mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-04-20 14:22:08 +00:00
093a27c528
Add support for GPTQ Marlin kernels GPTQ Marlin extends the Marlin kernels to support common GPTQ configurations: - bits: 4 or 8 - groupsize: -1, 32, 64, or 128 - desc_act: true/false Using the GPTQ Marlin kernels requires repacking the parameters in the Marlin quantizer format. The kernels were contributed by Neural Magic to VLLM. We vendor them here for convenience.
89 lines
2.7 KiB
Python
89 lines
2.7 KiB
Python
import torch
|
|
import torch.distributed
|
|
|
|
from typing import Optional
|
|
|
|
from transformers import (
|
|
AutoTokenizer,
|
|
AutoConfig,
|
|
)
|
|
from text_generation_server.models import CausalLM
|
|
from text_generation_server.models.custom_modeling.neox_modeling import (
|
|
GPTNeoxForCausalLM,
|
|
)
|
|
from text_generation_server.utils import (
|
|
initialize_torch_distributed,
|
|
weight_files,
|
|
Weights,
|
|
)
|
|
|
|
|
|
class GPTNeoxSharded(CausalLM):
|
|
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
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
model_id,
|
|
revision=revision,
|
|
padding_side="left",
|
|
truncation_side="left",
|
|
trust_remote_code=trust_remote_code,
|
|
)
|
|
tokenizer.pad_token = tokenizer.eos_token
|
|
|
|
config = AutoConfig.from_pretrained(
|
|
model_id,
|
|
revision=revision,
|
|
trust_remote_code=trust_remote_code,
|
|
)
|
|
config.quantize = quantize
|
|
config.speculator = speculator
|
|
|
|
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
|
|
)
|
|
if config.quantize in ["gptq", "marlin"]:
|
|
weights._set_gptq_params(model_id, revision)
|
|
|
|
model = GPTNeoxForCausalLM(config, weights)
|
|
|
|
torch.distributed.barrier(group=self.process_group)
|
|
super(CausalLM, self).__init__(
|
|
model=model,
|
|
tokenizer=tokenizer,
|
|
requires_padding=True,
|
|
dtype=dtype,
|
|
device=device,
|
|
rank=rank,
|
|
world_size=world_size,
|
|
)
|
|
|
|
def forward(
|
|
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
|
|
):
|
|
outputs, speculative_logits = self.model.forward(
|
|
input_ids=input_ids,
|
|
attention_mask=attention_mask,
|
|
position_ids=position_ids,
|
|
past_key_values=past_key_values,
|
|
use_cache=True,
|
|
)
|
|
|
|
return outputs.logits, speculative_logits, outputs.past_key_values
|