GPT OSS 20b Perf Analysis

[mvasiljevicTT]

Run full model from tt-xla:

pytest -x -v tests/benchmark/test_llms.py::test_gpt_oss_20b_tp 

Run less layers:

pytest -x -v tests/benchmark/test_llms.py::test_gpt_oss_20b_tp --num-layers 4

Full model perf report based on nightly run from 3.3.2026:

Layers in gpt oss alternate between using sliding window and full attention. Perf report for two consecutive layers (one odd, one even):

Expecting notable improvments:

• SDPA fusing - expected to highly improve matmuls, remove some of the concat and slice ops
• RMS fusing - removes expensive fillpad before reduce, tilize/untilize, reshapes...

cc: @odjuricicTT

[mvasiljevicTT]

There are 19 FillPadDeviceOperations per layer:

• 2: 1 per RMS norm sum - ~300us each - will be solved with rms fusing
• 16: 8 per reduce scatter (one 8*5us, other 8*46us)
• 1: for final sum along experts of routing_weights * expert_output, ~600us

these are analyzed in this issue.