[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
@@ -17,6 +17,24 @@ from vllm.v1.worker.gpu.states import RequestState
from vllm.v1.worker.utils import AttentionGroup
class ModelSpecificAttnMetadata:
"""Base class for model-specific attention metadata."""
def get_extra_common_attn_kwargs(
self,
kv_cache_group_id: int,
num_reqs: int,
) -> dict[str, Any]:
return {}
def get_extra_attn_kwargs(
self,
attn_metadata_builder: Any,
num_reqs: int,
) -> dict[str, Any]:
return {}
class ModelState(ABC):
@abstractmethod
def __init__(
@@ -38,6 +56,13 @@ class ModelState(ABC):
def apply_staged_writes(self) -> None:
return None
def postprocess_state(
self,
input_batch: InputBatch,
num_sampled: torch.Tensor,
) -> None:
return None
@abstractmethod
def get_mm_embeddings(
self,