File size: 10,562 Bytes
2e4c7fe 51dbbd3 2e4c7fe 51dbbd3 2e4c7fe 51dbbd3 2e4c7fe 51dbbd3 2e4c7fe 51dbbd3 2e4c7fe 51dbbd3 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | """Emit a schema-conformant record per task, and validate it against task.schema.json.
This is the machine-readable contract for the suite. It is DERIVED from what is on disk -- the grader's
TOL/shape lists, the reference source, the catalog's family and difficulty -- so it cannot drift from
the tasks themselves.
The split that matters: DEFINITION (maths + graded interface + correctness policy) is separate from
WORKLOAD (the shapes). Today every task has one workload, "synthetic", sized so the kernel dominates.
A captured workload -- shapes recorded from a real serving or sampling run -- drops in beside it
without touching the definition or duplicating the task.
python3 _factory/audit_schema.py [task ...] -> writes TASKS.json, reports violations
"""
import ast
import json
import pathlib
import re
import sys
LANE = pathlib.Path(__file__).resolve().parent.parent
SCHEMA = pathlib.Path(__file__).resolve().parent / "schema" / "task.schema.json"
ONLY = set(sys.argv[1:])
COMPARATOR = {"tuple": "relative-frobenius-tuple-max", "rowwise": "rowwise",
"tensor": "relative-frobenius"}
def const(src, name, default=None):
"""Module-level constant by AST. A single-line regex misses the many multi-line shape lists."""
try:
tree = ast.parse(src)
except SyntaxError:
return default
for n in tree.body:
if isinstance(n, ast.Assign):
# tuple unpacking: BATCH, PREFILL, MAX_SEQ = 1, 2048, 4096
for t in n.targets:
if isinstance(t, ast.Tuple) and isinstance(n.value, ast.Tuple):
for tgt, val in zip(t.elts, n.value.elts):
if isinstance(tgt, ast.Name) and tgt.id == name:
try:
return ast.literal_eval(val)
except Exception:
pass
for t in n.targets:
if isinstance(t, ast.Name) and t.id == name:
try:
return ast.literal_eval(n.value)
except Exception:
pass
try: # non-literal (comprehension, arithmetic): evaluate it
ns = {}
exec(compile(ast.Module([n], []), "<c>", "exec"), ns)
return ns.get(name, default)
except Exception:
return default
return default
def emit(task):
d = LANE / task
v = d / "tests" / "verify_env.py"
if not v.exists():
return None
src = v.read_text()
metric = "GB/s" if "GB/s" in src else "TFLOP/s"
tol = const(src, "TOL")
if tol is None:
tol = const(src, "PERF_TOL")
comp = "relative-frobenius"
if const(src, "NAMES") is not None:
comp = "relative-frobenius-tuple-max"
if const(src, "ROW_PASS") is not None:
comp = "rowwise"
if tol == 0:
comp = "exact"
# three harness layouts: factory (fn = m.x), legacy (return m.x inside _load), megakernel
# (m.<entry_build> / m.<entry_step>)
entry = (re.search(r"fn = m\.(\w+)", src) or re.search(r"return m\.(\w+)", src)
or re.search(r"hs = m\.(\w+)", src) or re.search(r"got = m\.(\w+)", src)
or re.search(r'getattr\(m, "(\w+)"', src))
if entry is None: # dist-* graders resolve the entry via their spec
for sub in ("_dist_factory", "_factory", "_mega_factory"):
sp = LANE / sub / "specs" / (task.replace("-", "_") + ".py")
if sp.exists():
fm = re.search(r'func\s*=\s*"(\w+)"', sp.read_text())
if fm:
entry = fm
break
dims = []
graded = const(src, "GRADER_SHAPES", []) or []
correct = const(src, "CORRECT_SHAPES", []) or []
measure = const(src, "MEASURE_SHAPES", []) or []
# the megakernel family has no shape tuples: its workload is an architecture config plus a
# decode regime, so express that as a single graded point rather than pretending it has none.
if not graded:
cfg = const(src, "CFG")
if isinstance(cfg, dict):
dims_mk = ["layers", "d", "n_q", "n_kv", "hd", "vocab"]
pt = [cfg.get(k, 0) for k in dims_mk] + [const(src, "BATCH", 1),
const(src, "PREFILL", 0),
const(src, "DECODE_STEPS", 0)]
dims = dims_mk + ["batch", "prefill", "decode_steps"]
graded = correct = [pt]
def _params(fn_src):
"""Positional dim names of a function: drop defaulted params (chunk_size=64) and `seed`."""
