[Misc] Add extra checking to torchair_graph_config. (#1939)

### What this PR does / why we need it?

cherry-pick #1675  to main
This PR adds validation checking to torchair_graph_config for better
reliability.

Co-authored-by: whx-sjtu <2952154980@qq.com>

### Does this PR introduce _any_ user-facing change?

No

### How was this patch tested?


- vLLM version: v0.10.0
- vLLM main:
2836dd73f1

Signed-off-by: 22dimensions <waitingwind@foxmail.com>

vllm_ascend/ascend_config.py

@@ -76,6 +76,31 @@ class TorchairGraphConfig:
             raise ValueError(
                 "graph_batch_sizes_init is only valid when graph_batch_sizes is empty"
             )
+        if not self.enabled:
+            if self.use_cached_graph:
+                raise RuntimeError(
+                    "use_cached_graph is valid only when Torchair graph mode is enabled"
+                )
+            if self.graph_batch_sizes:
+                raise RuntimeError(
+                    "graph_batch_sizes is valid only when Torchair graph mode is enabled"
+                )
+            if self.graph_batch_sizes_init:
+                raise RuntimeError(
+                    "graph_batch_sizes_init is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_multistream_mla:
+                raise RuntimeError(
+                    "enable_multistream_mla is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_multistream_moe:
+                raise RuntimeError(
+                    "enable_multistream_moe is valid only when Torchair graph mode is enabled"
+                )
+            if self.enable_kv_nz:
+                raise RuntimeError(
+                    "enable_kv_nz is valid only when Torchair graph mode is enabled"
+                )
 
 
 class AscendSchedulerConfig:

vllm_ascend/models/deepseek_v2.py

@@ -313,7 +313,8 @@ class CustomDeepseekV2MoE(nn.Module):
         ascend_config = get_ascend_config()
         self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
         self.enable_multistream_moe = \
-            ascend_config.torchair_graph_config.enable_multistream_moe
+            ascend_config.torchair_graph_config.enable_multistream_moe and \
+            self.torchair_graph_enabled
 
         self.gate = ReplicatedLinear(config.hidden_size,
                                      config.n_routed_experts,

vllm_ascend/ops/fused_moe.py

@@ -1232,7 +1232,8 @@ class AscendFusedMoE(FusedMoE):
 
         self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
         self.enable_multistream_moe = \
-            ascend_config.torchair_graph_config.enable_multistream_moe
+            ascend_config.torchair_graph_config.enable_multistream_moe and \
+            self.torchair_graph_enabled
 
         if self.scoring_func != "softmax" and not self.use_grouped_topk:
             raise ValueError("Only softmax scoring function is supported for "