Usage Stats Collection (#2852)

This commit is contained in:
yhu422
2024-03-28 22:16:12 -07:00
committed by GitHub
parent 7bc94a0fdd
commit d8658c8cc1
15 changed files with 362 additions and 24 deletions
+9 -4
View File
@@ -1,6 +1,6 @@
"""Utilities for selecting and loading models."""
import contextlib
from typing import Type
from typing import Tuple, Type
import torch
import torch.nn as nn
@@ -25,7 +25,8 @@ def _set_default_torch_dtype(dtype: torch.dtype):
torch.set_default_dtype(old_dtype)
def _get_model_architecture(model_config: ModelConfig) -> Type[nn.Module]:
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.
@@ -36,17 +37,21 @@ def _get_model_architecture(model_config: ModelConfig) -> Type[nn.Module]:
for arch in architectures:
model_cls = ModelRegistry.load_model_cls(arch)
if model_cls is not None:
return model_cls
return (model_cls, arch)
raise ValueError(
f"Model architectures {architectures} are not supported for now. "
f"Supported architectures: {ModelRegistry.get_supported_archs()}")
def get_architecture_class_name(model_config: ModelConfig) -> str:
return _get_model_architecture(model_config)[1]
def get_model(model_config: ModelConfig, device_config: DeviceConfig,
**kwargs) -> nn.Module:
lora_config = kwargs.get("lora_config", None)
vision_language_config = kwargs.get("vision_language_config", None)
model_class = _get_model_architecture(model_config)
model_class = _get_model_architecture(model_config)[0]
# Get the (maybe quantized) linear method.
linear_method = None