From 7938d121196d87a710b93e55ce4cefa6a11bedd8 Mon Sep 17 00:00:00 2001 From: DorBernsohn <47590570+DorBernsohn@users.noreply.github.com> Date: Mon, 23 Mar 2026 13:35:45 +0200 Subject: [PATCH] [Bugfix] Fix CPU backend crash in KV cache block zeroing (#37550) Signed-off-by: DorBernsohn --- vllm/v1/worker/cpu_model_runner.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/vllm/v1/worker/cpu_model_runner.py b/vllm/v1/worker/cpu_model_runner.py index a945aec390..4015b7722c 100644 --- a/vllm/v1/worker/cpu_model_runner.py +++ b/vllm/v1/worker/cpu_model_runner.py @@ -88,6 +88,11 @@ class CPUModelRunner(GPUModelRunner): def _sync_device(self) -> None: pass + def _zero_block_ids(self, block_ids: list[int]) -> None: + # CPU attention assigns -INF to logits at invalid positions, + # so stale KV cache data never affects computation. + pass + def get_dp_padding(self, num_tokens: int) -> tuple[int, torch.Tensor | None]: # Note: For CPU backend, dp padding is not required for now. return 0, None