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/attention/sfa_v1.py

@@ -345,7 +345,6 @@ class AscendSFAImpl(MLAAttentionImpl):
         ascend_config = get_ascend_config()
         self.enable_shared_expert_dp = ascend_config.enable_shared_expert_dp
         self.enable_prefetch = ascend_config.weight_prefetch_config.enabled
-        self.enable_kv_nz = ascend_config.torchair_graph_config.enable_kv_nz
         self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
 
         assert self.indexer is not None, "Indexer is required for DSA."
@@ -534,7 +533,7 @@ class AscendSFAImpl(MLAAttentionImpl):
         # npu_kv_rmsnorm_rope_cache needs [B, N, S, D]
         kv_no_split = kv_no_split.view(
             B, N, S, self.kv_lora_rank + self.qk_rope_head_dim)
-        cache_mode = "PA_NZ" if self.enable_kv_nz else "PA"
+        cache_mode = "PA"
 
         if self.enable_sfa_cp:
             assert slots_cp is not None