File size: 7,490 Bytes
2e4c7fe | 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 | """kbench — one entry point for the kernel-generation suite.
Before this, the suite was eight loose scripts with different calling conventions (build.py took a spec
path, validate.sh took a GPU then task names, the audits took task names or nothing, publish took
nothing). Same operations, unified:
kbench list [--tier T3] [--family moe] [--metric GB/s] [--gpus 2] [--json]
kbench show <task>
kbench build <spec.py | task-name | --all>
kbench validate <task...> [--gpu N]
kbench audit [--sizes] [--quality] [--gates] [--span] [--schema] [--all] [task...]
kbench catalog regenerate CATALOG.json + README.md
kbench publish [--dry-run] stage the human-readable tree and push to the Hub
kbench stats
Everything delegates to the existing tools; nothing is reimplemented.
"""
import argparse
import json
import pathlib
import subprocess
import sys
HERE = pathlib.Path(__file__).resolve().parent
LANE = HERE.parent
ROOT = LANE.parent
def _catalog():
p = LANE / "CATALOG.json"
return json.loads(p.read_text()) if p.exists() else []
def _run(cmd, **kw):
return subprocess.run(cmd, shell=isinstance(cmd, str), **kw).returncode
def _spec_for(task):
for sub in ("_factory", "_mega_factory", "_dist_factory"):
p = LANE / sub / "specs" / (task.replace("-", "_") + ".py")
if p.exists():
return p
return None
def cmd_list(a):
rows = _catalog()
for f, v in (("tier", a.tier), ("family", a.family), ("metric", a.metric)):
if v:
rows = [r for r in rows if (r.get(f) or "").lower().find(v.lower()) >= 0]
if a.gpus:
rows = [r for r in rows if r.get("gpus") == a.gpus]
if a.json:
print(json.dumps(rows, indent=2)); return 0
for r in rows:
us = f"{r['roofline_us']:.0f}us" if r.get("roofline_us") else "—"
print(f"{r.get('tier','—'):<3} {r.get('metric') or '—':<8} {us:>8} {r['name']}")
print(f"\n{len(rows)} tasks", file=sys.stderr)
return 0
def cmd_show(a):
r = next((x for x in _catalog() if x["name"] == a.task), None)
if not r:
print(f"no such task: {a.task}", file=sys.stderr); return 1
print(f"{r['name']}\n family {r['family']}\n difficulty{'':1}{r.get('tier','—')} "
f"{r.get('tier_why','')}\n metric {r.get('metric')}\n roofline "
f"{r.get('roofline_us')} us\n gpus {r.get('gpus')}")
sp = _spec_for(a.task)
print(f" spec {sp.relative_to(LANE) if sp else '(hand-written, no spec)'}")
print(f" keywords {', '.join(r.get('keywords', []))}")
print(f"\n{r.get('description','')}")
return 0
def cmd_build(a):
if a.all:
n = f = 0
for sub, b in (("_factory", "build.py"), ("_mega_factory", "build.py")):
for s in sorted((LANE / sub / "specs").glob("*.py")):
if s.name.startswith("_"):
continue
ok = _run([sys.executable, str(LANE / sub / b), str(s)],
stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL) == 0
n, f = n + ok, f + (not ok)
if not ok:
print(f" BUILD FAILED {s.name}", file=sys.stderr)
print(f"built {n}, failed {f}")
return 1 if f else 0
for t in a.target:
p = pathlib.Path(t)
if not p.exists():
p = _spec_for(t)
if p is None:
print(f"no spec for {t}", file=sys.stderr); return 1
b = LANE / ("_mega_factory" if "_mega_factory" in str(p) else "_factory") / "build.py"
if _run([sys.executable, str(b), str(p)]) != 0:
return 1
return 0
def cmd_validate(a):
mega = [t for t in a.task if (LANE / t / "tests" / "verify_env.py").exists()
