Spark-vllm-docker older version/commit. Is it viable?

instanttensor broke gemma4-26b-a3b-it as far as i found,

safetensors makes it usable and after 2 days I can say it for sure.

But… it’s slower!

Latest vLLM bundled with eugr spark-vllm-docker in almost 75% slower.
I’m not sure which version I used before the pull I did 3 days ago.

I think it was vLLM 0.19.1rc0 + mod gemma4 and was a lot faster.
I would like to revert to it cause it worked better.
8/10 concurrent request (3-5k tokens) was insanely faster than “today”.
Grafana shows the previous version had a 3x aggregte troughput.

Is there a way to build an old commit afte pulling it?
As far as I undestand the flow in build-*.sh, the compiled wheels are downloaded from an “untracked” url; they are not part of the commit.

If I try to rebuild using paramether -vllm-ref v0.19.1rc0 I suspect I will get error cause Dockerfile whill not match the wheels downloaded.
Are my assumption wrong?

I would like to do some test but I cannot “stop” the current unit to make some test.

Do some of you know if is possible to get/build/use an old commit and go back in time?

you can run the build command with --vllm-ref, the default value is main

example:./build-and-copy.sh --tag v0.19.2rc0 --vllm-ref v0.19.2rc0

And then apply the recipe with the mod

Did u try and got success?
Did you made a --cleanup before?
Aniway, I’ll try after the current processing batch… ;) anxiety…

The command works, not cleanup needed, just a different tag.

Are you sure you are not hitting the performance regression bug? Try to power down the Spark, unplug the usb-c cable from it, wait 10 seconds, plug it back in and start it back up.

I updated to latest vLLM v.20.x and applied latest update to OS (via Nvidia Sync). Latest operation rebooted the device.
Yeah! I’m back to past reference performance for PP/TG.
:)