[V1] [Hybrid] Refactor mamba state shape calculation; enable V1 via cli (#20840)

Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
This commit is contained in:
Thomas Parnell
2025-07-15 04:04:35 -07:00
committed by GitHub
parent c586b55667
commit 3534c39a20
14 changed files with 441 additions and 353 deletions
+20
View File
@@ -22,6 +22,7 @@ from .interfaces_base import is_pooling_model
if TYPE_CHECKING:
from vllm.attention import AttentionMetadata
from vllm.config import VllmConfig
from vllm.model_executor.models.utils import WeightsMapper
from vllm.sequence import IntermediateTensors
@@ -481,6 +482,25 @@ class IsHybrid(Protocol):
, also indicates that the model's hf_config has
'layers_block_type' """
@classmethod
def get_mamba_state_shape_from_config(
cls,
vllm_config: "VllmConfig",
use_v1: bool = True,
) -> tuple[tuple[int, int], tuple[int, int, int]]:
"""Calculate shapes for Mamba's convolutional and state caches.
Args:
vllm_config: vLLM config
use_v1: Get shapes for V1 (or V0)
Returns:
Tuple containing:
- conv_state_shape: Shape for convolutional state cache
- temporal_state_shape: Shape for state space model cache
"""
...
@runtime_checkable
class _IsHybridType(Protocol):