Add fused top-K softmax kernel for MoE (#2769)

csrc/pybind.cpp

@@ -48,8 +48,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &rotary_embedding,
     "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
 
-#ifndef USE_ROCM
   // Quantization ops
+#ifndef USE_ROCM
   ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
   ops.def("awq_dequantize", &awq_dequantize, "Dequantization for AWQ");
 #endif