[Bugfix] Fix level-2 sleep/wake/reload with enable_lora=True (#39935)

Signed-off-by: Silen Naihin <silen.naihin@gmail.com>
Co-authored-by: Dakai An <dakaian108@gmail.com>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
This commit is contained in:
Silen Naihin
2026-08-06 09:13:10 +08:00
committed by GitHub
co-authored by Dakai An Jiangyun Zhu
parent 8cfa01cdd2
commit b50fdebce0
11 changed files with 252 additions and 4 deletions
+18 -3
View File
@@ -56,7 +56,7 @@ from vllm.forward_context import (
set_forward_context,
)
from vllm.logger import init_logger
from vllm.lora.layers import LoRAMapping, LoRAMappingType
from vllm.lora.layers import BaseLayerWithLoRA, LoRAMapping, LoRAMappingType
from vllm.model_executor.layers.attention import Attention, MLAAttention
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.model_executor.layers.fused_moe.all2all_utils import get_ep_all2all_manager
@@ -250,6 +250,16 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
def _get_parameter_for_reload(model: nn.Module, name: str) -> nn.Parameter:
"""Resolve checkpoint names without changing the model's module tree."""
module_name, _, parameter_name = name.rpartition(".")
module = model.get_submodule(module_name)
if isinstance(module, BaseLayerWithLoRA):
module = module.base_layer
return module.get_parameter(parameter_name)
AttnMetadataDict: TypeAlias = dict[str, AttentionMetadata]
# list when ubatching is enabled
PerLayerAttnMetadata: TypeAlias = list[AttnMetadataDict] | AttnMetadataDict
@@ -5535,7 +5545,10 @@ class GPUModelRunner(
)
model = self.get_model()
weights_to_load = {name for name, _ in model.named_parameters()}
weights_to_load = {
name.replace(".base_layer.", ".") if self.lora_config else name
for name, _ in model.named_parameters()
}
counter_before_reloading = time.perf_counter()
# load weights from disk if none are provided
@@ -5572,10 +5585,12 @@ class GPUModelRunner(
)
loaded_weights = set()
for name, loaded_weight in weights_iterator:
param = model.get_parameter(name) # TODO: buffers?
param = _get_parameter_for_reload(model, name) # TODO: buffers?
param.copy_(loaded_weight)
loaded_weights.add(name)
self.reset_lora_state()
# logging and validation
counter_after_reloading = time.perf_counter()
diff_seconds = counter_after_reloading - counter_before_reloading