hf_text-generation-inference/server/text_generation_server/models/t5.py

116 lines
3.4 KiB
Python

import torch
import torch.distributed
from typing import List, Optional, Tuple
from transformers import (
AutoTokenizer,
AutoConfig,
)
from text_generation_server.models import Seq2SeqLM
from text_generation_server.models.custom_modeling.t5_modeling import (
T5ForConditionalGeneration,
)
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
class T5Sharded(Seq2SeqLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
speculator: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
dtype = torch.float16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32 if dtype is None else dtype
config = AutoConfig.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
)
config.quantize = quantize
config.speculator = speculator
tokenizer = AutoTokenizer.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
tokenizer.bos_token_id = config.decoder_start_token_id
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames,
device=device,
dtype=dtype,
process_group=self.process_group,
aliases={
"shared.weight": [
"encoder.embed_tokens.weight",
"decoder.embed_tokens.weight",
]
},
)
model = T5ForConditionalGeneration(config, weights)
torch.distributed.barrier(group=self.process_group)
super(Seq2SeqLM, self).__init__(
model_id=model_id,
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
def forward(
self,
input_ids,
attention_mask,
decoder_input_ids,
decoder_attention_mask: Optional,
encoder_last_hidden_state: Optional,
past_key_values: Optional = None,
) -> Tuple[
torch.Tensor,
torch.Tensor,
List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]],
]:
# Model Forward
outputs, speculative_logits = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
decoder_input_ids=decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_outputs=encoder_last_hidden_state,
past_key_values=past_key_values,
use_cache=True,
)
return (
outputs.logits,
speculative_logits,
outputs.encoder_last_hidden_state,
outputs.past_key_values,
)