txgsync's picture
|
download
raw
1 kB

Source layout

Folder Contents
engine/ The engines: Metal (mova_gpu.m, kernels_moe.metal) and CUDA (mova_cuda.c, kernels_moe.cu)
nslm/ The LFSR, the seed search (C, Metal and CUDA), the model folder reader and writer, the packer, the CPU reference forward
harness/ The command-line tools, the server, the tokenizer (ICU), the chat template, the tool-call parser, the platform layer
tests/ One test program for each part. The same tests run on each platform. GPU kernels go through kernel_backend.h
win/ POSIX shims for MSVC, and build.bat
tools/ The MLX loader, and Python tools for calibration and reference data
docs/ This documentation. Start at index.md

Rules for the code

  • The engine, the server and the packer are C99. Objective-C is only thin Metal glue. GPU code is Metal and CUDA.
  • Python is only for tools in tools/.
  • Each GPU kernel has a C reference, and a test compares the two.

Xet Storage Details

Size:
1 kB
·
Xet hash:
0f955c44413752292f062a1220daa4cb73c33e37adc66a4f75bdb77f982ab785

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.