From 8c742a66d1bdb92978a40f0420123bf0d07fd90d Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Sun, 31 Aug 2025 21:02:43 -0700 Subject: [PATCH] [Misc] Avoid redundant copy for encoder-only models (#24012) Signed-off-by: Woosuk Kwon --- vllm/v1/worker/gpu_model_runner.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/vllm/v1/worker/gpu_model_runner.py b/vllm/v1/worker/gpu_model_runner.py index f77373e8adaae..4a6856bf4fefc 100644 --- a/vllm/v1/worker/gpu_model_runner.py +++ b/vllm/v1/worker/gpu_model_runner.py @@ -827,13 +827,13 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin): blk_table_tensor = torch.zeros( (num_reqs, 1), dtype=torch.int32, - pin_memory=self.pin_memory, - device="cpu").to(self.device, non_blocking=True) - slot_mapping = torch.zeros((total_num_scheduled_tokens, ), - dtype=torch.int32, - pin_memory=self.pin_memory, - device="cpu").to(self.device, - non_blocking=True) + device=self.device, + ) + slot_mapping = torch.zeros( + (total_num_scheduled_tokens, ), + dtype=torch.int64, + device=self.device, + ) num_common_prefix_blocks = 0 else: blk_table = self.input_batch.block_table[kv_cache_group_id]