[Core] Support inplace model weights loading (#18745)

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
This commit is contained in:
22quinn
2025-06-02 17:38:50 +08:00
committed by GitHub
parent b9f61e1387
commit 9760fd8f6a
13 changed files with 240 additions and 288 deletions
@@ -1,9 +1,12 @@
# SPDX-License-Identifier: Apache-2.0
from abc import ABC, abstractmethod
import torch
import torch.nn as nn
from vllm.config import LoadConfig, ModelConfig, VllmConfig
from vllm.model_executor.model_loader.utils import (
initialize_model, process_weights_after_loading, set_default_torch_dtype)
class BaseModelLoader(ABC):
@@ -18,7 +21,22 @@ class BaseModelLoader(ABC):
raise NotImplementedError
@abstractmethod
def load_model(self, *, vllm_config: VllmConfig,
def load_weights(self, model: nn.Module,
model_config: ModelConfig) -> None:
"""Load weights into a model. This standalone API allows
inplace weights loading for an already-initialized model"""
raise NotImplementedError
def load_model(self, vllm_config: VllmConfig,
model_config: ModelConfig) -> nn.Module:
"""Load a model with the given configurations."""
raise NotImplementedError
device_config = vllm_config.device_config
target_device = torch.device(device_config.device)
with set_default_torch_dtype(model_config.dtype):
with target_device:
model = initialize_model(vllm_config=vllm_config,
model_config=model_config)
# Quantization does not happen in `load_weights` but after it
self.load_weights(model, model_config)
process_weights_after_loading(model, model_config, target_device)
return model.eval()