|
from typing import Any, Dict, List |
|
|
|
import torch |
|
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline |
|
|
|
dtype = torch.bfloat16 |
|
|
|
|
|
class EndpointHandler: |
|
def __init__(self, path=""): |
|
|
|
self.tokenizer = AutoTokenizer.from_pretrained(path) |
|
self.model = AutoModelForCausalLM.from_pretrained( |
|
path, device_map="auto", torch_dtype=dtype |
|
) |
|
if self.tokenizer.pad_token is None: |
|
self.tokenizer.pad_token = self.tokenizer.eos_token |
|
|
|
self.pipeline = pipeline( |
|
"text-generation", model=self.model, tokenizer=self.tokenizer |
|
) |
|
self.ce = torch.nn.CrossEntropyLoss( |
|
ignore_index=self.tokenizer.pad_token_id, reduction="none" |
|
) |
|
|
|
def compute_log_likelihood(self, lm_logits, input_ids): |
|
predictions = lm_logits[..., :-1, :].contiguous() |
|
target_ids = input_ids[..., 1:].contiguous() |
|
|
|
ce_loss = self.ce( |
|
predictions.view(-1, predictions.size(-1)), |
|
target_ids.view(-1), |
|
) |
|
return -ce_loss.view_as(target_ids)[0] |
|
|
|
def __call__(self, data: Any): |
|
inputs = data.pop("inputs", data) |
|
parameters = data.pop("parameters", None) |
|
input_tokens = self.tokenizer.batch_encode_plus( |
|
[inputs], return_tensors="pt", padding=False |
|
) |
|
for t in input_tokens: |
|
if torch.is_tensor(input_tokens[t]): |
|
input_tokens[t] = input_tokens[t].to(torch.cuda.current_device()) |
|
|
|
logits = self.model( |
|
input_ids=input_tokens["input_ids"], |
|
attention_mask=input_tokens["attention_mask"], |
|
)[0] |
|
log_likelihood = self.compute_log_likelihood( |
|
logits, input_tokens["input_ids"] |
|
) |
|
return (logits, log_likelihood) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|