| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import torch.nn as nn |
| 4 | |
| 5 | from vllm.config import ModelConfig |
| 6 | from vllm.config.load import LoadConfig |
| 7 | from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase |
| 8 | from vllm.model_executor.model_loader.base_loader import BaseModelLoader |
| 9 | from vllm.model_executor.model_loader.reload.layerwise import ( |
| 10 | _get_original_loader, |
| 11 | get_layerwise_info, |
| 12 | ) |
| 13 | from vllm.model_executor.model_loader.reload.meta import materialize_layer |
| 14 | from vllm.model_executor.model_loader.reload.types import LayerReloadingInfo |
| 15 | from vllm.model_executor.model_loader.reload.utils import get_layer_tensors |
| 16 | from vllm.model_executor.model_loader.weight_utils import ( |
| 17 | initialize_single_dummy_weight, |
| 18 | ) |
| 19 | |
| 20 | |
| 21 | class DummyModelLoader(BaseModelLoader): |
| 22 | """Model loader that will set model weights to random values.""" |
| 23 | |
| 24 | def __init__(self, load_config: LoadConfig): |
| 25 | super().__init__(load_config) |
| 26 | if load_config.model_loader_extra_config: |
| 27 | raise ValueError( |
| 28 | f"Model loader extra config is not supported for " |
| 29 | f"load format {load_config.load_format}" |
| 30 | ) |
| 31 | |
| 32 | def download_model(self, model_config: ModelConfig) -> None: |
| 33 | pass # Nothing to download |
| 34 | |
| 35 | def load_weights(self, model: nn.Module, model_config: ModelConfig) -> None: |
| 36 | for layer in model.modules(): |
| 37 | info = get_layerwise_info(layer) |
| 38 | if info.can_load(): |
| 39 | self._process_online_quant_layer(layer, info) |
| 40 | else: |
| 41 | # NOTE(woosuk): For accurate performance evaluation, we assign |
| 42 | # random values to the weights. |
| 43 | # Initialize only local tensors: state_dict() would recursively |
| 44 | # initialize descendants that model.modules() also visits. |
| 45 | # Exclude nonpersistent buffers, matching state_dict(). |
| 46 | for name, tensor in get_layer_tensors(layer).items(): |
| 47 | if name not in layer._non_persistent_buffers_set: |
| 48 | initialize_single_dummy_weight(tensor) |
| 49 | |
| 50 | def _process_online_quant_layer( |
| 51 | self, |
| 52 | layer: nn.Module, |
| 53 | info: LayerReloadingInfo, |
| 54 | ) -> None: |
| 55 | """Materialize, apply dummy weights, and run quantization processing.""" |
| 56 | materialize_layer(layer, info) |
| 57 | |
| 58 | for tensor in get_layer_tensors(layer).values(): |
| 59 | initialize_single_dummy_weight(tensor) |
| 60 | |
| 61 | for param in get_layer_tensors(layer).values(): |
| 62 | param.weight_loader = _get_original_loader(param) |
| 63 | |
| 64 | quant_method = getattr(layer, "quant_method", None) |
| 65 | if isinstance(quant_method, QuantizeMethodBase): |
| 66 | quant_method.process_weights_after_loading(layer) |
| 67 | |
| 68 | info.reset() |