From 095adf1fdc93d5554a3afbf26e97b399309c80db Mon Sep 17 00:00:00 2001 From: Yifan Qiao Date: Mon, 6 Jul 2026 07:36:15 -0700 Subject: [PATCH] [Bugfix] Fix int32 overflow in triton_decode_attention page offsets (#47671) Signed-off-by: Yifan Qiao --- vllm/v1/attention/ops/triton_decode_attention.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/v1/attention/ops/triton_decode_attention.py b/vllm/v1/attention/ops/triton_decode_attention.py index dbe3c5705d..6aec1db2e5 100644 --- a/vllm/v1/attention/ops/triton_decode_attention.py +++ b/vllm/v1/attention/ops/triton_decode_attention.py @@ -133,7 +133,7 @@ def _fwd_kernel_stage1( + offs_n // PAGE_SIZE, mask=offs_n < split_kv_end, other=0, - ) + ).to(tl.int64) # page_number * page stride overflows int32 kv_in_page = offs_n % PAGE_SIZE offs_buf_k = ( (kv_page_number * stride_buf_kpbs + kv_in_page * stride_buf_kbs)[ @@ -375,7 +375,7 @@ def _fwd_grouped_kernel_stage1( mask=offs_n < split_kv_end, other=0, cache_modifier=".ca", - ) + ).to(tl.int64) # page_number * page stride overflows int32 kv_off_k = ( kv_page_number * stride_buf_kpbs + (offs_n % PAGE_SIZE) * stride_buf_kbs )