137 lines
6.3 KiB
Python
137 lines
6.3 KiB
Python
"""Utilities for selecting and loading models."""
|
|
import contextlib
|
|
from typing import Type
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from vllm.config import DeviceConfig, ModelConfig
|
|
from vllm.model_executor.models import ModelRegistry
|
|
from vllm.model_executor.weight_utils import (get_quant_config,
|
|
initialize_dummy_weights)
|
|
|
|
|
|
@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) -> Type[nn.Module]:
|
|
architectures = getattr(model_config.hf_config, "architectures", [])
|
|
# Special handling for quantized Mixtral.
|
|
# FIXME(woosuk): This is a temporary hack.
|
|
if (model_config.quantization is not None
|
|
and "MixtralForCausalLM" in architectures):
|
|
architectures = ["QuantMixtralForCausalLM"]
|
|
|
|
for arch in architectures:
|
|
model_cls = ModelRegistry.load_model_cls(arch)
|
|
if model_cls is not None:
|
|
return model_cls
|
|
raise ValueError(
|
|
f"Model architectures {architectures} are not supported for now. "
|
|
f"Supported architectures: {ModelRegistry.get_supported_archs()}")
|
|
|
|
|
|
def get_model(model_config: ModelConfig, device_config: DeviceConfig,
|
|
**kwargs) -> nn.Module:
|
|
lora_config = kwargs.get("lora_config", None)
|
|
model_class = _get_model_architecture(model_config)
|
|
|
|
# Get the (maybe quantized) linear method.
|
|
linear_method = None
|
|
if model_config.quantization is not None:
|
|
quant_config = get_quant_config(model_config)
|
|
capability = (9, 0)
|
|
# capability = torch.cuda.get_device_capability() avoid capability error
|
|
capability = capability[0] * 10 + capability[1]
|
|
if capability < quant_config.get_min_capability():
|
|
raise ValueError(
|
|
f"The quantization method {model_config.quantization} is not "
|
|
"supported for the current GPU. "
|
|
f"Minimum capability: {quant_config.get_min_capability()}. "
|
|
f"Current capability: {capability}.")
|
|
supported_dtypes = quant_config.get_supported_act_dtypes()
|
|
if model_config.dtype not in supported_dtypes:
|
|
raise ValueError(
|
|
f"{model_config.dtype} is not supported for quantization "
|
|
f"method {model_config.quantization}. Supported dtypes: "
|
|
f"{supported_dtypes}")
|
|
linear_method = quant_config.get_linear_method()
|
|
|
|
with _set_default_torch_dtype(model_config.dtype):
|
|
# Create a model instance.
|
|
# The weights will be initialized as empty tensors.
|
|
try:
|
|
# with torch.device contextmanager need torch >= 2.0
|
|
with torch.device(device_config.device):
|
|
if hasattr(model_class, "supported_lora_modules"):
|
|
model = model_class(model_config.hf_config, linear_method,
|
|
lora_config)
|
|
elif lora_config:
|
|
raise ValueError(
|
|
f"Model {model_class.__name__} does not support LoRA, "
|
|
"but LoRA is enabled. Support for this model may "
|
|
"be added in the future. If this is important to you, "
|
|
"please open an issue on github.")
|
|
else:
|
|
model = model_class(model_config.hf_config, linear_method)
|
|
if model_config.load_format == "dummy":
|
|
# NOTE(woosuk): For accurate performance evaluation, we assign
|
|
# random values to the weights.
|
|
initialize_dummy_weights(model)
|
|
else:
|
|
# Load the weights from the cached or downloaded files.
|
|
model.load_weights(model_config.model, model_config.download_dir,
|
|
model_config.load_format, model_config.revision)
|
|
# for torch < 2.0
|
|
except:
|
|
if hasattr(model_class, "supported_lora_modules"):
|
|
model = model_class(model_config.hf_config, linear_method,
|
|
lora_config)
|
|
elif lora_config:
|
|
raise ValueError(
|
|
f"Model {model_class.__name__} does not support LoRA, "
|
|
"but LoRA is enabled. Support for this model may "
|
|
"be added in the future. If this is important to you, "
|
|
"please open an issue on github.")
|
|
else:
|
|
model = model_class(model_config.hf_config, linear_method)
|
|
model = model.cuda()
|
|
if model_config.load_format == "dummy":
|
|
# NOTE(woosuk): For accurate performance evaluation, we assign
|
|
# random values to the weights.
|
|
initialize_dummy_weights(model)
|
|
else:
|
|
# Load the weights from the cached or downloaded files.
|
|
model.load_weights(model_config.model, model_config.download_dir,
|
|
model_config.load_format, model_config.revision)
|
|
return model.eval()
|
|
# TODO align
|
|
"""
|
|
with torch.device(device_config.device):
|
|
if hasattr(model_class, "supported_lora_modules"):
|
|
model = model_class(model_config.hf_config, linear_method,
|
|
lora_config)
|
|
elif lora_config:
|
|
raise ValueError(
|
|
f"Model {model_class.__name__} does not support LoRA, "
|
|
"but LoRA is enabled. Support for this model may "
|
|
"be added in the future. If this is important to you, "
|
|
"please open an issue on github.")
|
|
else:
|
|
model = model_class(model_config.hf_config, linear_method)
|
|
if model_config.load_format == "dummy":
|
|
# NOTE(woosuk): For accurate performance evaluation, we assign
|
|
# random values to the weights.
|
|
initialize_dummy_weights(model)
|
|
else:
|
|
# Load the weights from the cached or downloaded files.
|
|
model.load_weights(model_config.model, model_config.download_dir,
|
|
model_config.load_format, model_config.revision)
|
|
return model.eval()
|
|
""" |