hf_text-generation-inference/server/text_generation/models/santacoder.py

import torch
import torch.distributed

from typing import Optional, List, Tuple
from transformers import AutoTokenizer, AutoModelForCausalLM

from text_generation.models import CausalLM

FIM_PREFIX = "<fim-prefix>"
FIM_MIDDLE = "<fim-middle>"
FIM_SUFFIX = "<fim-suffix>"
FIM_PAD = "<fim-pad>"
EOD = "<|endoftext|>"


class SantaCoder(CausalLM):
    def __init__(self, model_name: str, quantize=False):
        if torch.cuda.is_available():
            device = torch.device("cuda")
            dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
        else:
            if quantize:
                raise ValueError("quantization is not available on CPU")

            device = torch.device("cpu")
            dtype = torch.float32

        tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="left")
        tokenizer.add_special_tokens(
            {
                "additional_special_tokens": [
                    EOD,
                    FIM_PREFIX,
                    FIM_MIDDLE,
                    FIM_SUFFIX,
                    FIM_PAD,
                ],
                "pad_token": EOD,
            }
        )

        self.model = AutoModelForCausalLM.from_pretrained(
            model_name,
            torch_dtype=dtype,
            load_in_8bit=quantize,
            trust_remote_code=True,  # required
        ).to(device).eval()

        super(CausalLM, self).__init__(
            tokenizer=tokenizer,
            device=device,
        )

    def decode(self, generated_ids: List[int]) -> str:
        # Do not skip special tokens as they are used for custom parsing rules of the generated text
        return self.tokenizer.decode(
            generated_ids, skip_special_tokens=False, cleanup_tokenization_spaces=False
        )