diff --git a/cutedsl/moe_pipeline.py b/cutedsl/moe_pipeline.py index 52bbf6a8..49717d39 100644 --- a/cutedsl/moe_pipeline.py +++ b/cutedsl/moe_pipeline.py @@ -200,9 +200,8 @@ def run_nvfp4_moe( # SiLU(gate) * up (BF16 — nonlinear requires BF16) # ════════════════════════════════════════════════════════════════ intermediate = l1_out.shape[1] - half = intermediate // 2 # 3072 - gate = l1_out[:, :half] - up = l1_out[:, half:] + gate = l1_out[:, :intermediate] + up = l1_out[:, intermediate:] activated = torch.nn.functional.silu(gate) * up # (num_slots, half) BF16 # ════════════════════════════════════════════════════════════════