jakeatx/llamampere is a fork aimed at improving llama.cpp performance on Ampere-series GPUs, particularly the RTX 3090 and 3090 Ti. A v0.6.2.2 merge commit landed in this repository on November 5, 2026. It is the merge of PR #84, and its content is a change whereby the launcher no longer forces PLE prompt prefetch.
Prefetch is the act of fetching needed pages ahead of time while a prompt is being read. The removed setting was a mapped page prefetch that uses two workers (fault mode, LLAMA_PLE_PF_AHEAD=1000000), and it now remains in the engine default state, which is off. Passing LLAMA_PLE_PREFETCH=N can restore it by specifying N read workers. The message of GitHub commit 1c34224 states that with this change, reading a 100K-token prompt became approximately 30~45% faster.
The commit statistics are recorded as 5 files, 1 line added, 4 lines deleted, and in the presented patch, 4 lines are erased from a single file, flash-next.env, and 1 line is added. The 30~45% is a statement written by the maintainer in the commit message, and no comparative benchmark table or measurement environment is presented alongside it. There is no license in this record either.
So this change is closer to a default adjustment than a feature addition. It amounts to sharing the judgment that turning prefetch off is advantageous for prompt reading, and it leaves behind a form that can be reverted if desired. The repository description states that it supports more than 200K context with 262K full context and a 12GB configuration. However, from the record alone it is difficult to confirm what difference this change made under those conditions.