jamesHD2001/DenseMamba-1.3B

DenseRetNet-1.3BM

An unofficial pretraining checkpoints for DenseRetNet-1.3B of paper DenseMamba: https://arxiv.org/abs/2403.00818, the trainig data is 15B tokens randomly samples from The Pile dataset.

• recurrent generation examples:

import torch
import transformers
model_name_or_path = '/path to model'
MAX_NEW_TOKENS = 256
inference_dtype = torch.float16

generation_config = transformers.GenerationConfig(
    do_sample=False,
    max_new_tokens=MAX_NEW_TOKENS,
)

tokenizer = transformers.AutoTokenizer.from_pretrained(model_name_or_path, use_fast=False, trust_remote_code=True)
config = transformers.AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
model = transformers.AutoModelForCausalLM.from_pretrained(
    model_name_or_path, torch_dtype=torch.float16, trust_remote_code=True)  # .cuda()
model.cuda()
model = model.half()
model.eval()
input_sents = 'I have a dream'
inputs = tokenizer(input_sents, return_tensors="pt", truncation=True, max_length=2048)
output = model.generate(input_ids=inputs["input_ids"].cuda(),
                   generation_config=generation_config,
                   return_dict_in_generate=True,
                   output_scores=True
                   )
output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True)
print(output)