Add type check (fp16) in the forward pass

csrc/flash_attn/fmha_api.cpp

@@ -130,6 +130,9 @@ mha_fwd(const at::Tensor &qkv,         // total x num_heads x 3 x head_size, tot
     bool is_dropout = p_dropout > 0.0;
     Launch_params<Fused_multihead_attention_fprop_params> launch_params(dprops, stream, is_dropout, return_softmax);
 
+    TORCH_CHECK(qkv.dtype() == torch::kFloat16);
+    TORCH_CHECK(cu_seqlens.dtype() == torch::kInt32);
+
     TORCH_CHECK(qkv.is_cuda())
     TORCH_CHECK(cu_seqlens.is_cuda())