title: the same probe on the L4, default server against the deterministic flag
source: manuals/sglang-101/capture/gpu/out-l4/livecheck-2026-10-23.txt
lines: 7-18
lang: text
note: Read the pattern of the default row letter by letter, then the backend that the flag chose on this GPU.
---
Probe: one prompt 5 times alone, then 5 times in a batch of 16, temperature 0.

server                                    distinct texts   pattern
default                                   3                A, B, B, B, B, C, B, B, B, B
--enable-deterministic-inference          1                all ten the same

On the default server the first request alone differs and the four later requests alone match each other.
One batched request differed. On BF16 Qwen/Qwen3-8B (the earlier L4 run), the batched requests gave two more texts.

Deterministic mode with no --attention-backend, startup log:
Attention backend not specified. Falling back to 'fa3' for deterministic inference
The server ran on fa3 on SM 8.9 with no error. The default server on this GPU uses flashinfer.
