inference/research
verifiedhigh · 0.90updated 2026-10-03

vLLM issue #52663: FP8 on RDNA3/gfx1100 exceeds the 600 s engine-ready timeout on first start

forum

Fields

id
sources/gotcha-vllm-issue-52663-rocm-fp8-cold-start
Publisher
vLLM GitHub issues
Kind
forum
Published
2026-08-17
Accessed
2026-10-03
notes
Contains measured cold-start timings (1474.26 s first init vs 49 s warm) and the tuned-config JSONs for two RDNA3 device-name strings.

Notes

Contains measured cold-start timings (1474.26 s first init vs 49 s warm) and the tuned-config JSONs for two RDNA3 device-name strings.

Cross-references

This record references no other records.