File size: 8,488 Bytes
f6d0311 | 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 | #!/usr/bin/env python3
"""vLLM GSM8K eval for IAPO released checkpoints.
Runs one model per process/job. This avoids vLLM engine teardown issues and
gives a cleaner wall-clock generation measurement than the Transformers script.
"""
from __future__ import annotations
import argparse
import csv
import gc
import json
import math
import os
import random
import re
import time
from pathlib import Path
from typing import Any
from datasets import load_dataset
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
SYSTEM_PROMPT = """A conversation between User and Assistant. The user asks a question, and the Assistant solves it.
The assistant first thinks about the reasoning process in the mind and then provides the user
with the answer. The reasoning process and answer are enclosed within <think> and
<answer> tags, respectively, i.e., <think> reasoning process here </think>
<answer> answer here </answer>.
The answer must be a single integer."""
FEWSHOT_USER = "What is 2+2?"
FEWSHOT_ASSISTANT = "<think>To calculate 2+2, we simply add the numbers together: 2 + 2 = 4.</think>\n<answer>4</answer>"
MODEL_SPECS = {
"base-7b": {"repo": "Qwen/Qwen2.5-7B-Instruct", "subfolder": None},
"iapo-7b-gsm8k": {"repo": "jonathanhe123/iapo", "subfolder": "Qwen2.5-7B-Instruct_GSM8K"},
"base-0.5b": {"repo": "Qwen/Qwen2.5-0.5B-Instruct", "subfolder": None},
"iapo-0.5b-gsm8k": {"repo": "jonathanhe123/iapo", "subfolder": "Qwen2.5-0.5B-Instruct_GSM8K"},
}
def normalize_answer(s: str | None) -> str:
if s is None:
return ""
s = str(s).strip().split("=")[-1]
s = s.replace(",", "").replace("$", "").strip()
s = re.sub(r"\\boxed\{([^{}]+)\}", r"\1", s)
return s.strip().rstrip(".")
def extract_gold(answer: str) -> str:
return normalize_answer(answer.split("####")[-1])
def extract_pred(text: str) -> str:
xml = re.search(r"<answer>(.*?)</answer>", text, flags=re.DOTALL | re.IGNORECASE)
if xml:
return normalize_answer(xml.group(1))
answer_line = re.findall(r"Answer:\s*([-+]?\d[\d,]*(?:\.\d+)?)", text, flags=re.IGNORECASE)
if answer_line:
return normalize_answer(answer_line[-1])
nums = re.findall(r"[-+]?\d[\d,]*(?:\.\d+)?", text)
return normalize_answer(nums[-1]) if nums else ""
def local_model_path(repo: str, subfolder: str | None, cache_dir: str | None) -> str:
if not subfolder:
return repo
patterns = [f"{subfolder}/*"]
root = snapshot_download(repo_id=repo, allow_patterns=patterns, cache_dir=cache_dir)
return str(Path(root) / subfolder)
def make_prompt(tokenizer: Any, question: str) -> str:
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": FEWSHOT_USER},
{"role": "assistant", "content": FEWSHOT_ASSISTANT},
{"role": "user", "content": question},
]
return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--model-key", required=True, choices=sorted(MODEL_SPECS))
parser.add_argument("--sample-size", type=int, default=128)
parser.add_argument("--num-return-sequences", type=int, default=8)
parser.add_argument("--max-new-tokens", type=int, default=256)
parser.add_argument("--temperature", type=float, default=1.0)
parser.add_argument("--top-p", type=float, default=1.0)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--dtype", default="bfloat16")
parser.add_argument("--gpu-memory-utilization", type=float, default=0.90)
parser.add_argument("--max-model-len", type=int, default=1024)
parser.add_argument("--output-dir", default="outputs/hf_gsm8k_vllm")
parser.add_argument("--cache-dir", default=os.environ.get("HF_HOME"))
args = parser.parse_args()
random.seed(args.seed)
spec = MODEL_SPECS[args.model_key]
