ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 09:09:50 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
ik_llama.cpp/github-data/pull_requests/210 - Repack also experts.md
Thomas eaa2510a28 Add GitHub data: filename sanitization (#640)
2025-07-23 13:31:53 +02:00

684 B
Raw Permalink Blame History

🔀 #210 - Repack also experts

Author ikawrakow
State ❌ Closed
Created 2025-02-19
Updated 2025-02-19

Description

When I implemented run time repacking, I required the tensor to be 2D to be eligible for repacking, I guess to simplify the code. But I forgot about MoE models, where expert weights are in 3D tensors.

This PR fixes that. This leads to very significant performance gains. E.g., for DeepSeek-Lite quantized with IQ4_XS, we get PP-512 = 545 t/s on the main branch, and PP-512 = 677 t/s with this PR when using run time repacking.