From 7901109ea5c9794a5e7e01481343235e3a042971 Mon Sep 17 00:00:00 2001 From: linhaifeng <1371675203@qq.com> Date: Wed, 21 Jan 2026 00:13:39 +0800 Subject: [PATCH] [Bugfix] Fix Off-by-one error in _num_tokens_to_min_blocks calculation (#32603) Signed-off-by: linhaifeng <1371675203@qq.com> --- tests/kernels/utils.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tests/kernels/utils.py b/tests/kernels/utils.py index ccdacf40c4..7763be0cb5 100644 --- a/tests/kernels/utils.py +++ b/tests/kernels/utils.py @@ -609,7 +609,7 @@ def _num_tokens_to_min_blocks(num_tokens: int, block_size: int) -> int: Compute the minimum number of blocks required to hold num_tokens tokens, given block_size """ - return (num_tokens + block_size) // block_size + return (num_tokens + block_size - 1) // block_size def make_empty_slot_mapping_tensor(device: torch.device | str): @@ -694,7 +694,7 @@ def make_block_tables_slot_mapping( For a sequence with num_tokens tokens the minimum number of required KV cache blocks is - num_blocks = (num_tokens + block_size) // block_size + num_blocks = (num_tokens + block_size - 1) // block_size Then the minimum KV cache size in blocks is