40 lines
1.2 KiB
Python
40 lines
1.2 KiB
Python
|
"""Utilities for selecting and loading models."""
|
||
|
|
import contextlib
|
||
|
|
from typing import Tuple, Type
|
||
|
|
|
||
|
|
import torch
|
||
|
|
from torch import nn
|
||
|
|
|
||
|
|
from vllm.config import ModelConfig
|
||
|
|
from vllm.model_executor.models import ModelRegistry
|
||
|
|
|
||
|
|
|
||
|
|
@contextlib.contextmanager
|
||
|
|
def set_default_torch_dtype(dtype: torch.dtype):
|
||
|
|
"""Sets the default torch dtype to the given dtype."""
|
||
|
|
old_dtype = torch.get_default_dtype()
|
||
|
|
torch.set_default_dtype(dtype)
|
||
|
|
yield
|
||
|
|
torch.set_default_dtype(old_dtype)
|
||
|
|
|
||
|
|
|
||
|
|
def get_model_architecture(
|
||
|
|
model_config: ModelConfig) -> Tuple[Type[nn.Module], str]:
|
||
|
|
architectures = getattr(model_config.hf_config, "architectures", [])
|
||
|
|
# Special handling for quantized Mixtral.
|
||
|
|
# FIXME(woosuk): This is a temporary hack.
|
||
|
|
mixtral_supported = [
|
||
|
|
"fp8", "compressed-tensors", "gptq_marlin", "awq_marlin"
|
||
|
|
]
|
||
|
|
|
||
|
|
if (model_config.quantization is not None
|
||
|
|
and model_config.quantization not in mixtral_supported
|
||
|
|
and "MixtralForCausalLM" in architectures):
|
||
|
|
architectures = ["QuantMixtralForCausalLM"]
|
||
|
|
|
||
|
|
return ModelRegistry.resolve_model_cls(architectures)
|
||
|
|
|
||
|
|
|
||
|
|
def get_architecture_class_name(model_config: ModelConfig) -> str:
|
||
|
|
return get_model_architecture(model_config)[1]
|