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
fp8: repack tool refuses loudly on zero repack-target tensors
18fe6d1 monotophic committed 4 months ago
fp8: refuse (not crash) on fmt=6/7 in qt_addrow/qt_matvec_rows
70cc594 monotophic committed 4 months ago
fp8: site-level wire regression test for qt_wire_mmap/qt_unwire_mmap
363342d monotophic committed 4 months ago
fp8: wire qt_scale_bytes() into qt_wire_mmap/qt_unwire_mmap
5528589 monotophic committed 4 months ago
fp8: invert THE DESIGN LANDMINE ambiguous-shape collision to fmt=1
c709a44 monotophic committed 4 months ago
fp8: repack tool (Z.ai FP8 shards -> byte-preserved fmt=7 container)
e621f6d monotophic committed 4 months ago
fp8: Metal mm_gemv kernel for fmt=7 native FP8-e4m3 passthrough
6ca61a1 monotophic committed 4 months ago
fp8: CPU read path + matmul_fp8, as fmt=7 (public ordinal, #524)
41fe5a7 monotophic committed 4 months ago
Merge pull request #709 from JustVugg/docs/readme-multi-model
292ed4c Vincenzo Fornaro committed 3 months ago
Merge pull request #457 from monotophic/metal/grouped-gemv
df5486c Vincenzo Fornaro committed 3 months ago
tests: byte-exact .coli_ssd fixtures; pin CRLF rejection in the shared vectors (#379)
0d15ba2 monotophic committed 3 months ago
test_ssd_probe.c: pin the largest-shard picker; unhook the sparse arm's precondition (#379)
543a8e6 monotophic committed 3 months ago
docs: st_dev volume identity is best-effort; document the new veto reasons (#379)
7e7f0a6 monotophic committed 3 months ago
resource_plan.py/doctor.py: say why a cache is pending, not "no probe yet" (#379)
d611598 monotophic committed 3 months ago
openai_server.py/coli: key the cap shim on model arch; --cap absent means absent (#379)
1d3862f monotophic committed 3 months ago
colibri.c: veto under-allocated shards; name the veto honestly; largest shard (#379)
56ab889 monotophic committed 3 months ago
docs: the .coli_ssd probe cache -- v2 format, veto, volume identity (#379)
0f3bffc monotophic committed 3 months ago
openai_server.py/coli: keep the cap-0 sentinel away from non-glm engines (#379)
b9114f4 monotophic committed 3 months ago
resource_plan.py: trust only a v2 SSD cache recorded on this volume (#379)
04e85d9 monotophic committed 3 months ago
colibri.c: steer the storage probe to verified-cold ranges; strict v2 cache (#379)
a3d76fc monotophic committed 3 months ago
doctor.py/resource_plan.py: surface the cached SSD probe (#379, S4)
42d72bc monotophic committed 4 months ago
colibri.c/coli/openai_server.py: platform-aware cache defaults + precedence (#379, S2/S3)
2e3cf86 monotophic committed 4 months ago