adapt to main2main for model runner v2 (#7578)

### What this PR does / why we need it?
This PR aims to adapt to newest commit of vllm main branch for model
runner v2. please refer to
https://github.com/vllm-project/vllm-ascend/issues/5208
### Does this PR introduce _any_ user-facing change?
no

### How was this patch tested?

- vLLM version: v0.18.0
- vLLM main:
ed359c497a

---------

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

vllm_ascend/worker/v2/model_states/default.py

@@ -20,6 +20,7 @@
 from typing import Any
 
 import torch
+from vllm.config.compilation import CUDAGraphMode
 from vllm.v1.kv_cache_interface import KVCacheConfig
 from vllm.v1.worker.gpu.model_states.default import DefaultModelState
 from vllm.v1.worker.utils import AttentionGroup
@@ -34,18 +35,28 @@ class AscendModelState(DefaultModelState):
     def prepare_attn(
         self,
         input_batch: AscendInputBatch,
+        cudagraph_mode: CUDAGraphMode,
         block_tables: tuple[torch.Tensor, ...],
         slot_mappings: torch.Tensor,
         attn_groups: list[list[AttentionGroup]],
         kv_cache_config: KVCacheConfig,
+        for_capture: bool = False,
     ) -> dict[str, Any]:
         """Override prepare_attn method because `build_attn_metadata` is different from vllm."""
+        if cudagraph_mode == CUDAGraphMode.FULL:
+            # Use padded sizes - padding is handled by model_runner.prepare_attn.
+            num_reqs = input_batch.num_reqs_after_padding
+            num_tokens = input_batch.num_tokens_after_padding
+        else:
+            # For piecewise cudagraphs and eager, use unpadded sizes.
+            num_reqs = input_batch.num_reqs
+            num_tokens = input_batch.num_tokens
         query_start_loc_cpu = torch.from_numpy(input_batch.query_start_loc_np)
         max_query_len = input_batch.num_scheduled_tokens.max().item()
         attn_metadata = build_attn_metadata(
             attn_groups=attn_groups,
-            num_reqs=input_batch.num_reqs,
-            num_tokens=input_batch.num_tokens,
+            num_reqs=num_reqs,
+            num_tokens=num_tokens,
             query_start_loc_gpu=input_batch.query_start_loc,
             query_start_loc_cpu=query_start_loc_cpu,
             max_query_len=max_query_len,