File size: 5,606 Bytes
682c27c 0828337 682c27c 0828337 682c27c 0828337 682c27c 0828337 682c27c 0828337 682c27c 0828337 682c27c 0828337 682c27c | 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 | #!/usr/bin/env python3
"""HumanEval via raw /v1/completions (no chat template).
For models whose chat template is broken/missing (e.g. MiniCPM5-2B whose
jinja renders prompts the model answers with emptiness): classic base-model
HE protocol — prompt = problem["prompt"], raw completion, extract code.
Same sampling protocol otherwise (temp/top_p/top_k/min_p/presence via HE_*).
Server lifecycle: same as run_humaneval (HUMANEVAL_SERVE_MODEL self-serve
with auto-kill, or external HUMANEVAL_SERVER). Extra server flags via
HUMANEVAL_SERVER_ARGS.
"""
import json
import os
import signal
import socket
import subprocess
import sys
import time
from urllib.parse import urlparse
import requests
from human_eval.data import read_problems, write_jsonl
from human_eval.evaluation import evaluate_functional_correctness
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import protocol
SERVER_URL = os.environ.get("HUMANEVAL_SERVER", "http://127.0.0.1:28082")
SERVE_MODEL = os.environ.get("HUMANEVAL_SERVE_MODEL")
SERVER_BIN = protocol.SERVER_BIN
REPO_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
OUTPUT_FILE = os.environ.get(
"HUMANEVAL_OUT",
os.path.join(REPO_ROOT, "eval_results", "humaneval_raw.jsonl"),
)
def _f(name, default):
try:
return float(os.environ.get(name, default))
except ValueError:
return float(default)
# Same protocol module as the chat runner (this file used to default
# max_tokens to 1024, the chat one to 2048 — silent 2x split).
SAMPLE = {k: v for k, v in protocol.sample_params().items() if k != "max_tokens"}
def _port_busy(port):
with socket.socket() as s:
s.settimeout(1)
return s.connect_ex(("127.0.0.1", port)) == 0
def _serve_own_model():
port = urlparse(SERVER_URL).port or 28082
if _port_busy(port):
raise SystemExit(f"refusing to serve: port {port} busy (strict queue)")
log = open("/tmp/he_raw_serve_%d.log" % port, "w")
extra = os.environ.get("HUMANEVAL_SERVER_ARGS", "").split()
p = subprocess.Popen(
[SERVER_BIN, "-m", SERVE_MODEL, "--port", str(port)]
+ protocol.SERVER_ARGS + extra,
stdout=log, stderr=subprocess.STDOUT)
for _ in range(120):
try:
r = requests.get(f"{SERVER_URL}/health", timeout=5)
if r.status_code == 200:
print(f"serving {SERVE_MODEL} on {port}", flush=True)
return p
except Exception:
pass
if p.poll() is not None:
raise SystemExit("own server died during load")
time.sleep(5)
p.send_signal(signal.SIGTERM)
raise SystemExit(f"own server on {port} never healthy — killed")
def generate(prompt, max_tokens=None):
if max_tokens is None:
max_tokens = protocol.sample_params()["max_tokens"]
resp = requests.post(
f"{SERVER_URL}/v1/completions",
json={"prompt": prompt, "max_tokens": max_tokens, **SAMPLE},
timeout=300,
)
resp.raise_for_status()
return resp.json()["choices"][0]["text"]
def extract_code(text):
text = text.strip()
if "```python" in text:
text = text.split("```python")[1]
if "```" in text:
text = text.split("```")[0]
return text.strip()
def main():
os.makedirs(os.path.dirname(OUTPUT_FILE), exist_ok=True)
srv = _serve_own_model() if SERVE_MODEL else None
try:
try:
r = requests.get(f"{SERVER_URL}/health", timeout=10)
r.raise_for_status()
except Exception as e:
raise SystemExit(f"server not healthy at {SERVER_URL}: {e}")
problems = read_problems()
if os.environ.get("HE_LIMIT"):
keep = sorted(problems)[:int(os.environ["HE_LIMIT"])]
problems = {k: problems[k] for k in keep}
print(f"HE_LIMIT: first {len(problems)} tasks only", flush=True)
results = []
total = len(problems)
for i, (task_id, problem) in enumerate(sorted(problems.items())):
if i % 10 == 0:
try:
r = requests.get(f"{SERVER_URL}/health", timeout=10)
r.raise_for_status()
except Exception as e:
raise SystemExit(
f"server died mid-run at task {i}/{total}: {e}. "
f"Results so far DISCARDED (not scored).")
print(f"[{i+1}/{total}] {task_id} ... ", end="", flush=True)
try:
raw = generate(problem["prompt"])
code = extract_code(raw) or raw
results.append({"task_id": task_id, "completion": code})
print("OK")
except Exception as e:
print(f"ERROR: {e}")
results.append({"task_id": task_id, "completion": ""})
time.sleep(0.1)
write_jsonl(OUTPUT_FILE, results)
print(f"\nSaved {len(results)} to {OUTPUT_FILE}")
side = protocol.write_meta(OUTPUT_FILE, protocol.battery_meta(
SERVER_URL, SERVE_MODEL,
extra=os.environ.get("HUMANEVAL_SERVER_ARGS", "").split()))
print(f"protocol sidecar: {side}")
print("\n--- Evaluating pass@1 ---")
r = evaluate_functional_correctness(sample_file=OUTPUT_FILE, k=[1],
n_workers=4)
print(f"pass@1: {r}")
finally:
if srv is not None:
srv.send_signal(signal.SIGTERM)
print("own server killed — GPU free", flush=True)
if __name__ == "__main__":
main()
|