File size: 12,593 Bytes
0f775e2 4161512 0f775e2 4161512 0f775e2 4161512 0f775e2 4161512 0f775e2 4161512 0f775e2 2e4c7fe 0f775e2 2e4c7fe 0f775e2 2e4c7fe 0f775e2 2e4c7fe 0f775e2 2e4c7fe 0f775e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 | """Generate CATALOG.json + README.md for the kernels/ suite.
The task directories stay FLAT on purpose: _factory/build.py, _factory/validate.sh and
_factory/audit_sizes.py all address tasks as LANE/<name>, and harbor is pointed at a task path
directly. Nesting them would break the build chain for no functional gain. Navigability comes from
this catalog instead.
Everything here is derived from what is already on disk -- task.toml keywords, the grader's metric and
GRADER_SHAPES, and canonical_work -- so it cannot drift from the tasks themselves. Re-run after adding
tasks: python3 _factory/make_catalog.py
"""
import ast
import json
import pathlib
import re
LANE = pathlib.Path(__file__).resolve().parent.parent
H200_TFLOPS, H200_GBPS, LINK_GBPS = 700.0, 4800.0, 50.0
# Ordered: first match wins. Keyed on task.toml keywords, falling back to the name.
FAMILIES = [
# Includes the E2 "enabling primitives": persistent-kernel building blocks that are part of the
# megakernel family but do not carry the megakernel- prefix.
("Megakernel — whole-model fusion & primitives", lambda n, k: n.startswith("megakernel-") or
"megakernel" in k or "persistent-kernel" in k or re.match(
r"(gridwide-barrier|instruction-interpreter|async-weight-prefetch|warp-specialized|"
r"persistent-|cross-layer-fusion)", n)),
("Distributed — multi-GPU collectives", lambda n, k: n.startswith("dist-")),
# NOTE: match these on the NAME, before the model-specific families. Keyword matching on
# "wan"/"hunyuanvideo" is far too greedy -- VAE and sparse-attention tasks tag those keywords
# merely because they use Wan/HunyuanVideo shapes, and get mis-filed into the DiT families.
("Video — 3D causal VAE / tokenizer", lambda n, k: re.match(
r"(causal-conv3d|conv3d-|groupnorm3d|temporal-(up|down)sample|spatial-upsample|trilinear|"
r"wavelet|video-latent|latent-normalize|depthwise-separable-conv3d|vae-)", n)),
("Video — sparse / efficient attention", lambda n, k: re.match(
r"(sliding-tile|sta-|radial-|svg-|block-sparse-video|video-attn|frame-anchor|"
r"temporal-strided|spatial-window|hierarchical-coarse|causal-video|sparse-attn-mask|"
r"latent-patchify|tile-permute|adaptive-sparsity|compressed-kv-video|rolling-window-video|"
r"video-cfg-zero-star|"
r"attn-density|sparse-block-worklist|pyramid-kv)", n)),
("Image generation — FLUX / SD3 MMDiT", lambda n, k: re.match(r"(flux-|sd3-)", n)),
("Video — CogVideoX / Mochi / LTX", lambda n, k: re.match(r"(cogvideox-|mochi-|ltx-)", n)),
("Video — Wan DiT", lambda n, k: n.startswith("wan-")),
("Video — HunyuanVideo MMDiT", lambda n, k: n.startswith("hunyuan-")),
("Diffusion — sampling, scheduling, caching", lambda n, k: "diffusion" in k and "video-diffusion" not in k
or re.match(r"(flow-match|dpmsolver|ddim|teacache|feature-cache|cache-hit|residual-diff|"
r"cfg-|noise-add|latent-blend|latent-interp|scheduler-|step-distill|sigma-)", n)),
("Multimodal & audio", lambda n, k: re.match(
r"(vision-patch|clip-|mrope|image-token|any-res|mm-embed|audio-|whisper|siglip|dinov2|vit-|tts-|"
