[Quantization] Consolidate dummy format logic into DummyModelLoader (#38637)

Signed-off-by: Josephasafg <ajgard7@gmail.com>
This commit is contained in:
Asaf Gardin
2026-03-31 22:20:45 +00:00
committed by GitHub
parent cc671cb110
commit 3dc01ef352
2 changed files with 42 additions and 15 deletions
@@ -4,8 +4,19 @@ import torch.nn as nn
from vllm.config import ModelConfig
from vllm.config.load import LoadConfig
from vllm.model_executor.layers.quantization.base_config import QuantizeMethodBase
from vllm.model_executor.model_loader.base_loader import BaseModelLoader
from vllm.model_executor.model_loader.weight_utils import initialize_dummy_weights
from vllm.model_executor.model_loader.reload.layerwise import (
_get_original_loader,
get_layerwise_info,
)
from vllm.model_executor.model_loader.reload.meta import materialize_layer
from vllm.model_executor.model_loader.reload.types import LayerReloadingInfo
from vllm.model_executor.model_loader.reload.utils import get_layer_tensors
from vllm.model_executor.model_loader.weight_utils import (
initialize_dummy_weights,
initialize_single_dummy_weight,
)
class DummyModelLoader(BaseModelLoader):
@@ -23,6 +34,31 @@ class DummyModelLoader(BaseModelLoader):
pass # Nothing to download
def load_weights(self, model: nn.Module, model_config: ModelConfig) -> None:
# NOTE(woosuk): For accurate performance evaluation, we assign
# random values to the weights.
initialize_dummy_weights(model, model_config)
for layer in model.modules():
info = get_layerwise_info(layer)
if info.can_load():
self._process_online_quant_layer(layer, info)
else:
# NOTE(woosuk): For accurate performance evaluation, we assign
# random values to the weights.
initialize_dummy_weights(layer, model_config)
def _process_online_quant_layer(
self,
layer: nn.Module,
info: LayerReloadingInfo,
) -> None:
"""Materialize, apply dummy weights, and run quantization processing."""
materialize_layer(layer, info)
for tensor in get_layer_tensors(layer).values():
initialize_single_dummy_weight(tensor)
for param in get_layer_tensors(layer).values():
param.weight_loader = _get_original_loader(param)
quant_method = getattr(layer, "quant_method", None)
if isinstance(quant_method, QuantizeMethodBase):
quant_method.process_weights_after_loading(layer)
info.reset()