pallas-evolve

Plugin pallas-evolve from marketplace dracospark in Neuro-Spark-AI/dracospark. Commands read from the repo's manifest — 25 days ago.

Skills (7)

• analyze

  Use when analyzing batch TPU kernel evaluations or caller-level regressions — compare variants, investigate bottleneck hypotheses, select research lineages, and write evidence-backed reports.

• init-kernel

  Use when initializing a new kernel optimization project from an upstream Pallas kernel repository (passed with --repo) — clones upstream, generates ref/template/test/config files, validates baseline

• profile-brief

  Use when generating a profile brief from TPU evaluation artifacts — reads eval_result.json, LLO/HLO/trace files, classifies bottleneck, derives optimization priorities, writes profile_brief.md

• reflect

  Use when recording batch kernel optimization learnings to AGENT.md — iterates over N variant results, extracts failure patterns and successful optimizations, writes local round summaries with optional GitHub tracking

• start

  Use when starting a Pallas kernel optimization session on a TPU-VM (v6e) over SSH — reads config, verifies baseline, records locally with optional GitHub tracking, runs profile-informed think-submit-analyze-reflect-compact loop with lineage tracking

• submit

  Use when submitting a batch of Pallas kernel variants for TPU evaluation over SSH — runs evaluate.py on a TPU-VM (e.g. v6e-16), collects all EVAL_RESULTs, scp's artifacts back

• transfer

  Use when migrating or reusing a kernel optimization method across models, shapes, runtimes or callers; adapt guarded implementations and record target validation. Not for data migration or automatic production deployment.

Similar plugins