Skip to content

docs: revert gpu value over-correction (!221) + trim README to front-door

Jérôme Revillard requested to merge docs/gpu-fix-readme-trim into main

Follow-up to !221 (merged) (which merged before these two corrections landed).

  1. Revert gpu value changes — !221 (merged) changed VLLM_MAX_MODEL_LEN (65536→2048/4096) and VLLM_GPU_UTILIZATION values in gpu.md / install-guide / choosing-models, based on a reviewer treating env.t4/env.rtx6000 as the sole truth. That was wrong: env.t4/env.rtx6000 are GPU-profile OVERRIDES; the docker-compose DEFAULT is VLLM_MAX_MODEL_LEN=65536 (compose:639), which is what the docs reflected. Values restored to original (65536; GPU_UTILIZATION 0.6). The var-NAME fix (VLLM_GPU_UTIL → VLLM_GPU_UTILIZATION) is kept — that one is verified correct.

  2. Trim README to a front-door (254→82 lines) — removed project-structure / architecture / features / deployment / development sections that each have a dedicated page on the docs site (the "one doc, one place" principle). README now: badges, intro, why, quick start, docs-site pointer, contributing, license, resources.

Build verified clean. (env.t4/env.rtx6000 left intact per operator decision — a separate concern.)

Merge request reports