10-13-2026, 03:07 PM
I tested this tonight. On 5090 with PyTorch, I saw basically the same behavior. The surprise was that memory pressure showed up before throughput collapsed.
throughput matters, but tail latency matters too
|
prefill/decode separation in practice for Kimi
|
|
« Next Oldest | Next Newest »
|
Users browsing this thread: 1 Guest(s)