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
perf(deepseek_v4): qdq the attention input once — wq_a and wkv share it
7c6cdc8 JustVugg committed 3 months ago
perf(deepseek_v4): thread-local scratch for qdq activations (bit-identical)
853a90a JustVugg committed 3 months ago
Merge pull request #1075 from JustVugg/perf/qrow-hoist-k3
5d0986c Vincenzo Fornaro committed 3 months ago
perf(kimi_k3): hoist expert activation quantization to layer level (bit-identical)
a7f2f64 JustVugg committed 3 months ago
Merge pull request #1071 from JustVugg/perf/qrow-hoist
c945cd5 Vincenzo Fornaro committed 3 months ago
Merge pull request #964 from 8PotatoChip8/v4/expert-store-registry
8512368 Vincenzo Fornaro committed 3 months ago
perf: hoist expert activation quantization to layer level (bit-identical)
5e8e99d JustVugg committed 3 months ago
Merge pull request #1068 from terrizoaguimor/refactor/family-registry-dispatch
58e6c44 Vincenzo Fornaro committed 3 months ago
refactor(core): finish registry-owned family dispatch
2ecb5d4 Mario Gutierrez committed 3 months ago
build: install cluster.py; complete test_cluster_protocol prerequisites
6bfbd3e eli committed 3 months ago
docs: correct the fmt=6 down-rotation accounting in the expert_ffn comment
6608dca eli committed 3 months ago
ci: exercise the cluster parity gate against the fmt6/fmt4 fixtures
cea3779 eli committed 3 months ago
test: token-exact parity gate — local vs cluster-delegated sharding (#7)
4982f66 eli committed 3 months ago
tools: quantize glm_tiny routed experts to fmt=6 (E8/IQ3) and fmt=4 (grouped int4)
edcfd05 eli committed 3 months ago
fix: reload worker expert slot when slab is missing (use-after-free after failed load)
43ed2a7 eli committed 3 months ago
fix: route cluster worker through expert_ffn and rotate fmt=6 input
cf872b0 eli committed 3 months ago