sglang0.4.5.post1/python/sglang/srt/model_loader/utils.py

42 lines
1.3 KiB
Python

# Adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/model_executor/model_loader/utils.py
"""Utilities for selecting and loading models."""
import contextlib
from typing import Tuple, Type
import torch
from torch import nn
from sglang.srt.configs.model_config import ModelConfig
@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]:
from sglang.srt.models.registry import ModelRegistry
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]