sraivante's picture
tiny-superfast-agentic-MLM-6.5m-v1: 6.44M-param NumPy command parser, sample executor, benchmarks
76d666a verified
Raw History Blame Contribute Delete
1.5 kB
"""Parse commands and show what the sample executor would run. Nothing is executed.
python quickstart.py "volume 40 kar do" "connect to wifi Redmi Note 12 password hello@123"
python quickstart.py # interactive
"""
import json, os, platform, sys
# small matrices: a few BLAS threads are much faster than one per core (see eval/BENCHMARK.md). Must be set before numpy loads.
_threads = "2" if platform.machine().lower() in ("aarch64", "arm64") else "4"
os.environ.setdefault("OPENBLAS_NUM_THREADS", _threads); os.environ.setdefault("OMP_NUM_THREADS", _threads)
HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, os.path.join(HERE, "model")); sys.path.insert(0, os.path.join(HERE, "executor"))
from runtime import CommandModel # noqa: E402
import executor as EX # noqa: E402
model = CommandModel(os.path.join(HERE, "model"))
def show(text):
out = model.predict(text); meta = out.pop("_meta")
plan = EX.plan(out, None, text) # None = this machine's platform; text lets it ask about bare GPIO pins
print(json.dumps({"text": text, "command": out, "confidence": meta["confidence"], "latency_ms": meta["latency_ms"],
"risk": plan.risk, "plan": plan.action, "would_run": plan.command or plan.note}, ensure_ascii=False, indent=1))
if len(sys.argv) > 1:
for t in sys.argv[1:]: show(t)
else:
print("type a command (empty line to quit)")
while True:
t = input("> ").strip()
if not t: break
show(t)