upgrade torch npu version (#4433)

vLLM graph feature now rely on torch >=2.8. To make graph mode work, we
need upgrade torch version as well. For long term support, upgrade torch
to a newer one is good to go as well.

Related vLLM change: https://github.com/vllm-project/vllm/pull/25110


- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

tests/e2e/singlecard/test_bge_model.py

@@ -28,7 +28,7 @@ def test_bge_model_correctness():
     model_name = snapshot_download("BAAI/bge-m3")
     with VllmRunner(
             model_name,
-            task="embed",
+            runner="pooling",
             enforce_eager=True,
     ) as vllm_runner:
         vllm_outputs = vllm_runner.encode(queries)

tests/e2e/singlecard/test_embedding.py

@@ -28,7 +28,7 @@ def test_embed_models_correctness():
     model_name = snapshot_download("Qwen/Qwen3-Embedding-0.6B")
     with VllmRunner(
             model_name,
-            task="embed",
+            runner="pooling",
             enforce_eager=False,
     ) as vllm_runner:
         vllm_outputs = vllm_runner.encode(queries)

tests/e2e/singlecard/test_embedding_aclgraph.py

@@ -34,14 +34,14 @@ def test_aclgrpah_embed_models_correctness(model_name):
 
     with VllmRunner(
             model_name,
-            task="embed",
+            runner="pooling",
             enforce_eager=False,
     ) as vllm_aclgraph_runner:
         vllm_aclgraph_outputs = vllm_aclgraph_runner.encode(queries)
 
     with VllmRunner(
             model_name,
-            task="embed",
+            runner="pooling",
             enforce_eager=True,
     ) as vllm_runner:
         vllm_outputs = vllm_runner.encode(queries)