I build small tools and occasionally send patches upstream. Fascinated by all things AI.
| Project | PR |
|---|---|
| PyTorch | #201964 — support bytearray item and slice assignment and deletion in Dynamo (landed via 90b0a84) #201965 — support bytearray.join in Dynamo (landed via 599ec6e) #201966 — support bytearray.replace and bytearray.translate in Dynamo (landed via 2b01f6e) #205003 — support bytearray iterator __length_hint__ in Dynamo (landed via be6e8df)#205004 — support bytearray and bytearray iterator __reduce__ in Dynamo (landed via dc72472) |
| Triton | #12572 — reject generic linear encodings on non-AMD targets |
| llama.cpp | #29533 — split Vulkan matmul dispatch when it would exceed maxComputeWorkGroupCount #29766 — fix concurrent Vulkan first-device init #30711 — make Metal ggml_tri select instead of scaling by a 0/1 mask #30712 — implement the backward pass for TRI #31946 — remove the deprecated DIAG_MASK_INF / DIAG_MASK_ZERO ops #32454 — fuse NORM + MUL (+ ADD) into the CUDA norm kernel |
| flashinfer | #5763 — runtime page-size support for SM120 DeepSeek-V4 NVFP4 sparse-MLA decode; landed as the baseline inside #5914 #6139 — keep CuTe-DSL MoE autotune cache keys rank-invariant #9144 — retune when cached autotuner winners lack runner coverage |
| Mooncake | #4569 — pin BatchDesc across completion callbacks before freeBatchID #4839 — restore caller CUDA context after plugin copies #4840 — distinguish CUDA pointer errors from host memory |
| ktransformers | #2243 — AVX2 RAWINT4 routed-expert LoRA SFT backend; swiglu_limit plumbing fix |
| stable-diffusion.cpp | #2047 — add PixArt model family support #2062 — improve GGUF metadata parsing and reader selection #2074 — fix f16 overflow in Z-Image quantized matmuls on CUDA #2103 — keep MiniMax-H3 VAE weights resident across temporal chunks #2119 — demote resident params to disk residency when memory reclamation fails #2120 — halve VAE decode tiles on OOM retry instead of jumping to 256px #2124 — reclaim before accepting wrapped Vulkan iGPU memory estimates #2164 — add ADetailer support to server image generation #2167 — fix /v1/images/edits routing for reference-conditioned models#2196 — allow grouped temporal decoding for Wan2.2 TI2V 5B #2197 — balance singleton tails in short Wan decode plans #2317 — support Sol-H3-VAE decoding for MiniMax-H3 #2320 — fuse bias into INT8 CUTLASS epilogues on CUDA #2323 — fuse RMSNorm and RoPE in Sol-H3-VAE #2350 — trim MiniMax-H3 audio to video duration |
| ncnn | #7027 — Vulkan implementation of the Tile operator #7138 — x86 SIMD implementation of Softplus |
| sherpa-onnx | #3985 — forward provider config keys to the CUDA execution provider #4194 — MedASR CTC example for the Rust API #4197 — TeleSpeech CTC example for the Rust API |
| mimalloc | #1432 — pprof allocation failure handling and backtrace re-entrancy |
| exllamav3 | #406 — large-page memory for CPU MoE on Windows #521 — fix disk cache model load on Windows (os.statvfs is POSIX-only) |
| auto-round | #2456 — skip layers the AWQ GEMM kernel cannot serve in AWQ export |
- ncnn#7026 — skip GPU instance teardown at Windows process shutdown
- spawnfate — predicts Windows process-spawn failures before they happen (argv/PATH/PATHEXT/cmd.exe simulator, MCP-ready)
- compose-mica — wallpaper-aware window backdrop for Compose Desktop
- bililens — browser extension that turns Bilibili videos into long-form notes
- flutter_liquid_glasses — liquid glass material for Flutter
- uniwill-ec-charge-limit — battery charge-limit fix for Uniwill laptops
- tern-zh-cn — unofficial zh-CN localization for Tern


