From a1f9508dd41006672710eda29528cccb34d4a52e Mon Sep 17 00:00:00 2001 From: Benjamin Truong Date: Tue, 25 Aug 2026 19:56:42 +0700 Subject: [PATCH] [Bugfix] Fix int32 destination offset overflow in CUTLASS MoE pre-reorder (#35188) --- python/sglang/kernels/ops/moe/ep_moe_kernels.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/kernels/ops/moe/ep_moe_kernels.py b/python/sglang/kernels/ops/moe/ep_moe_kernels.py index 938b82951..9268572e3 100644 --- a/python/sglang/kernels/ops/moe/ep_moe_kernels.py +++ b/python/sglang/kernels/ops/moe/ep_moe_kernels.py @@ -241,7 +241,7 @@ def pre_reorder_triton_kernel_for_cutlass_moe( for idx in range(topk): expert_id = tl.load(token_topk_ids_ptr + idx) if expert_id != num_local_experts: - dst_idx = tl.load(token_src2dst_ptr + idx) + dst_idx = tl.load(token_src2dst_ptr + idx).to(tl.int64) tl.store(dst_ptr_offs + dst_idx * hidden_size, out_data, mask=mask)