vLLM#52663: FP8 on RDNA3 exceeds 600s engine-ready timeout (no tuned configs)

Fields

id

sources/src5-gotcha-vllm-52663-fp8-rdna3-cold-start

URL

https://github.com/vllm-project/vllm/issues/52663

Publisher

vLLM

Kind

forum

Published

2026-08-17

Accessed

2026-10-04

notes

FP8 checkpoint on RDNA3 (gfx1100) loads and serves fine but exceeds 600s engine-ready timeout on first start. Root cause: no tuned W8A8 block-FP8 configs shipped for AMD_Radeon_Graphics, causing per-shape Triton JIT compilation. Cold start: 1474s. Warm: 49s. Fix: VLLM_ENGINE_READY_TIMEOUT_S=1800. Demonstrates ROCm cold-start performance issue.

Notes

FP8 checkpoint on RDNA3 (gfx1100) loads and serves fine but exceeds 600s engine-ready timeout on first start. Root cause: no tuned W8A8 block-FP8 configs shipped for AMD_Radeon_Graphics, causing per-shape Triton JIT compilation. Cold start: 1474s. Warm: 49s. Fix: VLLM_ENGINE_READY_TIMEOUT_S=1800. Demonstrates ROCm cold-start performance issue.

Cross-references

This record references no other records.