functionstackx

Perfetto SGLang profiler trace during CI to make it easier to improve perf

1

as per lmsys xai sglang suggestion, https://github.com/sgl-project/sglang/actions/runs/18701520115/job/53331131552#step:4:402. prefer get the full log from start to shutdown. For the profilers, we are looking for torch profiles How to generate: https://docs.sglang.ai/developer_guide/benchmark_and_profiling.html#profile-a-server-with-sglang-profiler> Examples...

probably with `lm-eval` 1. create PoC to experiment with gsm8k h100 2. write design doc on how to integrate eval quality clean-ly into the codebase keeping in mind we want...

should be able to test specific conc @cquil11

lmsys sglang engineers r saying that the pareto frontier is not optimal as dp_attention, and other parallelism strategy is hard set and the only thing that get sweeped is fp4...

theoretical model tflop/s/gpu & MFU

1

WIP and have draft attention formula for llama70b, gptoss attention sink & deepseek mla but put on hold till we finish all the p00 refactors