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
qwen36 tier: name QT_MAX_ROWS, pin replica-block placement, int8-container docs
3e16e2c mfethe1 committed 2 months ago
Merge pull request #1383 from JustVugg/feat/glm53-brain-profile
f18a54a Vincenzo Fornaro committed 2 months ago
Merge pull request #1385 from JustVugg/feat/qwen36-brain-profile
a87ec7b Vincenzo Fornaro committed 2 months ago
inkling, olmoe: HITS after every turn, so the Brain tab lights up
6f33f18 JustVugg committed 2 months ago
qwen36: Brain and Profile tabs (EMAP, HITS, PROF on the serve protocol)
3365fbb JustVugg committed 2 months ago
glm53: keep the dashboard emitters out of main's late passes (MinGW GCC ICE)
a6fd3c2 JustVugg committed 2 months ago
glm53: keep timing locals plain and away from the omp pragma (MinGW GCC ICE)
dda96f2 JustVugg committed 2 months ago
fix(build): declare sse41_kernels.h/gsgemv.h/qgemv.h as Makefile prerequisites
49d54b1 Jaapjan Tinbergen committed 2 months ago
Merge pull request #1381 from JustVugg/fix/ctx-honest-and-visible
9bc991e Vincenzo Fornaro committed 2 months ago
feat(web): reasoning depth selector for GLM (off/low/medium/high/max)
791345d Diego Moraes committed 2 months ago
fix(ctx): the context ceiling is announced, refused when impossible, and reported with the right number (#1376)
af3e7cf JustVugg committed 2 months ago
Merge pull request #1377 from JustVugg/fix/windows-memory-budget
db7d034 Vincenzo Fornaro committed 2 months ago
fix(memory): measure available RAM on every platform, and mind the commit limit (#1375)
ec6656e JustVugg committed 2 months ago
Merge pull request #1238 from ZacharyZcR/feat/experiment-manifests
d18fdf1 Vincenzo Fornaro committed 2 months ago
Merge pull request #1358 from monotophic/evidence/raw-adapter
5b64bc7 Vincenzo Fornaro committed 2 months ago
Merge pull request #1374 from kreuzzelg/qwen36-tier-fp8-fixes
f2853a0 Vincenzo Fornaro committed 2 months ago
Merge pull request #1361 from kreuzzelg/qwen36-auto-place
6443bba Vincenzo Fornaro committed 2 months ago
Merge pull request #1360 from kreuzzelg/qwen36-tier-invariants
9f25fe8 Vincenzo Fornaro committed 2 months ago
tests(tier): the two remaining exp_bytes assertions follow the allocator footprint
823caa8 kreuzzelg committed 2 months ago
tier: fit the small-allocation footprint to the measured cudaMalloc curve
40ff645 kreuzzelg committed 2 months ago
tier: do not inherit the OpenMP master's CPU mask for the uploader and CUDA threads
792b6f2 kreuzzelg committed 2 months ago
tier: charge VRAM at cudaMalloc granularity; accept COLI_GPU
d0a382d kreuzzelg committed 2 months ago
qwen36 tier: fp8 streaming mode for experts that do not fit in RAM
979025c kreuzzelg committed 2 months ago
docs(qwen36 tier): the two card names in the calibration were swapped
7841f02 kreuzzelg committed 2 months ago