Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Commits on main
tools: whole-artifact digests in repack-manifest.json + verify-side check
cc4a589 monotophic committed 3 months ago
rans: harden the edges the review round reached — parse bounds, encoder bound, never-crash validator
4dc19b3 monotophic committed 3 months ago
docs: int4-rans256-g0 format specification (registry row proposed, not claimed)
a770262 monotophic committed 3 months ago
tests: rans codec/record oracle + repack/verify e2e battery
0877d8d monotophic committed 3 months ago
tools: int4-rans256-g0 repack + verify tools, optional librans_c bridge
e0ef310 monotophic committed 3 months ago
rans: in-tree rANS codec + int4-rans256-g0 chunk-record library
c0e7a80 monotophic committed 3 months ago
fix(win): test_st_mirror stale dirs + format-truncation warnings
83c6f08 Khalil Sleimi committed 3 months ago
Add --workers: parallel --indir conversion (default 1, no behavior change)
3bfcf9f RDouglasSharp committed 3 months ago
Merge pull request #665 from JustVugg/fix/amd-rocm-tooling-discovery-662-663
eb7a881 Vincenzo Fornaro committed 3 months ago
fix(tooling): recognize AMD/ROCm GPUs in doctor + resource_plan (#662, #663)
9109961 JustVugg committed 3 months ago
Merge pull request #661 from monotophic/kernel/avx512-matmul-i3
2a301ea Vincenzo Fornaro committed 3 months ago
quant: AVX-512 arm for matmul_i3 (fmt=5 int3-g64), high plane as __mmask64
23ca87a monotophic committed 3 months ago
Merge pull request #655 from steve-m/feat-e8-converter
fa599e0 Vincenzo Fornaro committed 3 months ago
Merge pull request #659 from JustVugg/fix/cuda-integrated-mem-budget-653
8c84da9 Vincenzo Fornaro committed 3 months ago
fix(cuda): resolve coli_cuda_device_integrated in the Windows DLL loader
176bb54 JustVugg committed 3 months ago
fix(cuda): correct RAM budget on integrated/unified-memory GPUs (#653)
bc6ccc4 JustVugg committed 3 months ago
tools: native SIMD encoder for fmt=6, 15x over numpy (byte-identical)
dbc312a Steve Markgraf committed 3 months ago
quant: AVX2 matmul_e8 — fmt=6 was 92% of decode on a scalar kernel
e2c3d94 Steve Markgraf committed 3 months ago
convert: block the quantizers by rows, and report progress on the local pass
8e2ee1f Steve Markgraf committed 3 months ago
tools: hoist the code-independent product out of the E8 sub-scale search (3x)
3544a12 Steve Markgraf committed 3 months ago