[FFI][ABI] Better String and Nested Container handling (#18311)

[FFI][ABI][REFACTOR] Better String and nested container handling

This PR improves the overall String/Bytes and nested container handling
It also fixes a bug for temp object recycling when temp object.

- Introduce formal API for string/bytes creation
- Updates the tuple/dict conversion to also preserve the torch stream
   - So if a function takes a list of torch.Tensor, torch stream will be setup in context
- Optimizes recursive argument conversion by moving most logic into c++

Author: tqchen

ffi/include/tvm/ffi/c_api.h

@@ -555,6 +555,25 @@ TVM_FFI_DLL int TVMFFITensorFromDLPackVersioned(DLManagedTensorVersioned* from,
  */
 TVM_FFI_DLL int TVMFFITensorToDLPackVersioned(TVMFFIObjectHandle from,
                                               DLManagedTensorVersioned** out);
+//---------------------------------------------------------------
+// Section: string/bytes support APIs.
+// These APIs are used to simplify the string/bytes construction
+//---------------------------------------------------------------
+/*!
+ * \brief Reinterpret the content of TVMFFIByteArray to String.
+ * \param input The TVMFFIByteArray to convert.
+ * \param out The output String owned by the caller, maybe a SmallStr or a Str object.
+ * \return 0 on success, nonzero on failure.
+ */
+TVM_FFI_DLL int TVMFFIStringFromByteArray(const TVMFFIByteArray* input, TVMFFIAny* out);
+
+/*!
+ * \brief Reinterpret the content of TVMFFIByteArray to Bytes.
+ * \param input The TVMFFIByteArray to convert.
+ * \param out The output Bytes owned by the caller, maybe a SmallBytes or a Bytes object.
+ * \return 0 on success, nonzero on failure.
+ */
+TVM_FFI_DLL int TVMFFIBytesFromByteArray(const TVMFFIByteArray* input, TVMFFIAny* out);
 
 //---------------------------------------------------------------
 // Section: dtype string support APIs.

ffi/pyproject.toml

@@ -17,7 +17,7 @@
 
 [project]
 name = "apache-tvm-ffi"
-version = "0.1.0a12"
+version = "0.1.0a13"
 description = "tvm ffi"
 
 authors = [{ name = "TVM FFI team" }]

ffi/python/tvm_ffi/_convert.py

@@ -40,13 +40,9 @@ def convert(value: Any) -> Any:
     automatically converted. So this function is mainly
     only used in internal or testing scenarios.
     """
-    if isinstance(value, core.Object):
+    if isinstance(value, (core.Object, core.PyNativeObject, bool, Number)):
         return value
-    elif isinstance(value, core.PyNativeObject):
-        return value
-    elif isinstance(value, (bool, Number)):
-        return value
-    elif isinstance(value, (list, tuple)):
+    elif isinstance(value, (tuple, list)):
         return container.Array(value)
     elif isinstance(value, dict):
         return container.Map(value)
@@ -67,6 +63,3 @@ def convert(value: Any) -> Any:
     else:
         # in this case, it is an opaque python object
         return core._convert_to_opaque_object(value)
-
-
-core._set_func_convert_to_object(convert)

ffi/python/tvm_ffi/_optional_torch_c_dlpack.py

@@ -384,7 +384,6 @@ def load_torch_c_dlpack_extension():
             ],
             extra_cflags=["-O3"],
             extra_include_paths=libinfo.include_paths() + cpp_extension.include_paths("cuda"),
-            verbose=True,
         )
         # set the dlpack related flags
         torch.Tensor.__c_dlpack_from_pyobject__ = mod.TorchDLPackFromPyObjectPtr()

ffi/python/tvm_ffi/cython/base.pxi

@@ -212,6 +212,8 @@ cdef extern from "tvm/ffi/c_api.h":
                                          TVMFFIByteArray* traceback) nogil
 
     int TVMFFITypeKeyToIndex(TVMFFIByteArray* type_key, int32_t* out_tindex) nogil
+    int TVMFFIStringFromByteArray(TVMFFIByteArray* input_, TVMFFIAny* out) nogil
+    int TVMFFIBytesFromByteArray(TVMFFIByteArray* input_, TVMFFIAny* out) nogil
     int TVMFFIDataTypeFromString(TVMFFIByteArray* str, DLDataType* out) nogil
     int TVMFFIDataTypeToString(const DLDataType* dtype, TVMFFIAny* out) nogil
     const TVMFFIByteArray* TVMFFITraceback(
@@ -284,6 +286,15 @@ cdef extern from "tvm_ffi_python_helpers.h":
         DLPackToPyObject* out_dlpack_importer
     ) except -1
 
+    int TVMFFIPyConstructorCall(
+        TVMFFIPyArgSetterFactory setter_factory,
+        void* chandle,
+        PyObject* py_arg_tuple,
+        TVMFFIAny* result,
+        int* c_api_ret_code,
+        TVMFFIPyCallContext* parent_ctx
+    ) except -1
+
     int TVMFFIPyCallFieldSetter(
         TVMFFIPyArgSetterFactory setter_factory,
         TVMFFIFieldSetter field_setter,