biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Extend collect_children and no_init_weights contexts (#32757)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-22 16:20:27 +08:00
committed by GitHub
parent 1bf1a34b19
commit 2b8a38b6d6
20 changed files with 444 additions and 257 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm/model_executor/model_loader/utils.py
View File

@@ -18,7 +18,7 @@ from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig,
QuantizeMethodBase,
)
from vllm.model_executor.models.interfaces import SupportsQuant, supports_multimodal
from vllm.model_executor.models.interfaces import SupportsQuant
from vllm.utils.platform_utils import is_pin_memory_available
logger = init_logger(__name__)
@@ -165,11 +165,7 @@ _MODEL_ARCH_BY_HASH = dict[int, tuple[type[nn.Module], str]]()
def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module], str]:
from vllm.model_executor.models.adapters import (
as_embedding_model,
as_seq_cls_model,
try_create_mm_pooling_model_cls,
)
from vllm.model_executor.models.adapters import as_embedding_model, as_seq_cls_model
architectures = getattr(model_config.hf_config, "architectures", [])
@@ -189,15 +185,6 @@ def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module],
)
convert_type = model_config.convert_type
if convert_type != "none" and supports_multimodal(model_cls):
logger.debug_once("Detected conversion of Multi Modal model.")
converted = try_create_mm_pooling_model_cls(model_cls)
if converted is not None:
logger.debug_once("Creating wrapper class to forward pooler.")
return converted, arch
else:
logger.debug_once("Attempting direct conversion.")
if convert_type == "none":
pass
elif convert_type == "embed":