[Feature] support aclgraph for model runner v2 (#7110)

### What this PR does / why we need it?
This PR aims to support aclgraph for model runner v2, please see RFC
#5208. The PR contains these modifications:
- adapt to newest commit of vllm main branch.
- supply a unified interface of extra forward context for both model
runner v1 and model runner v2.
- implement graph mode for main model. 

### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?

- vLLM version: v0.16.0
- vLLM main:
4034c3d32e

---------

Signed-off-by: Ronald1995 <ronaldautomobile@163.com>

vllm_ascend/worker/v2/utils.py

@@ -1,6 +1,11 @@
 from contextlib import contextmanager
 
 import torch
+import vllm
+from vllm.logger import logger
+
+from vllm_ascend.worker.v2.block_table import AscendBlockTables
+from vllm_ascend.worker.v2.model_states import init_asecnd_model_state
 
 
 @contextmanager
@@ -15,6 +20,34 @@ def torch_cuda_wrapper():
         torch.cuda.CUDAGraph = torch.npu.NPUGraph
         torch.cuda.graph = torch.npu.graph
         torch.cuda.synchronize = torch.npu.synchronize
+        torch.cuda.set_stream = torch.npu.set_stream
+        torch.cuda.current_device = torch.npu.current_device
+        torch.cuda.mem_get_info = torch.npu.mem_get_info
+        logger.info_once("Wrapping torch.cuda with torch.npu.")
+        yield
+    finally:
+        pass
+
+
+@contextmanager
+def block_table_wrapper():
+    try:
+        # vllm-ascend need to initialize slot mapping as torch.int32 dtype,
+        # but vllm default is torch.int64 dtype.
+        vllm.v1.worker.gpu.model_runner.BlockTables = AscendBlockTables
+        logger.info_once("Wrapping BlockTables with AscendBlockTables.")
+        yield
+    finally:
+        pass
+
+
+@contextmanager
+def model_states_wrapper():
+    try:
+        # prepare_attn in AscendModelState is different from vllm,
+        # we need to override init_model_state.
+        vllm.v1.worker.gpu.model_runner.init_model_state = init_asecnd_model_state
+        logger.info_once("Wrapping init_model_state with init_asecnd_model_state.")
         yield
     finally:
         pass