and "canonical_work" not in (LANE / t / "tests" / "verify_env.py").read_text()]
std = [t for t in a.task if t not in mega]
rc = 0
if std:
rc |= _run(["bash", str(HERE / "validate.sh"), str(a.gpu), *std])
if mega:
print("megakernel-family tasks need the mega harness "
"(reference as pass-through submission); see _mega_factory/CALIBRATION.md §7:",
", ".join(mega), file=sys.stderr)
return rc
def cmd_audit(a):
which = [k for k in ("sizes", "quality", "gates", "span", "schema") if getattr(a, k)] or (
["sizes", "quality", "schema"] if not a.all else ["sizes", "quality", "gates", "span", "schema"])
rc = 0
if "sizes" in which:
print("== sizing =="); rc |= _run([sys.executable, str(HERE / "audit_sizes.py"), *a.task])
if "quality" in which:
print("== quality =="); rc |= _run([sys.executable, str(HERE / "audit_quality.py"), *a.task])
if "schema" in which:
print("== schema =="); rc |= _run([sys.executable, str(HERE / "audit_schema.py"), *a.task])
if "gates" in which:
print("== gate discrimination ==")
lst = pathlib.Path("/tmp/_kb_gates.txt")
lst.write_text("\n".join(a.task or [r["name"] for r in _catalog()]) + "\n")
rc |= _run(f"GPU={a.gpu} bash {HERE/'gate_probe.sh'} {lst}")
if "span" in which:
print("== input span == (run per task inside its container; see _factory/span_check.py)")
return rc
def cmd_catalog(a):
return _run([sys.executable, str(HERE / "make_catalog.py")])
def cmd_publish(a):
if _run([sys.executable, str(HERE / "publish_kbench.py")]) != 0:
return 1
if a.dry_run:
print("staged only (--dry-run); not pushed"); return 0
push = ROOT.parent / "scripts" / "push_kbench.py"
print(f"stage complete. push with the Hub uploader ({push if push.exists() else 'see RUN.md'})")
return 0
def cmd_stats(a):
rows = _catalog()
from collections import Counter
for field, label in (("tier", "difficulty"), ("family", "family"), ("metric", "metric")):
c = Counter(r.get(field) or "—" for r in rows)
print(f"\n{label}:")
for k, v in c.most_common():
print(f" {v:5d} {k}")
print(f"\ntotal in kernels/: {len(rows)}")
return 0
def main():
p = argparse.ArgumentParser(prog="kbench", description=__doc__.split("\n")[0])
sub = p.add_subparsers(dest="cmd", required=True)
q = sub.add_parser("list"); q.set_defaults(fn=cmd_list)
q.add_argument("--tier"); q.add_argument("--family"); q.add_argument("--metric")
q.add_argument("--gpus", type=int); q.add_argument("--json", action="store_true")
q = sub.add_parser("show"); q.set_defaults(fn=cmd_show); q.add_argument("task")
q = sub.add_parser("build"); q.set_defaults(fn=cmd_build)
q.add_argument("target", nargs="*"); q.add_argument("--all", action="store_true")
q = sub.add_parser("validate"); q.set_defaults(fn=cmd_validate)
q.add_argument("task", nargs="+"); q.add_argument("--gpu", type=int, default=1)
q = sub.add_parser("audit"); q.set_defaults(fn=cmd_audit)
q.add_argument("task", nargs="*"); q.add_argument("--gpu", type=int, default=1)
for k in ("sizes", "quality", "gates", "span", "schema", "all"):
q.add_argument(f"--{k}", action="store_true")
sub.add_parser("catalog").set_defaults(fn=cmd_catalog)
q = sub.add_parser("publish"); q.set_defaults(fn=cmd_publish)
q.add_argument("--dry-run", action="store_true")
sub.add_parser("stats").set_defaults(fn=cmd_stats)
a = p.parse_args()
sys.exit(a.fn(a))
if __name__ == "__main__":
main()
|