File size: 22,756 Bytes
e1acc15 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 | """D2 G1 plumbing and amendment-7 K2 committed-trajectory proxy.
Default binding: dev position-0 match rate >= half the hash-bound live
second-pass position-0 acceptance (amendment 7b gross-error tripwire). Historical acap/1.82 mode is explicit and cannot admit
pilot training. See G1-LIVE.md for coordinator-only execution instructions.
"""
from __future__ import annotations
import argparse
import importlib
import collections
import json
import os
import random
import sys
import shlex
import time
from pathlib import Path
from decimal import Decimal
from typing import Any, Callable, TypedDict
from .live_reference import build_reference, verify_reference, file_hash, Reference
import torch
from . import shards as S
K2_REF = 1.82
TOL = 0.1
BINDING_SPLIT = "dev" # Amendment 7; other splits are information
# cards this lane may use (each under its own flock ~/dsv41-offload/locks/gpu-<8 hex>.lock, <= 14 min per hold);
# Historical G1 allocation; coordinator must recheck availability. Never c4e7a371 or 240876ad.
GPU_ALLOWED = ("GPU-bc62df2a-59cf-9828-7d43-558d1355bab7", "GPU-7744b44b-8a13-b06d-a67e-f958fd74b8fc")
ROOT = "/opt/dsv41/x86host-data/dspark"
# ------------------------------------------------------------------------------------------------------- plumbing
def cmd_plumbing(a: argparse.Namespace) -> None:
from .dataset import cmd_verify
cmd_verify(argparse.Namespace(out=a.shards))
with open(os.path.join(a.shards, "MANIFEST.json")) as fh:
man = json.load(fh)
rep: dict[str, Any] = {}
for split, sp in man["splits"].items():
cs = S.ShardCaptureSet([os.path.join(a.shards, f["file"]) for f in sp["files"]], stride=a.stride)
rep[split] = {"sequences": len(cs.rows), "samples": len(cs.samples), "plumbing_errors": len(cs.plumbing),
"first": cs.plumbing[:5]}
rep["pass"] = all(v["plumbing_errors"] == 0 for v in rep.values() if isinstance(v, dict))
print(json.dumps(rep, indent=1))
if not rep["pass"]:
raise SystemExit(1)
# ------------------------------------------------------------------------------------------------------ K2 drafts
class SeqSource:
"""What the K2 runner needs of one sequence: tokens, the first position with taps, a row reader, the blocks."""
def __init__(self, seq_id: str, tokens: torch.Tensor, first_row: int, rows: Callable[[int, int], torch.Tensor], blocks: list[int],
anchors: dict[int, int] | None = None):
self.seq_id, self.tokens, self.first_row, self.rows, self.blocks = seq_id, tokens, first_row, rows, blocks
self.anchors = anchors or {} # explicit anchor tokens (captures: the last anchor is past the rows)
def anchor(self, P: int) -> int:
a = self.anchors.get(P)
return int(self.tokens[P + 1]) if a is None else a
def check_gpu() -> torch.device:
vis = os.environ.get("CUDA_VISIBLE_DEVICES", "")
if vis not in GPU_ALLOWED:
raise SystemExit(f"CUDA_VISIBLE_DEVICES must be exactly one of {GPU_ALLOWED} (got {vis!r})")
if torch.cuda.device_count() != 1:
raise SystemExit(f"{torch.cuda.device_count()} visible devices, want 1")
uuid = str(getattr(torch.cuda.get_device_properties(0), "uuid", ""))
if uuid and uuid not in vis:
raise SystemExit(f"visible device uuid {uuid} is not {vis}")
return torch.device("cuda", 0)
def load_k2(root: str, dev: torch.device, slots: int) -> Any:
dspark = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if dspark not in sys.path:
sys.path.insert(0, dspark)
Checkpoint = importlib.import_module("worker.checkpoint").Checkpoint
DSparkDrafter = importlib.import_module("worker.model").DSparkDrafter
ck = Checkpoint(root)
m = DSparkDrafter(ck, dev, slots=slots)
if slots * 5 > m.max_rows:
raise SystemExit(f"{slots} slots x 5 drafts = {slots * 5} rows exceed the drafter's {m.max_rows}-row scratch")
m.capture(batch_sizes=tuple(range(1, slots + 1)), ks=(5,), insert_rows=(8,), n_outs=(5,), margins=False)
return m
def run_k2(m: Any, seqs: list[SeqSource], slots: int, deadline: float,
on_done: Callable[[str, dict[int, list[int]]], None]) -> int:
"""Draft every block of every sequence, `slots` sequences in lock step; on_done(seq_id, {P: drafts}) per finished
sequence. Returns the number of sequences finished before the deadline (unfinished ones are not reported)."""
