zzstoatzz.io / embedz
sentence embeddings in pure zig: bge-small with an HF-exact tokenizer and an SME matmul
Star
Fork 0
Watch

Commits on main

model: read architecture, pooling and normalization from the snapshot main
zzstoatzz committed
docs: precision modes, x86 results and the remote bench loop
zzstoatzz committed
int8: quantize activation rows in 16-lane vectors
zzstoatzz committed
int8: AVX2 VPMADDWD matmul for hosts without AVX-512 BF16
zzstoatzz committed
ops: keep the attention query tile within AVX2's 16 vector registers
zzstoatzz committed
ops: tile attention over four queries and vectorize layernorm
zzstoatzz committed
bf16: compile the kernel only under LLVM; test and benchmark it in CI
zzstoatzz committed
bf16: AVX-512 BF16 matmul with runtime dispatch and a load-time self test
zzstoatzz committed
model: run the last layer on CLS rows only; add a caller-clocked profiler
zzstoatzz committed
model: map model.safetensors instead of reading it into the heap
zzstoatzz committed
ci: trigger the first pipeline on spindle.zzstoatzz.io
zzstoatzz committed
embedz: bge-small-en-v1.5 encoder in zig with HF-exact tokenizer and SME matmul
zzstoatzz committed