Expand description
Moirai-2.0 inference on the Burn backend (Flex CPU).
Packed scaler, patch tokenizer, prefill + KV-cached decode with partial
RoPE, QK-norm attention, same-variate bias, SwiGLU FFN, quantile output
head, with weights loaded straight from GGUF via
zsfm_burn::GgufWeights.
Partial RoPE runs on host data via zsfm_burn::rope::apply_interleaved_rope
(Flex CPU: the roundtrip is cheap relative to attention).