Summary
Add a gate on compute-heavy sweep runs (full-sweep-fail-fast, full-sweep, non-canary-full-sweep-enabled, etc.) to prevent obviously-wrong / wasteful runs from consuming compute before they start.
Motivation
Compute-intensive sweeps are expensive, and there is currently no guardrail to stop a run that is clearly misconfigured or violating known bad patterns. This wastes runner time, increases queue times, and blocks other engineers' work. Stopping wasted compute here is P1.
Related: this ties into the broader InferenceX compute-improvement effort of stopping people from sweeping trtllm/atom before submitting/validating vllm/sglang, which increases queue time and blocks vllm/sglang work (ATOM engineers wasting other AMD engineers' time).
Approaches discussed
FORBIDDEN_HACK.md allow/deny list (preferred starting point): Maintain a simple markdown file of forbidden patterns/hacks that we append to as needed. A gate reads it and denies a run when a known bad pattern is detected. Cheap, deterministic, low false-positive risk.
- Run this gate only on fail-fast / canary runs (not every run) to limit overhead.
- LLM-based gate (fallback / later): Call an LLM to evaluate whether a run should be denied.
- Pros: catches novel bad patterns.
- Cons: risk of false positives; cost. If pursued, use a cheaper model (e.g. Haiku / Sonnet 5 / fable5) rather than an expensive review model. Estimated ~$1–2 per fable call, cheaper than the ~$20/review Claude enterprise PR code review.
Notes
- The CODEOWNER sign-off review is considered valuable and should stay.
- Keep the gate scoped to fail-fast / canary runs initially to control cost.
Scope
Start with the FORBIDDEN_HACK.md + gate on fail-fast/canary runs. Evaluate an LLM-based check later if needed.
Summary
Add a gate on compute-heavy sweep runs (
full-sweep-fail-fast,full-sweep,non-canary-full-sweep-enabled, etc.) to prevent obviously-wrong / wasteful runs from consuming compute before they start.Motivation
Compute-intensive sweeps are expensive, and there is currently no guardrail to stop a run that is clearly misconfigured or violating known bad patterns. This wastes runner time, increases queue times, and blocks other engineers' work. Stopping wasted compute here is P1.
Related: this ties into the broader InferenceX compute-improvement effort of stopping people from sweeping
trtllm/atombefore submitting/validatingvllm/sglang, which increases queue time and blocks vllm/sglang work (ATOM engineers wasting other AMD engineers' time).Approaches discussed
FORBIDDEN_HACK.mdallow/deny list (preferred starting point): Maintain a simple markdown file of forbidden patterns/hacks that we append to as needed. A gate reads it and denies a run when a known bad pattern is detected. Cheap, deterministic, low false-positive risk.Notes
Scope
Start with the
FORBIDDEN_HACK.md+ gate on fail-fast/canary runs. Evaluate an LLM-based check later if needed.