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
qwen38: run QSA per position at prefill, and time it by the clock
c7f5b7f laserpollo committed 1 month ago
qwen38: make the prefill chunk and its workspace runtime knobs
8745651 laserpollo committed 2 months ago
README: align Qwen3.8 entries with the shipped CUDA tier (1.12.0)
bff0f12 XBold committed 1 month ago
docs/qwen38: remove stale "no GPU backend" line (GPU tier merged in 1.12.0)
b9401a9 XBold committed 1 month ago
Merge remote-tracking branch 'upstream/dev' into feat/kimi-cuda-situ-pipeline
f2c305f ZacharyZcR committed 1 month ago
Merge pull request #1668 from JustVugg/perf/qwen38-idot-trunk
d99dc06 Vincenzo Fornaro committed 1 month ago
docs: qwen38 int8 trunk and vector FP8 kernel, measured; 1.12.1 notes for #1650 and #1668
cde709c JustVugg committed 1 month ago
test(systemone): resolve the scoring fixture during discovery
e82567a ZacharyZcR committed 2 months ago
Merge pull request #1650 from bokiko/fix/qwen38-stale-pin
8d546fa Vincenzo Fornaro committed 2 months ago
fix(v41): clamp output budgets and reject oversized prompts
71622eb ZacharyZcR committed 2 months ago
qwen38: the trunk threshold and skip list are read per load, not cached forever
dd10808 JustVugg committed 2 months ago
qwen38: rebuild the trunk table per load; the fixture checks run the int8 trunk too
85b7943 JustVugg committed 2 months ago
test_qwen38_native_weights: pin the FP8 dispatch to the table kernel
b8e23ac JustVugg committed 2 months ago