bugfix: casting int array to int32 for rope input arguments (#697)

To avoid the potential bugs when user pass LongTensor to rope APIs.
Also remove some files that are not used in current codebase.

Fixed a bug in AOT mode that `apply_rope_pos_ids_cos_sin_cache` was not
registered in pybind.

Author: yzh119

csrc/dispatch_type_code.h

@@ -254,6 +254,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
   m.def("apply_rope_pos_ids", &apply_rope_pos_ids, "Apply RoPE with positional ids");
   m.def("apply_llama31_rope_pos_ids", &apply_llama31_rope_pos_ids,
         "Apply Llama 3.1 style RoPE with positional ids");
+  m.def("apply_rope_pos_ids_cos_sin_cache", &apply_rope_pos_ids_cos_sin_cache,
+        "Apply RoPE with positional ids and cosine/sine cache");
 
   // sampling
   m.def("sampling_from_probs", &sampling_from_probs, "Sample from probabilities");

csrc/dispatch_utils.h

@@ -56,6 +56,8 @@ def _apply_rope(
     rope_theta: float,
 ) -> None:
     with q.device as device:
+        indptr = indptr.int()
+        offsets = offsets.int()
         get_rope_module().apply_rope(
             q,
             k,
@@ -104,6 +106,8 @@ def _apply_llama31_rope(
     old_context_len: float,
 ) -> None:
     with q.device as device:
+        indptr = indptr.int()
+        offsets = offsets.int()
         get_rope_module().apply_llama31_rope(
             q,
             k,
@@ -154,6 +158,7 @@ def _apply_rope_pos_ids(
     rope_theta: float,
 ) -> None:
     with q.device as device:
+        pos_ids = pos_ids.int()
         get_rope_module().apply_rope_pos_ids(
             q,
             k,
@@ -197,6 +202,7 @@ def _apply_rope_pos_ids_cos_sin_cache(
     interleave: bool,
 ) -> None:
     with q.device as device:
+        pos_ids = pos_ids.int()
         get_rope_module().apply_rope_pos_ids_cos_sin_cache(
             q,
             k,
@@ -242,6 +248,7 @@ def _apply_llama31_rope_pos_ids(
     old_context_len: float,
 ) -> None:
     with q.device as device:
+        pos_ids = pos_ids.int()
         get_rope_module().apply_llama31_rope_pos_ids(
             q,
             k,