Witllm/wit/inference.py

import pytorch_lightning as pl
import torch

from model.qwen_module import QwenModule
from model.modeling_wit import QwenRunner
from model.tokenization_qwen import QWenTokenizer
import numpy as np

import configuration
import dataset.dataset as ds
import dataset.node_tree as nt

if __name__ == "__main__":

    conf = configuration.TrainConfig()
    config = conf.model_config

    conf.name = "bigger"  # current train process name
    conf.pretrain_model_name = None  # "qwen/Qwen-1_8B-Chat"
    conf.learning_rate = 0.0001
    conf.use_tril_attention_mask = None
    conf.precision = "bf16-mixed"  # "precision:bf16-mixed,16-mixed,32-true"
    conf.train_batch_size = 16
    conf.val_batch_size = 4
    conf.num_proc = 8
    conf.max_epochs = 1000
    conf.strategy = "auto"
    conf.resume_from_ckpt_path = None
    conf.seed = 42
    conf.dataloader_works = 2

    conf.dataset.meaning.val_mask_level = [0, 1, 2]
    conf.dataset.meaning.val_mask_idx = [0, 0, -1]

    config.vocab_size = 256
    config.hidden_size = 128  # 128 1024 2048  32
    config.num_hidden_layers = 3  # 6 12 24  3
    config.num_attention_heads = 16  # 8 8 16

    torch.manual_seed(conf.seed)

    checkpoint_path = "log/bigger/version_1/checkpoints/epoch=23-step=24792.ckpt"
    qwen = QwenModule.load_from_checkpoint(checkpoint_path=checkpoint_path)
    qwen.eval()

    runner = QwenRunner(qwen.llm)

    val = ds.InitValDataset(conf).dataset
    md = val.meaning_dataset

    map = md.get_meaning_map()

    item = md.get_token(0)
    nt.NodeTree(map.get_tree(md.get_meaning(0))).print()

    batch = torch.tensor([item[:-16]], dtype=torch.int64)
    batch = batch.cuda()
    # print(item)
    next_token = runner.ChatToken(batch)
    print(next_token.detach().cpu().numpy())