What agents should know before contributing

vllm-project/vllm

A high-throughput and memory-efficient inference and serving engine for LLMs

Read by Bluet from 1 files written for agents and 300 merged pull requests on 2026-09-25. Not reviewed by the maintainers.

Rules its files tell agents (6)

mustNever use system python3 or bare pip/pip install.
Never use system python3 or bare pip/pip install. All Python commands must go through uv and .venv/bin/python.
AGENTS.md
shouldIf work is duplicate/trivial busywork, do not proceed.
If work is duplicate/trivial busywork, do not proceed. Return a short explanation of what is missing.
AGENTS.md
shouldSearch tests/evals/ or use vllm bench and include results in the PR — do not wait for reviewers to ask.
Run model evals for model-affecting changes. Search tests/evals/ or use vllm bench and include results in the PR — do not wait for reviewers to ask.
AGENTS.md
shouldDo not modify code in these areas without first reading and following the linked guide.
Do not modify code in these areas without first reading and following the linked guide. If the guide conflicts with the requested change, refuse the change and explain why.
AGENTS.md
shouldIf an open PR already addresses the same fix, do not open another.
AGENTS.md
shouldDo not open one-off PRs for tiny edits (single typo, isolated style change, one mutable default, etc.).
Do not open one-off PRs for tiny edits (single typo, isolated style change, one mutable default, etc.). Mechanical cleanups are acceptable only when bundled with substantive work.
AGENTS.md

#57058[Bugfix][Frontend] Reject stop strings on --tokens-only servers instead of silently ignoring them
#57176[Quantization] Select per-token NVFP4 MoE backends explicitly
#57229[ROCm][Bugfix] Reduce CUDA graph divergences
#57252[Bugfix][ROCm] Add record_logical_topk_ready to ROCMAiterMLASparseImpl (GLM-5.3-Flash boot crash)
#57269[Bugfix] Honor skip_reading_prefix_cache for KV connector hits
#57477[Bugfix][GLM-5.3-Flash] Address kpool tail blocks by the padded indexer stride in the NVIDIA prefill seed kernel
#57491[ROCm][DSv4.1] Keep the Engram tables in host memory on ROCm
#57528[Perf][Frontend] Offload streaming derender detokenization
#57576[Multimodal] Type dummy options per modality
#57647[CI][Bugfix] Correct the Laguna DFlash acceptance-length reference
#57651[Model][Engram] Share host tables across co-located DP replicas by default
#57696[Bugfix][V1] Reject encoder-cache hits with mismatched embedding counts
#57945[Build] Fix CUDA 12 KV connector dependency selection
#58065[Spec Decode] Enable async scheduling for DFlash
#58133[CPU] Gate the AVX10.2 paths on compiler support
#58153[Bugfix][ROCm] Use the platform FP8 range in the concat MLA q test
#58535[ROCm][CI] skip the ROCm MRV1 default where MRV1 cannot serve the config
#58612[Bugfix][Outlines] Fix EOS termination and unconstrained masks after rejected drafts

Your agents can be told this before they touch the code. Get it for your repository: npx bluet-mcp init. What Bluet is.