FAME v0: MiniMax M2.7 attention kernel profiling

[vedularaghu]

Summary

• Adds FAME (Fireworks Attention/MoE) v0 profiling framework for MiniMax M2.7 single-layer attention
• End-to-end Jupyter notebook that orchestrates nsys trace collection, ncu kernel profiling, and roofline report generation via Docker
• Includes profiling results for prefill (seq_len=4096) and decode (past_kv=4096) on B200 GPU

What's included

Path	Description
fame/v0/configs/profile.yaml	Profiling configuration
fame/v0/load_layer.py	Builds one transformer layer from HF weights
fame/v0/run_attention.py	Runs prefill/decode with NVTX instrumentation
fame/v0/fame_v0.ipynb	Main notebook: nsys/ncu orchestration + report
fame/v0/workspace/*_nsys.json	Per-kernel nsys timing results
fame/v0/workspace/*_ncu.{csv,json}	ncu metrics (DRAM bytes, durations)
fame/v0/workspace/report.{md,json}	Final roofline report

Notes

• B200 sm__ops_path_tensor FLOPs counters return 0; report uses analytical FLOPs from model dimensions instead
• ncu must be run as root inside the Docker container due to RmProfilingAdminOnly=1 on the host
• Kernel-to-NVTX mapping uses launch-overlap heuristic to handle async GPU kernel dispatch

Test plan

• Notebook cells 1-13 execute end-to-end in Docker container
• nsys Phase A produces per-kernel timing for all 7 attention sub-kernels
• ncu Phase B collects DRAM bytes and kernel durations
• Phase C report shows analytical FLOPs, measured bytes, and arithmetic intensity

Made with Cursor