Skip to content

Compatibility

vLLM Optimizer targets Python 3.11–3.12 and native Linux experiment hosts with NVIDIA GPUs. This page records dated evidence rather than promising that every vLLM/CUDA combination works.

Validation matrix

The required L40 and H100 native-Linux runs need dedicated hosts and are not available in this workspace. They remain explicitly unvalidated:

Date GPU GPUs Mode Driver / CUDA Python vLLM / GuideLLM Model Outcome
2026-09-01 L40 1 native Linux Not run: host unavailable
2026-09-01 L40 2+ tensor parallel Not run: host unavailable
2026-09-01 H100 1 native Linux Not run: host unavailable
2026-09-01 H100 2+ tensor parallel Not run: host unavailable
2026-09-01 RTX 3080 1 WSL2 attempt 596.36 / CUDA 13.2 3.12.3 0.28.0 / 0.7.3 facebook/opt-125m Failed: vLLM reports UVA is not available
2026-09-02 RTX 3080 1 WSL2 V0 smoke 596.36 3.12.3 0.19.0 / 0.7.3 local OPT-125M Passed: two repeats/backend, 10/10 requests/backend, clean drains

The RTX 3080 rows are environment diagnostics, not supported native-Linux validation. The 0.28.0 attempt produced no accepted benchmark result. The V0 smoke verified streaming, final counters, request totals, reports, cleanup, and offline reclassification, but used a vLLM version outside the supported range.

The matrix still needs evidence for explicit ports, interruption cleanup, long generation, and at least one multi-GPU run on each supported native-Linux GPU.

Both GuideLLM 0.7.3 and vllm bench serve from vLLM 0.28.0 are the versions used by the reproducibility procedure. Workload arguments are passed through, so consult the documentation for the installed engine version when adopting newer options.

On the WSL2 attempt, vLLM 0.28.0 also exposed these optional server settings:

env:
  VLLM_USE_V2_MODEL_RUNNER: "0"
  VLLM_USE_FLASHINFER_SAMPLER: "0"

The first disables a runner that required unavailable unified virtual addressing; the second avoids a sampler requiring a local CUDA compiler. Native Linux systems may not need either setting.

CUDA compatibility

vLLM Optimizer does not bundle CUDA and does not require one exact CUDA release. The working combination is determined by vLLM, its PyTorch wheel, the NVIDIA driver, the GPU architecture, and any locally compiled kernels. Therefore, vLLM Optimizer cannot promise support for every CUDA version.

Follow the vLLM GPU installation guide for the vLLM release you install. Also check NVIDIA's CUDA compatibility table: newer CUDA toolkits require a sufficiently recent driver, and some PTX or new features require a newer driver even within a compatible major family.

The py3-none-any wheel is one universal pure-Python artifact for Linux, Windows, and macOS. It is not three OS-specific binaries. Native vLLM execution is unsupported on Windows; configuration and stored-result inspection work.