projects

Coursework systems projects, plus my upstream open-source work on LLM training and inference frameworks.


open source

Upstream contributions to LLM training and inference frameworks. Details and numbers are on the cv page.

verl-omni · diffusion/omni RL post-training · committer · 15 PRs merged

  • #495 · merged · [trainer, tests, doc] feat: async teacher scheduling for diffusion OPD on v1 separate_async
  • #498 · RFC · one-step-off teacher scheduling for diffusion OPD on the v1 async trainer
  • #513 · merged · [trainer, tests] fix: union colocated reward output into the v1 diffusion trainer batch
  • #493 · merged · [trainer, recipe, tests, doc] feat: multi-teacher OPD on the v1 sync diffusion trainer
  • #482 · merged · [trainer, cfg, tests] fix: give each ray worker group its own rendezvous port slice
  • #427 · merged · [worker, trainer, cfg, tests, doc] feat: multi-teacher routing and standalone teacher pool for diffusion OPD
  • #325 · merged · [worker, trainer, cfg, tests, doc] feat: teacher runtime MVP for diffusion OPD (#293)
  • #300 · merged · [trainer, algo, cfg] feat: add teacher-anchored continuous distillation losses for diffusion OPD
  • #292 · merged · [doc] fix: drop incorrect threshold caveat from rollout correction doc
  • #291 · merged · [trainer, doc, tests] feat: log rollout-train log-prob consistency metrics by default
  • #280 · RFC · rollout–training numerical consistency across SD3.5-Medium, Qwen-Image-20B, and BAGEL-7B
  • #281 · merged · [doc, cfg] chore: document reward-server profiling in reward docs
  • #279 · merged · [omni] fix: drop stale vllm-omni 0.22 step-count mapping in BAGEL adapter
  • #256 · merged · [trainer, reward, cfg, doc] feat: profile reward-model rollout servers
  • #252 · merged · [trainer, doc] feat: profile rollout servers + lightweight profiling recipe for FlowGRPO
  • #246 · merged · [reward] fix: drop duplicate reward/<key>/score metric for multi-reward training
  • #242 · merged · [reward] fix: drop duplicate critic/score metric for single-reward training

sglang-omni · Qwen3-Omni inference performance

  • #1161 · merged · [Qwen3-Omni Perf] Thinker prefill: batch multimodal embedding merge, remove per-request host syncs

vllm-omni · video diffusion parallelism & acceleration: SANA-Video and MAGI-2 Preview

  • #7174 · in review · [Diffusion] MAGI-2: regional compile coverage
  • #7187 · in review · [Bugfix][Diffusion] Keep nested layerwise blocks on the host
  • #5940 · merged · [Model] Add sequence parallelism to SANA-Video 2B
  • #5861 · merged · [Model] Add TP and CFG parallelism to SANA-Video 2B
  • #5882 · merged · [Model] Add Cache-DiT and CPU offload support for SANA-Video 2B
  • #5866 · merged · [Bugfix][Diffusion] Fix FLASH_ATTN cross-attention key-padding unpad
  • #5269 · merged · [Bugfix] Skip non-positive scheduled token spans in omni stage runners
  • #5059 · in review · [Core] Merge-on-load fast path for diffusion LoRA (shadow wrappers)