Files
2024-09-25 18:04:29 +08:00

76 lines
2.6 KiB
Python

# Copyright (c) Meta Platforms, Inc. and affiliates.
import sys
from math import ceil
import numpy as np
from vllm import SamplingParams
from torch.utils.data import DataLoader
from utils import TokenizedDataset, complete_code
import tasks
class Generator:
def __init__(self, model, tokenizer, args):
self.model = model
self.tokenizer = tokenizer
self.args = args
def generate(self, task_name):
if self.args.model == "Phind/Phind-CodeLlama-34B-v2" and task_name == "output_prediction":
task = tasks.get_task(task_name, cot=self.args.cot, phind_output=True)
else:
task = tasks.get_task(task_name, cot=self.args.cot, phind_output=False)
dataset = task.get_dataset()
if self.args.limit is not None:
dataset = dataset.select(range(self.args.limit))
dataset_rows = range(dataset.num_rows)
dataset = dataset.add_column("row_index", dataset_rows)
if self.args.end is None:
self.args.end = dataset.num_rows
dataset = dataset.select(range(self.args.start, self.args.end))
dataset_rows = range(dataset.num_rows)
# shuffle the dataset
if self.args.shuffle:
dataset_rows = np.random.permutation(dataset_rows)
dataset = dataset.select(dataset_rows)
n_tasks = dataset.num_rows
prompts = [self.args.prefix + task.get_prompt(dataset[i]) for i in range(n_tasks)]
task_ids = [item["id"] for item in dataset]
if not self.args.do_sample:
print("Greedy decoding ON: setting temperature=0.0")
self.args.temperature = 0.0
self.args.top_p = 1.0
sampling_params = SamplingParams(
n=self.args.batch_size,
temperature=self.args.temperature,
top_p=self.args.top_p,
top_k=self.args.top_k,
max_tokens=self.args.max_length_generation,
stop=task.stop_words,
)
print(f"{sampling_params = }")
generations, generations_raw = complete_code(task, self.model, sampling_params, prompts, task_ids)
references = [task.get_reference(dataset[i]) for i in range(n_tasks)]
if len(list(generations.values())[0]) > self.args.n_samples:
generations = {k: v[: self.args.n_samples] for k, v in generations.items()}
generations_raw = {k: v[: self.args.n_samples] for k, v in generations_raw.items()}
assert all([len(gen) == self.args.n_samples for gen in generations.values()]), f"{[len(gen) for gen in generations.values()]}"
return generations, generations_raw, references