dev = m.dev
W = m.s.window
queue = list(seqs)
active: dict[int, dict[str, Any]] = {}
finished = 0
while queue or active:
if time.time() > deadline:
break
for sl in range(slots):
if sl not in active and queue:
s = queue.pop(0)
m.reset_slots(torch.tensor([sl], device=dev))
active[sl] = {"s": s, "i": 0, "next": s.first_row, "out": {}}
if not active:
break
ins_slot, ins_pos, ins_taps = [], [], []
order = sorted(active)
for sl in order:
st = active[sl]
P = st["s"].blocks[st["i"]]
lo = max(st["next"], P - W + 1, st["s"].first_row)
if lo <= P:
ins_slot.append(torch.full((P + 1 - lo,), sl, dtype=torch.int64))
ins_pos.append(torch.arange(lo, P + 1, dtype=torch.int64))
ins_taps.append(st["s"].rows(lo, P + 1))
st["next"] = P + 1
if ins_slot:
taps = torch.cat(ins_taps).to(dev)
m.insert(torch.cat(ins_slot).to(dev), torch.cat(ins_pos).to(dev), m.main_x(taps))
Ps = [active[sl]["s"].blocks[active[sl]["i"]] for sl in order]
anchors = [active[sl]["s"].anchor(P) for sl, P in zip(order, Ps)]
ids, _ = m.draft_graphed(torch.tensor(order, device=dev), torch.tensor(anchors, device=dev),
torch.tensor(Ps, device=dev), 5)
ids = ids[: len(order)].cpu().tolist()
for sl, P, d in zip(order, Ps, ids):
st = active[sl]
st["out"][P] = [int(x) for x in d]
st["i"] += 1
if st["i"] == len(st["s"].blocks):
on_done(st["s"].seq_id, st["out"])
finished += 1
del active[sl]
return finished
def cmd_k2(a: argparse.Namespace) -> None:
dev = check_gpu()
with open(os.path.join(a.shards, "MANIFEST.json")) as fh:
man = json.load(fh)
os.makedirs(a.work, mode=0o700, exist_ok=True)
out_path = os.path.join(a.work, f"{a.split}.k2_drafts.jsonl")
done = set()
if os.path.exists(out_path):
with open(out_path) as fh:
done = {json.loads(line)["seq_id"] for line in fh}
seqs = []
for f in man["splits"][a.split]["files"]:
sh = S.Shard(os.path.join(a.shards, f["file"]))
for i in range(len(sh)):
s = sh.seq(i)
if s["seq_id"] in done:
continue
seqs.append(SeqSource(s["seq_id"], sh.tokens(i), s["r0"], _rows_of(sh, i), list(sh.blocks(i))))
print(f"{len(seqs)} sequences to draft ({len(done)} done) in {a.split}", flush=True)
if not seqs:
return
seqs.sort(key=lambda s: -len(s.blocks))
t0 = time.time()
deadline = t0 + a.deadline_s
m = load_k2(a.root, dev, a.slots)
print(f"K2 drafter loaded + graphs in {time.time() - t0:.0f} s", flush=True)
fh = open(out_path, "a")
os.chmod(out_path, 0o600)
def on_done(sid: str, out: dict[int, list[int]]) -> None:
fh.write(json.dumps({"seq_id": sid, "drafts": {str(p): d for p, d in out.items()}}) + "\n")
fh.flush()
n = run_k2(m, seqs, a.slots, deadline, on_done)
print(f"{n} of {len(seqs)} sequences drafted in {time.time() - t0:.0f} s"
+ ("" if n == len(seqs) else " (deadline: run again in the next hold to continue)"), flush=True)
def _rows_of(sh: S.Shard, i: int) -> Callable[[int, int], torch.Tensor]:
return lambda p0, p1: sh.taps(i, p0, p1)
