LLM: tips untuk CPU

From OnnoWiki
Revision as of 04:26, 17 July 2024 by Onnowpurbo (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

Kata CGPT: saat pake CPU, coba:

1. Batch Processing u. kurangi overhead & speedup embedding.
2. Kurangi presisi model; float32->float16/int8;  speedup tanpa korbankan akurasi.
3. Buat versi kecil dari model yg sama.
4. Multi-threading.
5. Gunakan Intel MKL / OpenBLAS.

saya pakai model intfloat pak, lumayan cepet di CPU https://huggingface.co/intfloat/multilingual-e5-large


kalo pdf bisanya saya parse dulu textnya atau pakai ocr, terus embeddingnya disimpan di postgre pakai pgvector (https://github.com/pgvector/pgvector) agak effort sih