CUDA update broke my throughput

[cobaltspark]

Curious what people here are seeing.

I'm mostly working with Kimi in PyTorch on B200. The part I'm trying to understand is INT4.

If you've tested something similar, what did you keep and what did you stop bothering with?

This is mostly for local work rather than public serving.

If I had to summarize the lesson so far, it would be that clean setup and observability beat clever guesses.

[dockember]

A couple things I'd check first. On dual 3090, the easiest win is often to back off the aggressive settings and see whether the problem survives.

[fusedtinker]

My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Qwen and compare from there. Once you know the plain path works, then add the faster path back in piece by piece.

If you post your exact settings, people here can probably narrow it down pretty fast.

[cobaltspark]

What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at token cost. Once you know the plain path works, then add the faster path back in piece by piece.

If you post your exact settings, people here can probably narrow it down pretty fast.

[amberpatch25]

Appreciate the detail. I don't think your conclusion is far off. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.

[latentgarden]

What worked for me was pretty simple. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at consumer GPUs.

My first guess would be this: On GB200, 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. The logs were somehow calmer than I was.

[phaseport36]

I'd approach it in a boring order. Start by verifying the driver/runtime combo, then test with a smaller prompt, then look at router policy. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.

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

[orbitbench6]

I'd approach it in a boring order. If you're on llama.cpp, I'd confirm the version first and then reduce variables until the behavior is obvious. Once you know the plain path works, then add the faster path back in piece by piece.

[thetapath]

Maybe, but I had the opposite result. I get why that seems right, but on B200 the bottleneck felt elsewhere.

[thetapath]

Interesting, I hadn't tried that. That matches what I saw on M3 Max.