RTX 5090 + vLLM on Windows WSL2: ~140 tok/s Qwen3-14B for multi-user serving
A community report shows vLLM CUDA graphs working on an RTX 5090 under WSL2 2.7.0, reaching ~140 tok/s on Qwen3-14B-AWQ vs ~111 tok/s with Ollama, with the full config.
2-10 peopleCodingWriting assistant$3k+Sensitivity: High (client/confidential)
Hardware
| Item | Approx. price (as of 2026-10-16) |
|---|---|
| PC with NVIDIA GeForce RTX 5090 (32GB), Windows 11 + WSL2 | Not stated in source |
Source prices are as stated by the source. Retailer offers: prices checked 2026-10-16, may change. Some links may be affiliate links; we may earn a commission at no extra cost to you. This never changes what we list.
Software stack
- WSL2 2.7.0 (pre-release)
- Ubuntu 22.04
- CUDA 12.8
- vLLM 0.17.1
Models used
- Qwen/Qwen3-14B-AWQ
Good at
- ~140 tok/s with CUDA graphs; 26% faster than Ollama on the same model
- Runs on an ordinary Windows workstation
Weak at
- Needed a pre-release WSL2 and specific fixes; enforce-eager mode was ~17 tok/s
- Author notes a boot issue with nvidia-cdi-refresh
Rough total cost
Not stated in source.
Getting started
- Update to WSL2 2.7.0 and current NVIDIA driver.
- Install CUDA 12.8 and vLLM 0.17.1 in Ubuntu.
- Serve an AWQ model without --enforce-eager.
- Mask nvidia-cdi-refresh if using nvidia-container-toolkit.
Source & credit
[Community] RTX 5090 (Blackwell sm_120) + WSL2 2.7.0: CUDA graphs work — GitHub (vLLM issues)
Summarized from the source; quotes are the source author's words. Check the original before buying.