ValueError: Double free! PhysicalTokenBlock(device=Device.GPU, block_number=1859, ref_count=0) is already freed.

[DHOFM]

Using Beam Search this Error can occur, if it is a large prompt. Workaround: in VLLM Core block_manager.py I returned 0 instead of the exception and beam search works:

def free(self, block: PhysicalTokenBlock) -> None: if block.ref_count == 0: return 0 raise ValueError(f"Double free! {block} is already freed.") block.ref_count -= 1 if block.ref_count == 0: self.free_blocks.append(block)