out = []
for raw in fn_src.split(","):
nm = raw.split("=")[0].strip()
if not nm or "=" in raw or nm in ("seed", "self") or nm.startswith("*"):
continue
out.append(nm)
return out
cw = ""
m = re.search(r"^def canonical_work\(([^)]*)\)", src, re.M)
if m:
dims = _params(m.group(1))
if not dims:
# legacy graders name the work function differently; the input generator's signature is the
# authoritative list of dims either way
mk = re.search(r"^def _(?:mk|make)\(([^)]*)\)", src, re.M)
if mk:
dims = _params(mk.group(1))
if m:
r = re.search(r"return (.+)", src[m.start():])
cw = r.group(1).strip() if r else ""
# Most authoritative of all: how the grader itself unpacks a shape tuple. A legacy shape can carry
# config the input generator does not name as a parameter (flash-attn-backward's causal flag).
if graded:
want = len(graded[0])
if len(dims) != want:
for pat in (r"for\s+\w+,\s*\(([^)]+)\)\s+in\s+enumerate\(\s*(?:GRADER|CORRECT)_SHAPES",
r"for\s+\(([^)]+)\)\s+in\s+(?:GRADER|CORRECT)_SHAPES",
r"\(([^)]+)\)\s*=\s*shp\b"):
u = re.search(pat, src)
if u:
cand = [x.strip() for x in u.group(1).split(",") if x.strip()]
if len(cand) == want:
dims = cand
break
rec = {
"name": task,
"family": CAT.get(task, {}).get("family", "Other"),
"description": CAT.get(task, {}).get("description", ""),
"keywords": CAT.get(task, {}).get("keywords", []),
"definition": {
"entry_point": entry.group(1) if entry else "",
"module": (re.search(r"MODULE_PATH = \"/app/([^\"]+)\"", src) or [None, ""])[1]
if "MODULE_PATH" in src else "",
"inputs": [], "outputs": [],
"reference": {"language": "python/pytorch",
"source": f"{task}/environment/reference.py"},
},
"workloads": {
"synthetic": {
"source": "synthetic",
"dims": dims,
"graded": [list(s) for s in graded],
"correctness": [list(s) for s in correct],
"measure": [list(s) for s in measure],
**({"roofline_us": CAT[task]["roofline_us"]}
if CAT.get(task, {}).get("roofline_us") is not None else {}),
}
},
"grading": {
"metric": metric,
"reward": {"kind": "absolute-uncapped", "canonical_work": cw},
"tolerance": {"value": tol if tol is not None else 0.0, "comparator": comp},
},
"environment": {
"gpus": CAT.get(task, {}).get("gpus", 1),
"min_compute_capability": "9.0",
"offline": True,
"note": "The exact GPU is deliberately not specified; the task tells the agent to query it.",
},
}
t = DIFF.get(task)
if t:
rec["difficulty"] = {"tier": t[0], "rationale": t[1],
"reviewed_by_hand": task in HANDREVIEWED}
rp = const(src, "ROW_PASS")
if rp is not None:
rec["grading"]["tolerance"]["row_pass"] = rp
return rec
CAT = {}
p = LANE / "CATALOG.json"
if p.exists():
CAT = {r["name"]: r for r in json.loads(p.read_text())}
DIFF, HANDREVIEWED = {}, set()
p = LANE / "_factory" / "difficulty.json"
if p.exists():
for r in json.loads(p.read_text()):
DIFF[r["name"]] = (r["tier"], r["why"])
try:
sys.path.insert(0, str(LANE / "_factory"))
from difficulty import OVERRIDE
HANDREVIEWED = set(OVERRIDE)
except Exception:
pass
def main():
tasks = sorted(x.name for x in LANE.iterdir()
if x.is_dir() and not x.name.startswith("_") and (x / "task.toml").exists())
if ONLY:
tasks = [t for t in tasks if t in ONLY]
recs, bad = [], []
for t in tasks:
r = emit(t)
if r is None:
bad.append((t, "no grader")); continue
for req in ("name", "family", "definition", "workloads", "grading"):
if req not in r:
bad.append((t, f"missing {req}"))
if not r["definition"]["entry_point"]:
bad.append((t, "no entry point discoverable"))
w = r["workloads"]["synthetic"]
if not w["graded"] or not w["correctness"]:
bad.append((t, "workload has no graded/correctness shapes"))
elif w["dims"] and any(len(sh) != len(w["dims"]) for sh in w["graded"]):
n = next(len(sh) for sh in w["graded"] if len(sh) != len(w["dims"]))
bad.append((t, f"{len(w['dims'])} dim names for a {n}-value shape"))
elif not w["dims"]:
bad.append((t, "workload shapes have no dim names"))
recs.append(r)
(LANE / "TASKS.json").write_text(json.dumps(recs, indent=2) + "\n")
try:
import jsonschema
sch = json.loads(SCHEMA.read_text())
for r in recs:
try:
jsonschema.validate(r, sch)
except jsonschema.ValidationError as e:
bad.append((r["name"], f"schema: {e.message[:90]}"))
note = "validated against task.schema.json"
except ImportError:
note = "jsonschema not installed; structural checks only"
print(f"emitted {len(recs)} task records -> TASKS.json ({note})")
if bad:
print(f" {len(bad)} issues:")
for t, m in bad[:15]:
print(f" {t}: {m}")
else:
print(" no issues")
return 1 if bad else 0
if __name__ == "__main__":
sys.exit(main())
|