From 620e8924d9c6b2a0b1d49ac0dcf2588fffcbe390 Mon Sep 17 00:00:00 2001 From: Yongye Zhu Date: Sun, 12 Apr 2026 20:08:08 -0400 Subject: [PATCH] [Bugfix] [Tests] Enforce `out` tensor device in `kernel/moe/test_cutedsl_moe.py` (#39644) Signed-off-by: Yongye Zhu --- tests/kernels/moe/test_cutedsl_moe.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/tests/kernels/moe/test_cutedsl_moe.py b/tests/kernels/moe/test_cutedsl_moe.py index 2a6f83695c..c3e7190ca0 100644 --- a/tests/kernels/moe/test_cutedsl_moe.py +++ b/tests/kernels/moe/test_cutedsl_moe.py @@ -142,7 +142,9 @@ def prepare_inputs( # Initialize the hidden_states_3d with ones instead of empty to avoid nan # issue. hidden_states_3d = torch.ones( - (num_experts, max(masked_m), hidden_states.shape[1]), dtype=hidden_states.dtype + (num_experts, max(masked_m), hidden_states.shape[1]), + dtype=hidden_states.dtype, + device=hidden_states.device, ) for i in range(num_experts): hidden_states_3d[i, : masked_m[i], :] = hidden_states[topk_idx.view(-1) == i] @@ -426,7 +428,7 @@ def test_flashinfer_cutedsl_moe_masked( w1_alpha = 1.0 / (input_global_scale * w1_global_scale) w2_alpha = 1.0 / (a2_global_scale * w2_global_scale) - out = torch.empty_like(hidden_states_3d) + out = torch.empty_like(hidden_states_3d, device=hidden_states.device) # Note: the 1st dim shouldn't be bs wk = torch.empty( num_experts,