r"qwen-vl|conformer-|hifigan-|istft-|snake-antialias|mel-)", n)),
# Each family carries a NAME fallback as well as keywords: several tasks predate the keyword
# vocabulary and would otherwise land in "Other".
("Linear attention & SSM", lambda n, k: "linear-attention" in k or "delta-rule" in k or "ssm" in k
or "mamba" in k or re.match(
r"(mamba|titans|ttt-|rwkv|gla-|gsa-|retention|comba|kda-|delta|hgrn|based-|lightning|"
r"log-linear|mesa|path-attn|simple-gla|hybrid-layer)", n)),
("MoE — routing & grouped GEMM", lambda n, k: "moe" in k or n.startswith("moe-")),
("Quantization & low-precision GEMM", lambda n, k: "quantization" in k or "low-precision" in k
or "gemm" in k or "fp8" in k),
("Training, optimizer & RL", lambda n, k: "training" in k or "optimizer" in k or "rl" in k
or re.match(r"(dpo-|grpo-|gae-|ppo-|muon|sequence-packing|gradient-accum|lora|dora|qlora|"
r"quantized-optimizer|fused-adamw|grad-global|distill-kl|entropy-bonus|"
r"adafactor|mtp-multi-head|activation-recompute|flow-match-loss)", n)),
("Sampling & speculative decoding", lambda n, k: "sampling" in k or "speculative" in k
or "decoding" in k or re.match(
r"(draft-tree|ngram|spec-decode|beam-search|min-p|repetition|guided-decoding|fused-topk|"
r"logits-gather)", n)),
("KV cache & paging", lambda n, k: "kv-cache" in k or "paged" in k or re.match(
r"(radix-prefix|sequence-unpad|paged-kv|kv-cache|int4-kv|prefix-cache|kv-layout|kv-block|"
r"kv-repage|grammar-jump|multi-lora|penalty-count|speculative-draft)", n)),
("Normalization, RoPE & elementwise fusion", lambda n, k: re.match(
r"(fused-residual-rmsnorm|fused-rmsnorm|rmsnorm-|layernorm-|rope-|yarn-|fused-qk-norm|"
r"dyt-|layerscale-|sandwich-norm|"
r"swiglu|attention-qk-norm|embedding-backward)", n)),
("Attention — text LLM", lambda n, k: "attention" in k or re.match(
r"(dsa-|moba|nsa-|mla-|chunked-prefill|prefix-lm|cascade|streaming|softcap|alibi|"
r"deepseek-|qwen3-next-gated|altup-|laurel-|"
r"diff-attention|flex-|forgetting|qk-clip|cross-attention|varlen|gqa-|attention-)", n)),
("Other", lambda n, k: True),
]
def parse_toml(p):
txt = p.read_text()
def grab(field, default=""):
m = re.search(rf'^{field}\s*=\s*"(.*)"\s*$', txt, re.M)
return m.group(1) if m else default
kws = []
m = re.search(r"^keywords\s*=\s*\[(.*?)\]", txt, re.M | re.S)
if m:
kws = [x.strip().strip('"') for x in m.group(1).split(",") if x.strip()]
g = re.search(r"^gpus\s*=\s*(\d+)", txt, re.M)
return grab("name", p.parent.name), grab("description"), kws, int(g.group(1)) if g else 1
def roofline(task_dir, name):
"""Replicates audit_sizes.py so the catalog cannot disagree with the size audit."""
v = task_dir / "tests" / "verify_env.py"
if not v.exists():
return None, None
src = v.read_text()
if "canonical_work" not in src:
return "tokens/s", None # megakernel family: sized by MegaSpec.floor_us()
ns, shapes = {}, None
try:
for node in ast.parse(src).body:
if isinstance(node, (ast.FunctionDef, ast.Assign)):
try:
exec(compile(ast.Module([node], []), "<c>", "exec"), ns)
except Exception:
pass
if isinstance(node, ast.Assign) and getattr(node.targets[0], "id", "") == "GRADER_SHAPES":
shapes = ast.literal_eval(node.value)
metric = "GB/s" if "GB/s" in src else "TFLOP/s"
big = max(ns["canonical_work"](*s) for s in shapes)
if metric == "TFLOP/s":
us = big / (H200_TFLOPS * 1e12) * 1e6
else:
bw = LINK_GBPS if name.startswith("dist-") else H200_GBPS
us = big / (bw * 2 ** 30) * 1e6
return metric, round(us, 1)
except Exception:
return None, None
def _slug(s):
"""GitHub heading anchor: lowercase, drop punctuation, each space -> one hyphen."""
