I build small tools and occasionally send patches upstream. Fascinated by all things AI.
| Project | PR |
|---|---|
| llama.cpp | #29533 โ split Vulkan matmul dispatch when it would exceed maxComputeWorkGroupCount #29766 โ serialise concurrent Vulkan first-device init so a second thread cannot use a half-built device |
| flashinfer | #5914 โ SM120 DeepSeek-V4 NVFP4 sparse-MLA decode; my runtime page-size work from #5763 incorporated as the baseline (co-authored in main) #6139 โ keep CuTe-DSL MoE autotune cache keys rank-invariant |
| Mooncake | #4569 โ pin BatchDesc across completion callbacks before freeBatchID to fix TransferEngine UAF #4839 โ restore caller CUDA context after plugin copies |
| ktransformers | #2243 โ AVX2 RAWINT4 routed-expert LoRA SFT backend (non-AMX x86 tier); swiglu_limit plumbing fix |
| stable-diffusion.cpp | #2047 โ PixArt-ฮฑ/ฮฃ model family support #2062 โ unblock Wan2.2 VACE GGUFs: fix 5-dim tensor loading + array metadata parsing #2074 โ fix f16 overflow in Z-Image quantized matmuls on CUDA #2103 โ keep MiniMax-H3 VAE weights resident across temporal chunks #2120 โ halve VAE decode tiles on OOM retry instead of jumping to 256px #2119 โ demote resident params to disk residency when memory reclamation fails #2124 โ reclaim before accepting wrapped Vulkan iGPU memory estimates #2167 โ fix /v1/images/edits routing for reference-conditioned models#2164 โ add ADetailer support to server image generation #2196 โ allow grouped temporal decoding for Wan2.2 TI2V 5B #2197 โ balance singleton tails in short Wan decode plans |
| exllamav3 | #406 โ large-page memory for CPU MoE on Windows #521 โ fix disk cache model load on Windows (os.statvfs is POSIX-only) |
| ncnn | #7027 โ Vulkan implementation of the Tile operator |
| sherpa-onnx | #3985 โ forward provider config keys to the CUDA execution provider |
| mimalloc | #1432 โ pprof allocation failure handling and backtrace re-entrancy |
| auto-round | #2456 โ keep layers the AWQ GEMM kernel cannot serve in fp16 during AWQ export |
- ncnn#7026 โ skip GPU instance teardown at Windows process shutdown
- Mooncake#4840 โ distinguish CUDA pointer errors from host memory errors
- spawnfate โ predicts Windows process-spawn failures before they happen (argv/PATH/PATHEXT/cmd.exe simulator, MCP-ready)
- compose-mica โ wallpaper-aware window backdrop for Compose Desktop
- bililens โ browser extension that turns Bilibili videos into long-form notes
- flutter_liquid_glasses โ liquid glass material for Flutter
- uniwill-ec-charge-limit โ battery charge-limit fix for Uniwill laptops