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: the BF16 and F32 resident rows on Vulkan (fmt 11, fmt 10)
7461fc3 JustVugg committed 1 month ago
inkling: f32 and bf16 residents on Vulkan too (fmt 10 / fmt 11)
dc83ac5 JustVugg committed 1 month ago
make: tests that include an engine link the Vulkan backend under VK=1
519e883 JustVugg committed 1 month ago
Merge remote-tracking branch 'origin/dev' into feat/vulkan-all-engines
eb83557 JustVugg committed 1 month ago
Merge pull request #1829 from JustVugg/fix/convert-qwen36-readme
66bc5da Vincenzo Fornaro committed 1 month ago
qwenimage: the DiT's int8 matrices on Vulkan, every image token per call
adf2a70 JustVugg committed 1 month ago
mimo: int8 dense matrices and an opt-in expert tier on Vulkan
cdabace JustVugg committed 1 month ago
deepseek_v4: VK=1 in Makefile.deepseek-v4, Vulkan init, hook and report
b57586e JustVugg committed 1 month ago
deepseek_v41: Vulkan init, dense hook and report (nothing to offload yet)
8523cb3 JustVugg committed 1 month ago
Merge pull request #1828 from JustVugg/feat/qwen3-moe
dc8426c Vincenzo Fornaro committed 1 month ago
vulkan: one entry point for every engine, and the engines link the backend
703facb JustVugg committed 1 month ago
docs/qwen36: Qwen3-Coder-30B-A3B, measured against the bf16 release
31ffe8c JustVugg committed 1 month ago
registry and gateway: Qwen3-Coder-30B-A3B and its REAP prunes
e69870b JustVugg committed 1 month ago
convert_qwen36: Qwen3-Coder containers, with a tiny oracle in CI
33d3466 JustVugg committed 1 month ago
qwen36: run Qwen3 MoE (qwen3_moe): no shared expert, ungated attention
6abaa52 JustVugg committed 1 month ago
Merge pull request #1827 from JustVugg/fix/serve-variant-model-id
fadf31a Vincenzo Fornaro committed 1 month ago