You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: README.md
+1-1Lines changed: 1 addition & 1 deletion
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -24,6 +24,6 @@ Server images are tagged in the format of `server-<ssl>-<avx>`.
24
24
-`noavx`: All AVX-related optimizations are disabled. Do not use this build unless you are working around some known bug, or running LLMs on a 10-year-old potato.
25
25
-`avx`: (Only) AVX instruction set is enabled. Might be useful if you are using some old CPUs that don't support AVX2.
26
26
-`avx2`: AVX2 instruction set is enabled. This build should support most modern/recent CPUs with reasonable performance.
27
-
-`avx512`: AVX512 and AVX512-VNNI instruction sets are enabled. Currently only some high-end or server-grade CPUs support these instruction sets, so check your hardware specs before using this build.
27
+
-`avx512`: AVX512 base instruction set is enabled. Currently only some high-end or server-grade CPUs support this instruction set, so check your hardware specs before using this build.
28
28
-`oneapi`: Experimental build with the Intel oneAPI compiler, inspired by ggerganov/llama.cpp#5067. Offers a ~30% speed increase (~20tok/s vs ~15tok/s) in prompt processing on my machine compared to `avx2` builds. Not updated daily.
0 commit comments