[NPU] fix conflict between empty_cache and use_mem_pool (#21507)

python/sglang/srt/model_executor/model_runner.py

@@ -1108,6 +1108,10 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                 self.remote_instance_transfer_engine_weight_info = (
                     self.loader.remote_instance_transfer_engine_weight_info
                 )
+        # Cache needs to be cleared after loading model weights (in the self.loader.load_model function).
+        # To avoid conflict with memory_saver_adapter.region, empty_cache operation is now moved here.
+        if _is_npu:
+            torch.npu.empty_cache()
         monkey_patch_vllm_parallel_state(reverse=True)
 
         # Publish metadata to ModelExpress if running as seed source

python/sglang/srt/model_loader/loader.py

@@ -707,8 +707,6 @@ class DefaultModelLoader(BaseModelLoader):
                 # parameters onto device for processing and back off after.
                 with device_loading_context(module, target_device):
                     quant_method.process_weights_after_loading(module)
-                if _is_npu:
-                    torch.npu.empty_cache()
 
 
 class LayeredModelLoader(DefaultModelLoader):

python/sglang/srt/utils/common.py

@@ -365,7 +365,7 @@ def get_int_env_var(name: str, default: int = 0) -> int:
 
 
 def support_triton(backend: str) -> bool:
-    return backend not in ["torch_native", "intel_amx"]
+    return backend not in ["torch_native", "intel_amx", "ascend"]
 
 
 _ENABLE_TORCH_INFERENCE_MODE = get_bool_env_var(