Stage-Separated Profiling with Warm KV Cache Support#9

Merged

TerrenceZhangX merged 13 commits into

microsoft:mainfrom

TerrenceZhangX:zhangt/stage-profiling

Mar 17, 2026

Commits on Mar 3, 2026

feat: add stage-separated profiling script (prefill vs decode)
Tong Zhang
committed
fix: concurrent batch sending + DP-attention trace stabilization
Terrence Zhang
committed
Consolidate stage profiling: unified run_configs.sh, utils/ package, black formatting
Terrence Zhang
committed
fix(cross_rank_agg): per-invocation min for symmetric comm kernels
Terrence Zhang
committed
lint
Terrence Zhang
committed
feat(profiling): add prefill sweep mode with exact token-ID prompts
Terrence Zhang
committed

Commits on Mar 4, 2026

feat(prefill): add bs dimension + OOM auto-skip + disable-chunked-prefill
Terrence Zhang
committed

Commits on Mar 17, 2026

refactor(stage-profiling): single-point API, dead code cleanup, shape validation, tests
Terrence Zhang
committed
fix: address Copilot review comments (#9)
Terrence Zhang
committed
fix: remove unused sys import and clarify unmatched-kernel docstring in shape_merge
Terrence Zhang
committed
fix: address Copilot review round 2+3
Terrence Zhang
committed
fix: address Copilot review round 4
Terrence Zhang
committed
Potential fix for pull request finding

TerrenceZhangX
and
Copilot
authored