inference/research
verifiedmedium · 0.70updated 2026-10-04

nlzy/vllm-gfx906 issue #4: W8A8-quantized models very slow prompt processing and token generation

forum

Fields

id
sources/comm-nlzy-vllm-gfx906-issue-4-w8a8
URL
https://github.com/nlzy/vllm-gfx906/issues/4
Publisher
GitHub (user vega20user)
Kind
forum
Accessed
2026-10-04
notes
Full vLLM benchmark_serving.py output and startup log included. Environment: PyTorch 2.7.0+rocm6.3, gfx906:sramecc+:xnack-, Qwen3-0.6B.w8a8, tp1. Issue locked, still open.

Notes

Full vLLM benchmark_serving.py output and startup log included. Environment: PyTorch 2.7.0+rocm6.3, gfx906:sramecc+:xnack-, Qwen3-0.6B.w8a8, tp1. Issue locked, still open.

Cross-references

This record references no other records.