A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 04:30 UTC
[Bugfix] Fix CPU backend crash in KV cache block zeroing (#37550)
Signed-off-by: DorBernsohn <dor.bernsohn@gmail.com>
DorBernsohn committed
Mar 23, 2026 at 13:35 UTC
7938d121196d87a710b93e55ce4cefa6a11bedd8
1 file changed
+5
vllm/v1/worker/cpu_model_runner.py
+5
@@ -88,6 +88,11 @@ class CPUModelRunner(GPUModelRunner):
88
def _sync_device(self) -> None:
89
pass
90
91
+ def _zero_block_ids(self, block_ids: list[int]) -> None:
92
+ # CPU attention assigns -INF to logits at invalid positions,
93
+ # so stale KV cache data never affects computation.
94
+ pass
95
+
96
def get_dp_padding(self, num_tokens: int) -> tuple[int, torch.Tensor | None]:
97
# Note: For CPU backend, dp padding is not required for now.
98
return 0, None