README.md

metadata

license: mit
tags:
  - synthetic
  - rl
  - multi-agent
  - cuda
  - triton
  - inference
  - kernel-optimization

kernelswarm — multi-agent kernel-optimization episodes (RL)

Synthetic dataset of multi-agent long-horizon optimization campaigns on AI-inference kernels. Each row is one whole episode: a lead orchestrator plus 2–20 specialist agents split the work, exchange dispatches/statuses/handoffs, submit full kernel candidates, and receive simulated compile/verify/bench feedback — with a reward attached to every environment result.

Inference-focused only. All ops are inference primitives (decode/prefill attention, quantized GEMM/GEMV, sampling, norms, KV-cache, MoE, Mamba).

⚠️ All benchmarks, compiles and correctness checks are simulated analytically — no kernel was ever compiled or executed on a real GPU.

Row format (one row = one episode)

{
  "episode_id": "flash_decode_h100_B=8,..._dispatch_74",
  "scenario": "dispatch",            // dispatch | parallel_sweep | pipeline | campaign
  "n_agents": 6,
  "op": "flash_decode", "impl": "cuda",
  "gpu": "h100", "arch": "sm_90a", "dtype": "fp16",
  "shape": {...}, "baseline_ms": 0.42, "best_ms": 0.006,
  "speedup": 66.4, "elapsed_s": 17335,
  "agents": [{"id": "a01", "role": "worker", "skill": 0.9,
              "tries": 8, "kept": 5, "best_ms": 0.008}, ...],
  "num_messages": 74,
  "messages": [ ... ]
}

Message kinds

kind	role	meaning
plan	lead	campaign split strategy
dispatch	lead	subtask assignment (to, subtask)
candidate	agent	full kernel source + params
result	env	compile/verify/bench outcome + reward
status	agent	terse progress report
review	reviewer	pass/fail verdict
verify	verifier	correctness check report
leaderboard	env	ranked agent best-times
merge	lead	pick winner, update shared best
redirect	lead	re-assign agent to winner's variant/region
handoff	agent	pass artifact to another agent
verdict	lead	final outcome + terminal reward

Rewards

• result msgs: min(1, log2(prev_best / new)) on kept improvement; -0.05 no-improvement, -0.1 compile error, -0.3 wrong result, -0.05 unsupported arch.
• verdict msg: log2(speedup) / 4 terminal bonus.

Scenarios

• dispatch — lead splits variants across workers, merges, redirects losers (2–8 agents)
• parallel_sweep — N workers each own a cfg slice of one variant; leaderboard + top-2 refine (4–20)
• pipeline — proposer → reviewer → verifier → tuner chain over variants (3–6)
• campaign — fleet dispatch over variant×cfg space + verifier rebench (5–20)

Coverage

• 17 ops (14 CUDA + 3 Triton impls): flash_decode, swa_decode, gemm_fp16, int8_gemm, fp8_gemm, int4_gemv, rmsnorm, rope, swiglu_act, topk_sample, nucleus_sample, mamba_scan, moe_scatter, kv_append, fused_qkv_rope, triton_flash_decode, triton_rmsnorm
• GPUs: h100, h200, a100, a10, rtx_a6000, rtx_5090 (arch features gated: wgmma/tma/fp8/fp4/cp.async)
• 2400 episodes · ~170k messages · avg 71 msgs/episode · avg 8.6 agents
• Simulated horizon: 1–5 h per episode

Files

• batches/batch_*.jsonl — 24 shards × 100 episodes
• episodes.jsonl — all episodes concatenated
• tasks_holdout.jsonl — (op, gpu, shape) task specs for eval
• stats.json — aggregate stats