serving on Blackwell: what are you seeing?

[rasterpoint]

Spent some time on this today and wanted to compare notes.

I've gone back and forth between TensorRT-LLM and TensorRT-LLM for Llama. The interesting difference for me has been quant drift.

Interested in tradeoffs more than fan loyalty.

[pixelcrate82]

Maybe, but I had the opposite result. I've seen people blame tokens/sec when the real issue was much more mundane. Not saying you're wrong, just that I wouldn't lock that in as the lesson yet.

[kernelgrid]

Maybe, but I had the opposite result. I've seen people blame tokens/sec when the real issue was much more mundane. Not saying you're wrong, just that

My first guess would be this: If you're on llama.cpp, I'd confirm the version first and then reduce variables until the behavior is obvious. Not glamorous advice, but repeatable baselines save a lot of time.

[batchcraft]

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

[kernelgrid]

A couple questions before I guess too hard. What model, what runtime, and roughly what context size?

[pixelcrate82]

Ran a quick check on my box. My result was mixed: first run looked bad, second run looked normal. The surprise was that memory pressure showed up before throughput collapsed.

[vectorport]

A couple questions before I guess too hard. Are you testing this on GB200 or something else?

[macrotinker]

that fixed it for me too.

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

[syshouse]

What exact setup are you running? Are you testing this on M3 Max or something else?

[northcraft90]

A couple things I'd check first. On 7900 XTX, the easiest win is often to back off the aggressive settings and see whether the problem survives. Not glamorous advice, but repeatable baselines save a lot of time.