# ----------------------------------------------------------------------------------------------------------- score
def match_len(drafts: list[int], truth: list[int | None]) -> int | None:
n = 0
for d, t in zip(drafts, truth):
if t is None:
return None
if d != t:
return n
n += 1
return n
def cluster_ci(vals_by_seq: dict[str, list[int]], iters: int = 2000, seed: int = 0) -> list[float]:
rng = random.Random(seed)
keys = list(vals_by_seq)
boots = []
for _ in range(iters):
pick = [vals_by_seq[keys[rng.randrange(len(keys))]] for _ in keys]
tot = sum(len(p) for p in pick)
boots.append(sum(sum(p) for p in pick) / max(tot, 1))
boots.sort()
return [round(boots[int(0.025 * iters)], 4), round(boots[int(0.975 * iters)], 4)]
def bucket(P: int) -> str:
return "<2K" if P < 2000 else "2-8K" if P < 8000 else "8-16K" if P < 16000 else ">=16K"
class Sequence(TypedDict):
tokens: list[int]
gen_start: int
L: int
family: str
def score(seqs: dict[str, Sequence], drafts: dict[str, dict[int, list[int]]], split: str = BINDING_SPLIT, *, live_reference: Reference | None = None,
mode: str = "live-reference") -> dict[str, Any]:
"""seqs: seq_id -> {tokens, gen_start, L, family}; drafts: seq_id -> {P: [5]}. G1(b) binds only on BINDING_SPLIT."""
rounds_by_seq, allpos_by_seq = {}, {}
by_fam, by_bucket = collections.defaultdict(list), collections.defaultdict(list)
cond_n, cond_hit = [0] * 5, [0] * 5
for sid, s in seqs.items():
d = drafts.get(sid)
if d is None:
raise ValueError(f"{sid}: no drafts")
tok, g, L = s["tokens"], s["gen_start"], s["L"]
def realized(P: int) -> list[int | None]:
return [int(tok[t]) if t < L else None for t in range(P + 2, P + 7)]
allpos = [x for x in (match_len(d[P], realized(P)) for P in range(g - 1, L - 2)) if x is not None]
rounds = []
P = g - 1
while P <= L - 3:
ml = match_len(d[P], realized(P))
if ml is None:
break
rounds.append(ml)
for i in range(5):
if ml >= i:
cond_n[i] += 1
cond_hit[i] += int(ml >= i + 1)
by_fam[s["family"]].append(ml)
by_bucket[bucket(P)].append(ml)
P += ml + 1
rounds_by_seq[sid] = rounds
allpos_by_seq[sid] = allpos
rr = [x for v in rounds_by_seq.values() for x in v]
aa = [x for v in allpos_by_seq.values() for x in v]
if not rr or not aa:
raise ValueError("no scoreable committed-trajectory rounds/positions")
mean = sum(rr) / len(rr)
return {"sequences": len(seqs), "rounds": len(rr), "k2_accepted_per_round": round(mean, 4),
"ci95_request_cluster": cluster_ci(rounds_by_seq),
"k2_accepted_all_positions": round(sum(aa) / len(aa), 4), "positions": len(aa),
"position0_match_rate": sum(x > 0 for x in rr) / len(rr),
"per_position_match_rate": [sum(x > i for x in rr) / len(rr) for i in range(5)],
"live_comparison": live_comparison(live_reference, split) if live_reference else None,
"conditional_acceptance": [round(h / n, 4) if n else None for h, n in zip(cond_hit, cond_n)],
