Files
ik_llama.cpp/github-data/pull_requests/225 - Examples _ Add new sweep-bench benchmark.md
2025-07-23 13:31:53 +02:00

1.6 KiB

🔀 #225 - Examples : Add new sweep-bench benchmark

Author saood06
State Closed
Created 2025-02-23
Updated 2025-04-26

Description

Port of 9488fbf1e4

This is a good tool to benchmark with as requested by #223.

As a very quick demo I generated this, just by running this ( ./llama-sweep-bench -c 2048 -ub 512 -m WizardLM-2-8x22B-IQ4_K_R4.gguf -ctk q8_KV -ctv q8_0 -fa --output-format jsonl and then sweep-bench-plot.py with the output).

performance_comparison_pp

performance_comparison_tg

  • Self-reported review complexity:
    • Low
    • Medium
    • High

💬 Conversation

👤 ikawrakow submitted a review the 2025-02-23 at 06:00:18: APPROVED

Thank you for this - can be very useful.


👤 ubergarm commented the 2025-04-26 at 18:01:12:

@saood06 thanks I'm a convert to llama-sweep-bench! It is indeed very useful.

I pushed a branch on my personal mainline llama.cpp fork just to use for testing performance across forks. I don't plan to open a PR to mainline, but just left it up there in case anyone else is using it. I'm guessing ik has something similar as we were comparing the new GLM-4 performance.

Thanks!