diff --git a/CHANGELOG.md b/CHANGELOG.md index 783aa13..5438b36 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -2,6 +2,7 @@ ## Unreleased +- Update the runtime extra to vLLM 0.32.x and GuideLLM 0.8.1 or later below 0.9. - Add `optimization.objective` to minimize a latency mean or percentile, or maximize a throughput metric, with manual `tune` spaces. Ranking, finalist selection, reports, retries, and reclassification follow its direction. diff --git a/docs/reference/compatibility.md b/docs/reference/compatibility.md index 03b9fb3..38a7289 100644 --- a/docs/reference/compatibility.md +++ b/docs/reference/compatibility.md @@ -4,9 +4,12 @@ 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. -The runtime extra targets stable vLLM **0.30.x** (`>=0.30,<0.31`) and GuideLLM -`>=0.7.3,<0.8`. vLLM 0.30.0 is the latest stable release checked on 2026-10-04; -see its [release notes](https://github.com/vllm-project/vllm/releases/tag/v0.30.0). +The runtime extra targets stable vLLM **0.32.x** (`>=0.32,<0.33`) and GuideLLM +`>=0.8.1,<0.9`. See the +[vLLM release notes](https://github.com/vllm-project/vllm/releases/tag/v0.32.0) +and [GuideLLM release notes](https://github.com/vllm-project/guidellm/releases/tag/v0.8.1). +This runtime pair passed the RTX 3080/WSL2 backend check recorded below; +earlier rows retain their tested versions. Install the current checkout on Linux with `pip install -e '.[runtime]'`. Historical evidence and captured requirement files below retain their original versions. @@ -26,6 +29,7 @@ available in this workspace. They remain explicitly unvalidated: | 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 | | 2026-09-13 | RTX 3080 | 1 | WSL2 showcase and baseline control | 616.64 / CUDA 13.0 | 3.12.3 | 0.28.0 / 0.7.3 installed | facebook/opt-125m | Completed: published vLLM Bench Serve evidence; showcase drift remains inconclusive | | 2026-10-04 | RTX 3080 | 1 | WSL2 temperature smoke | 616.92 / CUDA 13.0 | 3.12.3 | 0.30.0 / 0.7.4 | local OPT-125M | Passed: default zero, explicit zero, nonzero startup warnings, 6/6 vLLM and 4/4 GuideLLM requests, JSON/CSV/HTML, clean shutdown | +| 2026-11-17 | RTX 3080 | 1 | WSL2 runtime upgrade | 621.01 / CUDA 13.0 | 3.12.3 | 0.32.0 / 0.8.1 | local OPT-125M FP16 | Passed: 64/64 requests on the selected configuration per backend, temperature 0, P99 TTFT target, JSON/CSV/HTML, offline report/reclassification/retry planning, clean shutdown; short-context compatibility check | The RTX 3080 rows are environment diagnostics, not supported native-Linux validation. The September 1 attempt produced no accepted benchmark result; diff --git a/pyproject.toml b/pyproject.toml index f0d62ce..4d9f64e 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -23,7 +23,7 @@ classifiers = [ dependencies = ["Optuna>=4.0,<5", "PyYAML>=6.0"] [project.optional-dependencies] -runtime = ["vllm>=0.30,<0.31", "guidellm>=0.7.3,<0.8"] +runtime = ["vllm>=0.32,<0.33", "guidellm>=0.8.1,<0.9"] test = [ "build>=1.2,<2", "coverage[toml]>=7.6,<8",