support torchair mode (#2641)

### What this PR does / why we need it?
support torchair mode
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?

- vLLM version: v0.10.1.1
- vLLM main:
5438967fbc

Signed-off-by: zhangdepeng <zhangdepeng2@huawei.com>
Signed-off-by: p00465316 <panchao13@huawei.com>
Co-authored-by: zhangdepeng <zhangdepeng2@huawei.com>

vllm_ascend/ascend_config.py

@@ -70,6 +70,7 @@ class TorchairGraphConfig:
 
     def __init__(self, torchair_graph_config):
         self.enabled = torchair_graph_config.get("enabled", False)
+        self.mode = torchair_graph_config.get("mode", '')
         self.use_cached_graph = torchair_graph_config.get(
             "use_cached_graph", False)
         self.graph_batch_sizes = torchair_graph_config.get(
@@ -91,6 +92,9 @@ class TorchairGraphConfig:
                 "graph_batch_sizes_init is only valid when graph_batch_sizes is empty"
             )
         if not self.enabled:
+            if self.mode:
+                raise RuntimeError(
+                    "mode is valid only when Torchair graph mode is enabled")
             if self.use_cached_graph:
                 raise RuntimeError(
                     "use_cached_graph is valid only when Torchair graph mode is enabled"

vllm_ascend/torchair/torchair_model_runner.py

@@ -324,6 +324,8 @@ class NPUTorchairModelRunner(NPUModelRunner):
             communication_adaptation_310p()
 
         config = torchair.CompilerConfig()
+        if get_ascend_config().torchair_graph_config.mode:
+            config.mode = get_ascend_config().torchair_graph_config.mode
         config.experimental_config.frozen_parameter = True
         # enabling tiling_schedule_optimize on 300I Duo has some bugs, so we have to
         # disable it on 300I Duo platform now.