[Core][Bookkeeping Optimization] Update against numpy view of is_token_ids tensor (#27618)

Signed-off-by: Jialin Ouyang <[email protected]>

Author: Jialin

vllm/v1/worker/gpu_input_batch.py

@@ -108,9 +108,10 @@ def __init__(
             pin_memory=False,
         )
         self.token_ids_cpu = self.token_ids_cpu_tensor.numpy()
-        self.is_token_ids = torch.zeros(
+        self.is_token_ids_tensor = torch.zeros(
             (max_num_reqs, max_model_len), device="cpu", dtype=bool, pin_memory=False
         )
+        self.is_token_ids = self.is_token_ids_tensor.numpy()
         # Store prompt embeddings per request to avoid OOM from large upfront
         # allocation if max_model_len is big.
         # Maps req_index -> tensor of shape (num_prompt_tokens, hidden_size)

vllm/v1/worker/gpu_model_runner.py

@@ -1103,7 +1103,7 @@ def _prepare_inputs(
             out=self.input_ids.cpu[:total_num_scheduled_tokens],
         )
         if self.enable_prompt_embeds:
-            is_token_ids = self.input_batch.is_token_ids.flatten()
+            is_token_ids = self.input_batch.is_token_ids_tensor.flatten()
             torch.index_select(
                 is_token_ids,
                 0,