Datasets:
File size: 4,345 Bytes
b48c9f8 | 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 | """Serve each small model with vLLM on Kaggle's two T4s, run the stepmold bench, keep the results.
Runs top to bottom with no input, so it works as a "Save & Run All" commit. Each model's
results go to /kaggle/working/results/<label>.jsonl, appended and resumable, so a rerun only
does the case-runs that are missing.
"""
import hashlib
import json
import os
import signal
import subprocess
import sys
import tarfile
import time
import urllib.request
WORK = "/kaggle/working"
OUT = os.path.join(WORK, "results")
BENCH = os.path.join(WORK, "stepmold-bench")
STEPMOLD = os.path.join(WORK, "stepmold") # a clone at 38235ab: pip installs leave out examples/
PORT = 8000
# One sampling setup for every model, so three runs of a case show how consistent it is.
# 0.7 and top_p 0.8 are Qwen's recommended non-thinking settings. Qwen3 runs with thinking
# off, the same way the stepmold GPU Space runs Qwen3 8B.
QWEN = {"temperature": 0.7, "top_p": 0.8}
NO_THINK = {"chat_template_kwargs": {"enable_thinking": False}}
MODELS = [
# label, Hugging Face id, tensor-parallel size, request extras
("qwen2.5-0.5b-instruct", "Qwen/Qwen2.5-0.5B-Instruct", 1, QWEN),
("qwen2.5-1.5b-instruct", "Qwen/Qwen2.5-1.5B-Instruct", 1, QWEN),
("qwen3-4b", "Qwen/Qwen3-4B", 1, dict(QWEN, **NO_THINK)),
("qwen2.5-7b-instruct", "Qwen/Qwen2.5-7B-Instruct", 2, QWEN),
("qwen3-8b", "Qwen/Qwen3-8B", 2, dict(QWEN, **NO_THINK)),
]
def wait_ready(proc, timeout=1500):
deadline = time.time() + timeout
while time.time() < deadline:
if proc.poll() is not None:
raise RuntimeError("vLLM exited with code %s" % proc.returncode)
try:
with urllib.request.urlopen("http://localhost:%d/v1/models" % PORT, timeout=5) as reply:
if reply.status == 200:
return
except Exception:
pass
time.sleep(5)
raise TimeoutError("vLLM did not come up within %ds" % timeout)
def serve(model, tp, log):
cmd = [
"vllm", "serve", model,
"--dtype", "half", # a T4 has no bf16
"--tensor-parallel-size", str(tp),
"--max-model-len", "4096",
"--gpu-memory-utilization", "0.90",
"--port", str(PORT),
]
return subprocess.Popen(cmd, stdout=log, stderr=subprocess.STDOUT, start_new_session=True)
def stop(proc):
try:
os.killpg(proc.pid, signal.SIGTERM)
proc.wait(timeout=120)
except subprocess.TimeoutExpired:
os.killpg(proc.pid, signal.SIGKILL)
except ProcessLookupError:
pass
time.sleep(15) # let the GPUs release their memory before the next model loads
def run_model(label, model, tp, extra):
started = time.time()
with open(os.path.join(OUT, label + ".vllm.log"), "w") as log:
proc = serve(model, tp, log)
try:
wait_ready(proc)
print("%s: server ready in %.0fs" % (label, time.time() - started), flush=True)
subprocess.run([
sys.executable, os.path.join(BENCH, "bench.py"),
"--stepmold", STEPMOLD,
"--model", "openai:http://localhost:%d/v1#%s" % (PORT, model),
"--label", label,
"--arms", "stepmold,plain_schema,plain",
"--runs", "3",
"--packs", "all",
"--concurrency", "16",
"--extra-body", json.dumps(extra),
"--out", os.path.join(OUT, label + ".jsonl"),
], check=False)
except Exception as exc:
print("%s: FAILED: %s" % (label, exc), flush=True)
finally:
stop(proc)
print("%s: finished in %.0fs" % (label, time.time() - started), flush=True)
def main(only=None):
os.makedirs(OUT, exist_ok=True)
digest = hashlib.sha256(open(os.path.join(BENCH, "bench.py"), "rb").read()).hexdigest()[:12]
print("bench.py sha256", digest, flush=True)
for label, model, tp, extra in MODELS:
if not only or label in only:
run_model(label, model, tp, extra)
with tarfile.open(os.path.join(WORK, "stepmold-bench-kaggle-results.tar.gz"), "w:gz") as tar:
tar.add(OUT, arcname="results")
print("packed", os.path.join(WORK, "stepmold-bench-kaggle-results.tar.gz"), flush=True)
if __name__ == "__main__":
main(set(sys.argv[1:]) or None)
|