BlockManager.can_append may be incorrect.

[PeterDing]

nano-vllm/nanovllm/engine/block_manager.py

Lines 94 to 95 in cde3fc2

	def can_append(self, seq: Sequence) -> bool:
	return len(self.free_block_ids) >= (len(seq) % self.block_size == 1)

It should be:

    def can_append(self, seq: Sequence) -> bool:
        return len(self.free_block_ids) >= ((len(seq) + self.block_size - 1) // self.block_size)

[GeeeekExplorer]

The method only requires a new block allocation when the sequence length reaches exactly 1 element past a block boundary (i.e., when len(seq) % block_size == 1). In all other cases, the sequence can continue using the currently allocated blocks.

[xiaohajiayou]

https://github.com/GeeeekExplorer/nano-vllm/blob/cde3fc22c2694ba8052f7df0b5e05ef7e10c8603/nanovllm/engine/block_manager.py#L97toL113
I initially thought there might be an issue here, but upon closer inspection, the logic in this code needs to be considered in conjunction with the may_append function. Cases where the sequence length exactly matches the block boundary also need to enter this function for hash computation.