#3087 My experience updating from 0.1.43.1 to 0.1.45 and possible lessons from it

说明

System: AMD 7900X, 64GB RAM, RTX 4070 Ti Super (16GB)

I upgraded from 0.1.43.1 to 0.1.45 expecting a performance improvement or no change similar to previous upgrades and so I was surprised when the decode rate fell from 85-95 tokens/s to about 65-75 tokens/s. Still very good compared with other servers but not what Strata had already shown me was possible.

The root cause turned out to be that I need to increase `--vram-reserve-mib` from 700 to about 1600. This reduced the number of experts that can fit in VRAM but restored the performance. The hint to do this was in the log but not immediately obvious unless you have a problem and start looking for it.

And it made me realise that Strata is running on a delicate balance of many options which I am sure many of us cannot keep up with, and some people may be running sub-optimally without even knowing it because Strata can give good-ish results even when misconfigured. I had previously been right on the brink and this release pushed me over the edge.

After that I asked Qwen to read the Strata log and the docs and it came up with many additional good suggestions for my circumstances and gave me an in-depth explanation of what the `--vram-reserve-mib` message meant. It also proactively compared my results against logged community benchmarks to check I was back "in tune".

As Strata gains more and more options and flags it is getting more delicately balanced - especially as everyone has their own different blend of RAM/VRAM/GPU make/SSD performance and so each release can upset something unexpected even with all the pre-release testing.

So here are a couple of suggestions:

1. Provide a prompt as part of the "getting started" info that users can give to their model once it is running that guides it to read the Strata log and the documentation in Strata/docs and offers warnings and suggestions about whether performance is running comfortably (without thrashing/paging etc). This would augment what `--calibrate` can do for you.
 
2. Provide an .md file in the docs folder specifically to help agents do this kind of diagnosis. (Arguably it is not needed given how well they can already diagnose Strata issues from DETAILS.md but the code authors might see ways it is not picking things up from the existing documents that could be very succinctly explained to agents rather than humans. This file could also give actions or references to other docs for each message in the log.)