Posts: 341
Threads: 51
Joined: Sep 2026
Reputation:
0
Posting this in case it saves someone else a little time.
I'm helping a small team set up a local workflow. We don't need a giant stack, but we do care about MoE serving.
Would rather learn from your scars than collect my own all at once.
measure twice, benchmark three times
Posts: 211
Threads: 21
Joined: Sep 2026
Reputation:
0
I tested this tonight. On GB200 with PyTorch, I saw basically the same behavior. The main thing that moved was latency, not quality.
7900 XTX / ROCm / somehow still optimistic
Posts: 37
Threads: 4
Joined: Sep 2026
Reputation:
0
My first guess would be this: Before rebuilding anything, I'd test one clean baseline with Mistral and compare from there.
Posts: 34
Threads: 6
Joined: Sep 2026
Reputation:
0
What worked for me was pretty simple. On A6000, the easiest win is often to back off the aggressive settings and see whether the problem survives. Once you know the plain path works, then add the faster path back in piece by piece.
Posts: 34
Threads: 5
Joined: Aug 2026
Reputation:
0
I'd approach it in a boring order. Before rebuilding anything, I'd test one clean baseline with DeepSeek and compare from there. A lot of these threads end up being one bad config plus a pile of conclusions drawn from it.
Posts: 40
Threads: 4
Joined: Sep 2026
Reputation:
0
Good point. I ended up in about the same place using PyTorch. Worth rechecking your defaults though, because one bad assumption can make the rest of the thread look confusing.
small models can still be good
Posts: 37
Threads: 2
Joined: Aug 2026
Reputation:
0
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. Not glamorous advice, but repeatable baselines save a lot of time.
Posts: 39
Threads: 2
Joined: Sep 2026
Reputation:
0
Need a little more detail here. Are you testing this on A6000 or something else?