out_dir = Path(args.output_dir)
out_dir.mkdir(parents=True, exist_ok=True)
model_path = local_model_path(spec["repo"], spec["subfolder"], args.cache_dir)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
tokenizer.padding_side = "left"
dataset = load_dataset("openai/gsm8k", "main", split="test")
indices = list(range(len(dataset)))
random.Random(args.seed).shuffle(indices)
indices = sorted(indices[: args.sample_size])
examples = [dataset[i] for i in indices]
prompts = [make_prompt(tokenizer, ex["question"]) for ex in examples]
golds = [extract_gold(ex["answer"]) for ex in examples]
env = {
"event": "env",
"model_key": args.model_key,
"repo": spec["repo"],
"subfolder": spec["subfolder"],
"model_path": model_path,
"sample_size": len(examples),
"num_return_sequences": args.num_return_sequences,
"max_new_tokens": args.max_new_tokens,
}
print(json.dumps(env), flush=True)
llm = LLM(
model=model_path,
tokenizer=model_path,
trust_remote_code=True,
dtype=args.dtype,
gpu_memory_utilization=args.gpu_memory_utilization,
max_model_len=args.max_model_len,
seed=args.seed,
)
sampling = SamplingParams(
n=args.num_return_sequences,
temperature=args.temperature,
top_p=args.top_p,
max_tokens=args.max_new_tokens,
seed=args.seed,
)
t0 = time.perf_counter()
outputs = llm.generate(prompts, sampling, use_tqdm=True)
gen_seconds = time.perf_counter() - t0
per_question_correct = []
per_question_lengths = []
details_path = out_dir / f"{args.model_key}_generations.jsonl"
with details_path.open("w") as f:
for ex_idx, (dataset_idx, gold, request_output) in enumerate(zip(indices, golds, outputs)):
correct_row = []
length_row = []
for sample_idx, completion in enumerate(request_output.outputs):
text = completion.text
token_ids = completion.token_ids or []
pred = extract_pred(text)
ok = int(pred == gold)
correct_row.append(ok)
length_row.append(len(token_ids))
f.write(json.dumps({
"model": args.model_key,
"dataset_index": dataset_idx,
"sample_in_group": sample_idx,
"gold": gold,
"prediction": pred,
"correct": ok,
"completion_tokens": len(token_ids),
"completion": text,
}) + "\n")
per_question_correct.append(correct_row)
per_question_lengths.append(length_row)
k_values = [k for k in [1, 2, 4, 8, 16, 32] if k <= args.num_return_sequences]
generated_count = len(examples) * args.num_return_sequences
summary = {
"event": "summary",
"model": args.model_key,
"repo": spec["repo"],
"subfolder": spec["subfolder"],
"sample_size": len(examples),
"num_return_sequences": args.num_return_sequences,
"max_new_tokens": args.max_new_tokens,
"generated_count": generated_count,
"generation_seconds": gen_seconds,
"avg_seconds_per_completion": gen_seconds / max(1, generated_count),
}
for k in k_values:
pass_at_k = sum(int(any(row[:k])) for row in per_question_correct) / len(per_question_correct)
length_at_k = sum(sum(row[:k]) for row in per_question_lengths) / (len(per_question_lengths) * k)
summary[f"pass@{k}"] = pass_at_k
summary[f"length@{k}"] = length_at_k
summary[f"ratio@{k}"] = pass_at_k / length_at_k if length_at_k else math.nan
summary_path = out_dir / f"{args.model_key}_summary.json"
csv_path = out_dir / f"{args.model_key}_summary.csv"
summary_path.write_text(json.dumps(summary, indent=2))
with csv_path.open("w", newline="") as f:
writer = csv.DictWriter(f, fieldnames=sorted(summary))
writer.writeheader()
writer.writerow(summary)
print(json.dumps(summary), flush=True)
print(json.dumps({
"event": "wrote_outputs",
"summary": str(summary_path),
"csv": str(csv_path),
"details": str(details_path),
}), flush=True)
del llm
gc.collect()
if __name__ == "__main__":
main()
|