"by_family": {k: {"rounds": len(v), "mean": round(sum(v) / len(v), 4)} for k, v in sorted(by_fam.items())},
"by_context": {k: {"rounds": len(v), "mean": round(sum(v) / len(v), 4)} for k, v in sorted(by_bucket.items())},
"per_sequence_mean_range": [round(min(sum(v) / len(v) for v in rounds_by_seq.values() if v), 4),
round(max(sum(v) / len(v) for v in rounds_by_seq.values() if v), 4)],
"g1": g1_verdict(sum(x > 0 for x in rr) / len(rr) if mode == "live-reference" else mean,
split, live_reference=live_reference, mode=mode)}
def g1_verdict(value: float, split: str, *, live_reference: Reference | None = None,
mode: str = "live-reference") -> dict[str, object]:
if mode == "historical-acap":
reference, binding_split = K2_REF, "acap"
valid = True
elif mode == "live-reference":
if live_reference is None:
raise ValueError("amendment-7 live-reference receipt required")
reference, binding_split = live_reference["position0_acceptance"], BINDING_SPLIT
valid = live_reference["status"] == "VALID"
else:
raise ValueError("unknown G1 mode")
binding = split == binding_split
if mode == "live-reference":
threshold = Decimal(str(reference)) * Decimal("0.5")
passed = Decimal(str(value)) >= threshold
rule: dict[str, object] = {"minimum_live_fraction": 0.5, "minimum_proxy_rate": float(threshold),
"tripwire_passed": passed}
else:
passed = abs(Decimal(str(value))-Decimal(str(reference))) <= Decimal(str(TOL))
rule = {"tolerance": TOL, "within_tolerance": passed}
return {"mode": mode, "amendment": "7b" if mode == "live-reference" else None,
"statistic": "position0_acceptance" if mode == "live-reference" else "accepted_per_round",
"reference": reference, "split": split, "binding": binding, **rule,
"pass": passed if binding and valid else None,
"status": ("PASS" if passed else "FAIL") if binding and valid else "INCONCLUSIVE" if not valid else "INFORMATION",
"live_reference_sha256": live_reference["sha256"] if live_reference else None}
def live_comparison(live: Reference, split: str) -> dict[str, object]:
return {"position0_acceptance": live["position0_acceptance"],
"accepted_per_round": live["accepted_per_round"],
"per_position_acceptance": live["per_position_acceptance"],
"foreign_share": live["foreign_share"], "status": live["status"],
"biases": {
"depth_0_censoring": "Live conf policy can verify zero drafts; proxy position 0 is uncensored.",
"window": "Proxy fixed window 5 and live adaptive windows select different round starts.",
"split_coverage": f"Proxy covers {split}; live counters cover the whole second pass.",
"foreign_share": "Live counters include foreign traffic; VALID requires foreign share <= 2%.",
}}
def coordinator_procedure(shards: Path, work: Path, live: Path, gpu: str = GPU_ALLOWED[0]) -> dict[str, object]:
"""Render only. No filesystem reads, CUDA, subprocesses or host access."""
