[Model Runner V2] support qwen35 / mamba hybrid model (#35520)

Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
This commit is contained in:
zhrrr
2026-05-07 09:31:05 -07:00
committed by GitHub
parent 06a60d3dd0
commit 7a08b34fbf
14 changed files with 439 additions and 106 deletions
+14 -2
View File
@@ -111,7 +111,15 @@ class MultiModalRegistry:
return False
mm_config = model_config.get_multimodal_config()
info = self._create_processing_info(model_config, tokenizer=None)
try:
info = self._create_processing_info(model_config, tokenizer=None)
except ValueError:
logger.warning_once(
"Model %s is treated as multimodal but has no registered "
"multimodal processor; running in text-only mode.",
model_config.model,
)
return False
# Check if all supported modalities have limit == 0
if all(
@@ -170,7 +178,11 @@ class MultiModalRegistry:
from vllm.model_executor.model_loader import get_model_architecture
model_cls, _ = get_model_architecture(model_config)
assert hasattr(model_cls, "_processor_factory")
if not hasattr(model_cls, "_processor_factory"):
raise ValueError(
f"Model class {model_cls.__name__} has no registered "
"multimodal processor"
)
return cast("SupportsMultiModal", model_cls)
def _create_processing_ctx(