vLLM PR #56085: add RTX 4090 (SM89) W8A8 block-FP8 Triton GEMM configs - without them lookup falls back to a stock config

Fields

id

sources/comm-github-vllm-56085-4090-block-fp8-configs

URL

https://github.com/vllm-project/vllm/pull/56085

Publisher

vLLM GitHub GitHub pull requests

Kind

repo

Accessed

2026-10-04

notes

Tuned configs for five shapes observed serving Qwen3.8-27B-FP8 with TP=2 on RTX 4090 (N=8192/K=5120, N=5120/K=3072, N=17408/K=5120, N=5120/K=8704, N=7168/K=5120). Companion PR #54372 adds fused-MoE Triton configs for RTX 4090. Both state the same class of problem: consumer cards are absent from vLLM's tuned-config tables, so kernel selection falls back to defaults.

Notes

Tuned configs for five shapes observed serving Qwen3.8-27B-FP8 with TP=2 on RTX 4090 (N=8192/K=5120, N=5120/K=3072, N=17408/K=5120, N=5120/K=8704, N=7168/K=5120). Companion PR #54372 adds fused-MoE Triton configs for RTX 4090. Both state the same class of problem: consumer cards are absent from vLLM's tuned-config tables, so kernel selection falls back to defaults.

Cross-references

This record references no other records.