if gpu not in GPU_ALLOWED:
raise ValueError("G1 GPU must be in GPU_ALLOWED")
lock = f"$HOME/dsv41-offload/locks/gpu-{gpu[4:12]}.lock"
python = "$HOME/dsv41-offload/o2/venv/bin/python"
setup = "set -euo pipefail\numask 077\nsource ~/.config/ndm/paths.sh\nexport PYTHONPATH=\"$PWD/offload/dspark:$PWD\"\n"
holds: dict[str, str] = {}
scores: dict[str, str] = {}
for split in (BINDING_SPLIT, "heldout_legal", "heldout_general"):
args = shlex.join(["-m", "d2.g1", "k2", "--shards", str(shards), "--split", split,
"--work", str(work), "--deadline-s", "780"])
holds[split] = setup + (f"CUDA_VISIBLE_DEVICES={shlex.quote(gpu)} flock -n \"{lock}\" "
f"timeout --signal=TERM --kill-after=20 820 \"{python}\" {args}")
scores[split] = setup + "CUDA_VISIBLE_DEVICES='' \""+python+'" '+shlex.join([
"-m", "d2.g1", "score", "--shards", str(shards), "--split", split,
"--work", str(work), "--live-reference", str(live), "--out", str(work/f"{split}.json")])
return {"dry": True, "gpu": gpu, "lock": lock, "deadline_s": 780, "hold_limit_s": 840, "term_timeout_s": 820,
"holds": holds, "scores": scores, "binding_split": BINDING_SPLIT,
"outputs": [str(work/f"{s}.{ext}") for s in holds for ext in ("k2_drafts.jsonl", "json")]}
def cmd_procedure(a: argparse.Namespace) -> None:
print(json.dumps(coordinator_procedure(a.shards, a.work, a.live_reference, a.gpu), indent=2))
def cmd_live_reference(a: argparse.Namespace) -> None:
receipt = build_reference(a.snapshots, a.rates, a.paused_at, a.rollback_at)
with open(a.out, "x") as fh:
json.dump(receipt, fh, indent=2, allow_nan=False)
print(json.dumps(receipt, indent=2))
def cmd_score(a: argparse.Namespace) -> None:
with open(os.path.join(a.shards, "MANIFEST.json")) as fh:
man = json.load(fh)
seqs: dict[str, Sequence] = {}
for f in man["splits"][a.split]["files"]:
sh = S.Shard(os.path.join(a.shards, f["file"]))
for i in range(len(sh)):
s = sh.seq(i)
seqs[s["seq_id"]] = {"tokens": sh.tokens(i).tolist(), "gen_start": s["gen_start"], "L": s["L"],
"family": s["family"]}
drafts = {}
with open(os.path.join(a.work, f"{a.split}.k2_drafts.jsonl")) as fh:
for line in fh:
r = json.loads(line)
drafts[r["seq_id"]] = {int(p): d for p, d in r["drafts"].items()}
live = verify_reference(a.live_reference) if a.live_reference else None
rep = {"split": a.split, **score(seqs, drafts, a.split, live_reference=live, mode=a.mode),
"manifest_sha256": file_hash(Path(a.shards)/"MANIFEST.json"),
"k2_sha256": file_hash(Path(a.work)/f"{a.split}.k2_drafts.jsonl"),
"live_reference_path": str(a.live_reference.resolve()) if a.live_reference else None}
with open(a.out, "w") as fh:
json.dump(rep, fh, indent=1)
print(json.dumps(rep, indent=1))
# ------------------------------------------------------------------------------------------- pre-check on captures
def cmd_k2_captures(a: argparse.Namespace) -> None:
"""Worker drafts vs the engine's drafts on window-1 blocks whose 128-row window is all decode taps (no prefill rows
are needed): validates the K2 runner on real taps before any D2 data exists."""
