Systems engineer working on AI infrastructure, GPU kernels, and performance-critical software.
C++ · CUDA · Triton · Rust · HPC
I care about the part after the benchmark: whether a performance claim survives a different shape, a different machine, and a clean rebuild.
做 AI Infra / HPC 系统，关心性能结论能否被复现。
Open to: AI Infra / HPC / performance engineering roles and technical collaboration.
Website · Systems · Arena · Forks · Email
Three public organizations, each with a different job:
| Organization | What it is | |
|---|---|---|
| Systems | holtwood-systems | Original AI Infra / HPC work: kernels, inference runtime, serving control plane, reproducible evidence |
| Arena | holtwood-arena | Competition and benchmark record: CANN / FlagOS / Kunpeng / Muse, plus the public race lab |
| Forks | holtwood-forks | Upstream reading copies. No original work claimed; useful for source study and minimal repros |
The curated entry point is imshane.site/projects. GitHub is the implementation layer, not the only map.
| Project | What it is | Signal |
|---|---|---|
| paged-serving | Rust LLM serving control plane: paged KV, continuous batching, OpenAI-compatible SSE, C ABI to a real CUDA backend | control/data plane boundary |
| tiny-llm | CUDA/C++ inference runtime: GGUF, W8A16, tokenizer, paged KV, CUDA Graphs | end-to-end runtime |
| cuflash | FlashAttention from scratch in CUDA C++: scalar → WMMA tensor cores, FlashDecoding / Split-KV | kernel depth |
| trifuse | Fused Triton kernels (RMSNorm+RoPE, gated MLP, FlashAttention) registered through torch.library |
kernel breadth + integration |
| cuda-foundations | A reproducible CUDA operator path: SGEMM ladder → reusable kernels → inference components | teaching by construction |
| race | Performance engineering lab: Speed Demon, 1BRC, PPC, DIMACS, AHC; every number bound to hardware and commit | competition discipline |
Full map: holtwood-systems/portfolio.
Also: kvtier (KV cache tiering experiments) · ai-infra-radar (upstream contribution radar, personal).
I use competitions as a forcing function for measurement discipline.
- CANNJudge / Ascend 910B2 — CANN 2026 electricity-industry operator special (10 problems) and October ladder (4 problems), plus the earlier CANN challenge. Private workspaces are retained as archives; public methodology is in holtwood-arena.
- Kunpeng HPC — S2 convolution optimization (public, x86-verified; Kunpeng platform not re-measured) and S3 OpenFOAM static study (public, archived; not entered).
- FlagOS S2 — SGLang kernel optimization and MiniCPM5-2B inference throughput workspaces.
- Muse Spark Open Kernel Challenge — kernel and serving tracks.
Public workspaces: kunpeng-s2-conv · kunpeng-s3-openfoam · race.
Competition rules and other teams' work are respected: private workspaces stay private while the rules require it; public pages carry only sanitized, reproducible material.
I came to AI infrastructure through genomics and real-time systems: BGI (genomic-scale data workflows), Mindray (medical imaging), ZEGO (RTC). That background is why I care about memory bounds, I/O behavior, and end-to-end latency — not just kernel microbenchmarks.
The genomics work is archived under open-genomics in maintenance mode. The earlier application/tooling projects remain at holtwood-projects.
- Local / CPU correctness is not device correctness.
- Device correctness is not platform acceptance.
- A benchmark is only as useful as its hardware, shape, build, command, and variance record.
- If a claim does not survive a clean rebuild, it is not a result yet.
- Website: imshane.site
- GitHub: @holtwood
- Email: jiashuai.shi@qq.com