Threadripper build for inference

[homelabhenry]

Posting this in case it saves someone else a little time.

If you were starting from scratch today for this use case, what would you choose first and what would you ignore?

Simple and boring is a valid answer.

[batchyard]

Posting this in case it saves someone else a little time.

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at quant drift. Not glamorous advice, but repeatable baselines save a lot of time.

[tensorworks]

Posting this in case it saves someone else a little time.

My first guess would be this: On 5090, the easiest win is often to back off the aggressive settings and see whether the problem survives. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it. Turns out 'just one more test' was a lie.

[edgequill]

Posting this in case it saves someone else a little time.

I'd approach it in a boring order. If you're on TensorRT-LLM, I'd confirm the version first and then reduce variables until the behavior is obvious. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[labnest]

I tested this tonight. I couldn't reproduce the slowdown until I pushed context much higher. The surprise was that memory pressure showed up before throughput collapsed.

There are a few different conversations hiding inside this thread, so being explicit about the goal helps.

[rocmmixer]

I hit almost the same issue last night.

[latentpeak22]

Appreciate the detail. I don't think your conclusion is far off.

[sanevault]

What worked for me was pretty simple. On 4090, the easiest win is often to back off the aggressive settings and see whether the problem survives. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

[phasenode]

I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with Qwen and compare from there.

[sanevault]

Tried to reproduce it just now. On B200 with MLX, I saw basically the same behavior. The main thing that moved was latency, not quality.