s = re.sub(r"[^\w\s-]", "", s.lower())
return re.sub(r"\s", "-", s.strip())
def family_of(name, kws):
for fam, pred in FAMILIES:
try:
if pred(name, kws):
return fam
except Exception:
pass
return "Other"
def main():
diff = {}
dp = LANE / "_factory" / "difficulty.json"
if dp.exists():
for r in json.loads(dp.read_text()):
diff[r["name"]] = (r["tier"], r["why"])
rows = []
for d in sorted(p for p in LANE.iterdir() if p.is_dir() and not p.name.startswith("_")):
t = d / "task.toml"
if not t.exists():
continue
name, desc, kws, gpus = parse_toml(t)
metric, us = roofline(d, d.name)
tier, why = diff.get(d.name, ("", ""))
rows.append(dict(name=d.name, family=family_of(d.name, kws), tier=tier, tier_why=why,
metric=metric, roofline_us=us, gpus=gpus, keywords=kws, description=desc))
(LANE / "CATALOG.json").write_text(json.dumps(rows, indent=2) + "\n")
order = [f for f, _ in FAMILIES]
by_fam = {}
for r in rows:
by_fam.setdefault(r["family"], []).append(r)
out = [
"# Kernel-generation suite",
"",
f"**{len(rows)} tasks.** Each gives the agent a correct-but-slow reference and an empty stub; the",
"agent writes a fast GPU kernel.",
"",
" reward = 0 if the submission is incorrect",
" reward = achieved TFLOP/s or GB/s otherwise, UNCAPPED",
"",
"Correctness is the gate, speed is the reward. There is no oracle and no gold solution — the score",
"is an absolute hardware metric, so it is hardware-portable and nothing needs re-benchmarking.",
"",
"Task directories are deliberately **flat**: the factory (`_factory/build.py`), the validator",
"(`_factory/validate.sh`) and the size audit (`_factory/audit_sizes.py`) all address tasks as",
"`kernels/<name>`, and harbor is pointed at a task path directly. This catalog provides the",
"structure instead, and is generated from what is on disk (`python3 _factory/make_catalog.py`),",
"so it cannot drift.",
"",
"`roofline` is the implied runtime of a perfect kernel at the largest graded shape (700 TFLOP/s",
"bf16 / 4.8 TB/s HBM; multi-GPU tasks are bounded by the 50 GB/s interconnect instead). Every task",
"is above 250 us, so the kernel dominates rather than launch overhead.",
"",
"## Difficulty",
"",
"Tasks are tiered by **how hard it is to improve on the best available implementation**, not by",
"how complex they look. A dense GEMM is a one-liner and near-unbeatable; a five-pass elementwise",
"chain is trivial to describe and has most of its performance still on the table.",
"",
"| tier | what it takes to win | typical headroom |",
"|---|---|---|",
"| **T1** | fusing several passes into one; coalesced/vectorised access, intermediates in registers | large |",
"| **T2** | shared-memory tiling, warp reductions, an online single-pass reformulation, layout/swizzle changes | moderate |",
"| **T3** | async copy (cp.async/TMA), double buffering, warp specialisation, hand-written MMA with correct fragment layouts | real but only reachable this way |",
"| **T4** | out-engineering a vendor kernel that is already at the hardware limit | very little |",
"",
"## Contents",
"",
]
for fam in order:
if fam in by_fam:
anchor = _slug(fam)
out.append(f"- [{fam}](#{anchor}) — {len(by_fam[fam])}")
out.append("")
for fam in order:
if fam not in by_fam:
continue
out += [f"## {fam}", "", "| task | diff | metric | roofline | what the kernel does |",
"|---|---|---|---|---|"]
for r in sorted(by_fam[fam], key=lambda x: x["name"]):
us = f"{r['roofline_us']:.0f} us" if r["roofline_us"] else "—"
d = (r["description"] or "").replace("|", "\\|")
d = (d[:150] + "…") if len(d) > 150 else d
g = " **2-GPU**" if r["gpus"] > 1 else ""
out += [f"| `{r['name']}`{g} | {r.get('tier') or '—'} | {r['metric'] or '—'} | {us} | {d} |"]
out.append("")
out += [
"## Generators",
"",
"| dir | what it builds |",
"|---|---|",
"| `_factory/` | the standard single-function tasks. `AGENT_GUIDE.md` is the contract for adding one. |",
"| `_mega_factory/` | the megakernel family (stateful multi-step, measured fusion gates). `CALIBRATION.md` records every measured design decision. |",
"| `_dist_factory/` | the 2-GPU `dist-*` tasks. |",
"| `_parked/` | tasks that do not yet validate. **Excluded from the suite** — never shipped. |",
"",
]
(LANE / "README.md").write_text("\n".join(out))
print(f"CATALOG.json + README.md written: {len(rows)} tasks in {len(by_fam)} families")
for fam in order:
if fam in by_fam:
print(f" {len(by_fam[fam]):4d} {fam}")
if __name__ == "__main__":
main()
|