Skip to content

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

458b1d5
14b85bd
a0ba858
7f9a1af
5ed33b3
fix(dsv4): avoid sanitized AVX-512 stack load
JustVugg committed 3 months ago
9b6dba5
test(dsv4): cover reported cache capacities
JustVugg committed 3 months ago
0b953c2
perf(dsv4): bound expert miss selection cost
JustVugg committed 3 months ago
45fb8b2
9d6affa
perf(dsv4): consume scalar FP4 tiles in place
JustVugg committed 3 months ago
b644f27
perf(dsv4): specialize batch-one FP4 matvecs
JustVugg committed 3 months ago
13a4282
fix(make): build qwen36 before install
Jérôme Benoit committed 3 months ago
2305771
Merge pull request #1167 from JustVugg/fix/uring-mirror
Vincenzo Fornaro committed 3 months ago
5c822f7
fix(uring): complete mirror I/O semantics
JustVugg committed 3 months ago
2c2be5e
d2bb447
perf(dsv4): batch routed FP4 expert matmuls
JustVugg committed 3 months ago
a735c8f
74a5270
d014e30
perf(dsv4): index resident experts directly
JustVugg committed 3 months ago
3a217ef
6c22f98