Elffuss precomputed prefixes
Made with love in Málaga.
The state of a fixed system prompt after the model has read it: the attention cache rows and the short-convolution state, for one model and one cache format. An Elffuss app's system prompt never changes, so its state never changes either. Downloading it lets the first answer skip reading the whole system prompt, which on a phone or on a CPU is the longest wait before the first word.
Each file is named by its key: the model file, its size, the cache precision, CPU or GPU, and the exact tokens. The engine only loads a file that matches what it would compute itself, and checks it token by token before using it.
Format: EFP1, a little-endian u32 header length, a JSON header, then the raw parts, each aligned to 4 bytes.
| File | Model | Path | System prompt |
|---|---|---|---|
faa5e6f4….efp |
LFM3 1.7B QAD-Q4_0 | CPU, f32 cache | Elffuss chat, Spanish |
10f6c432….efp |
LFM3 1.7B QAD-Q4_0 | CPU, f32 cache | Elffuss chat, English |
4d1e63d2….efp |
LFM3 1.7B QAD-Q4_0 | CPU, f32 cache | Elffuss chat with shared memory, Spanish |
5854bfd9….efp |
LFM3 1.7B QAD-Q4_0 | CPU, f32 cache | Elffuss chat with shared memory, English |
Generated by Elffuss Engine (Apache-2.0) for LFM3 1.7B.