diff --git a/vllm/lora/ops/triton_ops/fused_moe_lora_op.py b/vllm/lora/ops/triton_ops/fused_moe_lora_op.py
index 771035691..9376b4e6d 100644
--- a/vllm/lora/ops/triton_ops/fused_moe_lora_op.py
+++ b/vllm/lora/ops/triton_ops/fused_moe_lora_op.py
@@ -139,7 +139,9 @@ def _fused_moe_lora_kernel(
     offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M).to(tl.int64)
     token_ind = stride_tl * lora_id + offs_token_id
     offs_token = tl.load(
-        sorted_token_ids_ptr + token_ind, token_ind < max_loras * stride_tl, 0
+        sorted_token_ids_ptr + token_ind,
+        mask=token_ind < max_loras * stride_tl,
+        other=num_valid_tokens,
     )
     token_mask = offs_token < num_valid_tokens
 
@@ -185,7 +187,7 @@ def _fused_moe_lora_kernel(
         b_ptrs += BLOCK_SIZE_K * SPLIT_K * stride_bk
 
     if MUL_ROUTED_WEIGHT:
-        moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0)
+        moe_weight = tl.load(topk_weights_ptr + offs_token, mask=token_mask, other=0.0)
         accumulator = accumulator * moe_weight[:, None]
     accumulator = accumulator.to(c_ptr.dtype.element_ty)
     # Write back the block of the output