Drop torchair (#4814)

aclgraph is stable and fast now. Let's drop torchair graph mode now.

TODO: some logic to adapt torchair should be cleaned up as well. We'll
do it in the following PR.

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: Mengqing Cao <cmq0113@163.com>

vllm_ascend/ops/fused_moe/moe_comm_method.py

@@ -99,8 +99,6 @@ class MoECommMethod(ABC):
             w2_scale: Optional[list[torch.Tensor]] = None,
             w1_scale_bias: torch.Tensor = None,
             w2_scale_bias: torch.Tensor = None,
-            # For TorchAir graph
-            is_torchair: bool = False,
             # For Cube/Vector parallel
             shared_experts: Optional[Any] = None,
             quantized_x_for_share: Optional[Any] = None,
@@ -283,8 +281,6 @@ class FusedAlltoAllCommImpl(MoECommMethod):
             w2_scale: Optional[torch.Tensor] = None,
             w1_scale_bias: torch.Tensor = None,
             w2_scale_bias: torch.Tensor = None,
-            # For TorchAir graph
-            is_torchair: bool = False,
             # For Cube/Vector parallel
             shared_experts: Optional[Any] = None,
             quantized_x_for_share: Optional[Any] = None,