tutorials (and usually it is noticeable on small dimensions). I also believe that since the new implementation of profiling should have less impact on runtime, the increase in performance in some cases should definitely be noticeable.