Imported from jasl/vllm-ds4-sm120-harness (AGENTS.md). Install upstream with npx skills add jasl/vllm-ds4-sm120-harness. Copyright stays with the author.

Agent Instructions

This repository is the public, repo-independent validation harness for the DeepSeek V4 SM12x work. Keep rules here safe for public GitHub visibility.

Scope

• These instructions apply to this harness repository.
• When editing a vLLM checkout from this context, read and follow that vLLM checkout's own AGENTS.md first. For vLLM files, the vLLM repository's rules take precedence over this harness repository's rules.
• vLLM pull requests require DCO sign-off on every commit. When creating, amending, cherry-picking, rebasing, squashing, or force-pushing commits for a vLLM branch, use git commit -s or otherwise ensure each resulting commit has a valid Signed-off-by: Name <email> trailer before pushing.
• Do not put private hostnames, IP addresses, usernames, local absolute paths, tokens, or oracle bundle locations in tracked files. Use ignored local notes such as *.local.md for machine-specific details.

Subagents

The user has authorized subagent use for implementation, review, and codebase exploration related to this repository and the associated vLLM work.

• Use subagents when they materially help: independent code review, parallel codebase exploration, verification planning, or disjoint implementation slices.
• Keep small, sequential edits local instead of spawning agents by reflex.
• When delegating code changes, give each worker an explicit ownership area and remind it that other edits may exist in the codebase.
• Review subagent results before relying on them; do not treat their output as verification by itself.

Testing And TDD

• TDD is encouraged for behavior changes: add a focused failing test first, verify it fails for the expected reason, implement, then verify it passes.
• Preserve tests that carry long-term value: behavior, protocol boundaries, regression cases, parser compatibility, validation, and acceptance gates.
• Before finishing vLLM changes, self-review the implementation and prune TDD scaffolding that does not provide durable maintenance value. Do not leave mechanics-only fixtures, temporary probes, or duplicate assertions just because they were useful during development.
• Always report the exact verification commands that were run, including any test skipped or blocked by missing environment support.

Harness Changes

• Keep the runtime stdlib-only unless there is a clear reason to expand it.
• Unit tests may use project test dependencies such as pytest.
• For script entrypoints, preserve explicit interpreter overrides such as PYTHON=/path/to/python so the harness can run against a target vLLM venv.
• Treat local macOS results as harness validation only. GPU-path validation must run on an appropriate remote SM120/SM121 environment.

SM120 Documentation

• For new SM120 / SM121 optimization notes, follow docs/sm120/README.md.
• Do not append new detailed experiment narratives to docs/sm120_optimization_notes.md; treat it as the legacy evidence archive.
• Put new experiment packages under docs/sm120/experiments/ and durable conclusions under docs/sm120/decisions/.
• Keep EP-on/off and prefix-cache-on/off profile sensitivity explicit in new experiment and decision notes.