dev = check_gpu()
if a.v1 not in sys.path:
sys.path.insert(0, a.v1)
CaptureSet = importlib.import_module("verify.captures").CaptureSet
cs = CaptureSet(a.captures)
if cs.plumbing:
raise SystemExit(f"capture plumbing errors: {cs.plumbing[:3]}")
reqs = sorted(r for r in cs.rows if "warmup" not in r)
seqs, eng = [], {}
for r in reqs:
rows = cs.rows[r]
g = min(s.P for s in cs.samples if s.req == r) + 1
L = max(rows) + 1
# every window row must carry decode taps (rounds of > 64 tokens, e.g. a decode batched with another
# request's prefill chunk, saved no taps)
smp = [s for s in cs.samples if s.req == r and s.P - 127 >= g
and not (st := cs.window_status(s))["missing"] and not st["no_taps"]]
if not smp:
continue
for s in smp:
eng[(r, s.P)] = (s.engine_drafts, s.realized)
tok = torch.tensor([rows[p].tok for p in range(L)])
seqs.append(SeqSource(r, tok, g, (lambda rows: lambda p0, p1: torch.stack([rows[p].taps for p in
range(p0, p1)]))(rows),
sorted({s.P for s in smp}), anchors={s.P: s.anchor for s in smp}))
t0 = time.time()
m = load_k2(a.root, dev, a.slots)
got = {}
n = run_k2(m, seqs, a.slots, t0 + a.deadline_s, lambda sid, out: got.update({(sid, p): d for p, d in out.items()}))
same, pos, lw, le = 0, [0] * 5, [], []
for k, (ed, real) in eng.items():
w = got[k]
same += int(w == ed)
for i in range(5):
pos[i] += int(w[i] == ed[i])
mw, me = match_len(w, real), match_len(ed, real)
if mw is not None and me is not None:
lw.append(mw)
le.append(me)
nb = len(eng)
rep = {"requests": n, "blocks": nb, "worker_vs_engine_block_exact": round(same / nb, 4),
"worker_vs_engine_pos_equal": [round(x / nb, 4) for x in pos],
"proxy_worker": round(sum(lw) / len(lw), 4), "proxy_engine": round(sum(le) / len(le), 4),
"scored": len(lw), "seconds": round(time.time() - t0, 1)}
with open(a.out, "w") as fh:
json.dump(rep, fh, indent=1)
print(json.dumps(rep, indent=1))
def main() -> None:
ap = argparse.ArgumentParser()
sp = ap.add_subparsers(dest="cmd", required=True)
p = sp.add_parser("plumbing")
p.add_argument("--shards", required=True)
p.add_argument("--stride", type=int, default=1)
k = sp.add_parser("k2")
k.add_argument("--shards", required=True)
k.add_argument("--split", default=BINDING_SPLIT)
k.add_argument("--work", required=True)
k.add_argument("--root", default=ROOT)
k.add_argument("--slots", type=int, default=6, help="sequences drafted together (x 5 rows <= the 32-row scratch)")
k.add_argument("--deadline-s", type=float, default=780.0)
s = sp.add_parser("score")
s.add_argument("--shards", required=True)
s.add_argument("--split", default=BINDING_SPLIT, help=f"G1(b) binds on {BINDING_SPLIT}; others are information")
s.add_argument("--work", required=True)
s.add_argument("--out", required=True)
s.add_argument("--mode", choices=("live-reference", "historical-acap"), default="live-reference")
s.add_argument("--live-reference", type=Path)
r = sp.add_parser("live-reference", help="CPU only: bind paused, pre-rollback snapshots")
r.add_argument("--snapshots", nargs="+", required=True, type=Path)
r.add_argument("--rates", nargs="+", required=True, type=Path)
r.add_argument("--paused-at", required=True)
r.add_argument("--rollback-at", required=True)
r.add_argument("--out", required=True)
c = sp.add_parser("k2-captures")
c.add_argument("--captures", required=True)
c.add_argument("--v1", required=True)
c.add_argument("--root", default=ROOT)
c.add_argument("--slots", type=int, default=4)
c.add_argument("--deadline-s", type=float, default=600.0)
c.add_argument("--out", required=True)
p = sp.add_parser("procedure", help="dry-only coordinator command renderer; no host access")
p.add_argument("--shards", required=True, type=Path)
p.add_argument("--work", required=True, type=Path)
p.add_argument("--live-reference", required=True, type=Path)
p.add_argument("--gpu", choices=GPU_ALLOWED, default=GPU_ALLOWED[0])
a = ap.parse_args()
{"plumbing": cmd_plumbing, "k2": cmd_k2, "score": cmd_score, "live-reference": cmd_live_reference, "procedure": cmd_procedure, "k2-captures": cmd_k2_captures}[a.cmd](a)
if __name__ == "__main__":
main()
|