[Perf] fix async copy for async scheduling (#4113)

### What this PR does / why we need it?
Only CPU tensors with `pin_memory=True` can be asynchronously copied to
the device. Currently, there are two instances where non-pinned CPU
tensors are being copied to the device, which will trigger synchronous
operations, reducing the expected benefits of asynchronous scheduling.

- vLLM version: v0.11.0
- vLLM main:
83f478bb19

Signed-off-by: realliujiaxu <realliujiaxu@163.com>

vllm_ascend/worker/block_table.py

@@ -100,7 +100,7 @@ class BlockTable:
         self.slot_mapping_cpu = torch.zeros(
             self.max_num_batched_tokens +
             2 * self.pcp_world_size * self.max_num_reqs,
-            dtype=torch.int64,
+            dtype=torch.int32,
             device="cpu",
             pin_memory=self.pin_memory)
         self.slot_mapping_np = self.slot_mapping_cpu.numpy()

vllm_ascend/worker/model_runner_v1.py

@@ -1866,11 +1866,10 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                 logits_indices = torch.from_numpy(
                     cu_num_tokens
                 ) * self.pcp_size - self.num_pcp_pads[:num_reqs] - 1
-                logits_indices = logits_indices.to(self.device,
-                                                   non_blocking=True)
-            else:
-                logits_indices = torch.from_numpy(cu_num_tokens - 1).to(
+                logits_indices = logits_indices.pin_memory().to(
                     self.device, non_blocking=True)
+            else:
+                logits_indices = self.query_start_loc[1:num_reqs + 1] - 1
         else:
             # Get the number of draft tokens for each request.
             # Iterate over the dictionary rather than all requests since not all