[V1] Remove V0 code paths for Hybrid models (#25400)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2025-09-23 08:26:13 -07:00
committed by GitHub
parent 2c58742dff
commit a903669e10
31 changed files with 352 additions and 2296 deletions
+1 -4
View File
@@ -20,10 +20,7 @@ class MambaBase(AttentionLayerBase):
# Contains the KV cache (mamba state) for the layer
# in the shape specified by `self.get_state_shape`.
# The outer list is for v0 PP virtual engine. Though this code path
# only runs for v1, we have to do this to unify with the interface
# of Attention + v0 PP.
kv_cache: list[Iterable[torch.Tensor]]
kv_cache: tuple[torch.Tensor, ...]
@abstractmethod
def get_state_shape(self) -> Iterable[tuple[int, ...]]: