ci: bench: add more ftype, fix triggers and bot comment

[phymbert]

Motivation

• PR comment pops up in all PRs even unrelated to speed, it is a little bit distracting.
• benchmark results vary probably because a fixed seed is not set
• add more model ftypes

Proposition

• reduce file workflow path trigger to only llama.cpp, ggml.c and cuda files.
• add seed param in k6 script.js
• reduce the comment to the first line, so it does not use too much space in the PR, add a warning notice
• add q8_0 and f16 phi-2 quants
• add more metrics in the commit status to later on show performance history

Tests

Tested here on a self-hosted GCP L4 ( sic ^) ) :

• TEST tmp modif phymbert/llama.cpp#5 (comment)
• https://github.com/phymbert/llama.cpp/actions/runs/8544724268

References

• server: bench: continuous performance testing #6233 (comment)

• @ggerganov please restart the github runner manager on the azure T4 node with ci: start-github-runner-manager.sh add model download phi f16 ggml-org/ci#3

[phymbert]

@ggerganov Georgi, as more and more models are MOE based, I suggest later on adding mixtral8x7b, thoughts ?

[ggerganov]

Yes, we can continue to scale with more models in the future. For now, let's give it some more time to see how the existing benchmarks perform and gather feedback about what's useful.

[phymbert]

Understood, please merge once you have restarted the github manager

[ggerganov]

Updated to master of https://github.com/ggml-org/ci and restarted