This repository contains CPU-only AI infrastructure trace simulation labs.
Generate Perfetto/Chrome Trace timelines that compare unfused MoE, grouped GEMM, and megakernel-like execution:
python3 -m moe_trace_sim.cli --config configs/small.yaml --out outputs/smallRead the design and usage docs:
docs/moe_trace_sim_design.mddocs/moe_trace_sim_user_guide.md
Each scenario config carries its own experiment and assumptions sections so the generated report states the synthetic latency priors and the gap to real GPU behavior.
Generate Perfetto/Chrome Trace timelines and static PNG schedule diagrams that compare GPipe, 1F1B, Interleaved 1F1B, ZeroBubble 1F1B, a deliberately bad MoE 1F1B overlap, Chimera bidirectional pipelines, DualPipe, and DualPipeV pipeline training schedules:
python3 -m pipeline_trace_sim.cli --config configs/pipeline_small.yaml --out outputs/pipeline_smallRead the design and usage docs:
docs/pipeline_trace_sim_design.mddocs/pipeline_trace_sim_user_guide.md