remove CudaUtils dependencies

Signed-off-by: Jinjie Liu <jjliu@baai.ac.cn>

python/triton_tvm_ffi/driver.py

@@ -1,8 +1,9 @@
 from __future__ import annotations
 
-from typing import Mapping, Type
+from ctypes import c_void_p
+from typing import Mapping, Tuple, Type
 from triton.backends.nvidia.driver import CudaDriver
-from .utils import get_device_properties
+from .utils import get_device_properties, load_binary
 
 
 class TVMFFIUtils(object):
@@ -13,12 +14,11 @@ class TVMFFIUtils(object):
 
     def __init__(self, *args, **kwargs) -> None:
         super().__init__(*args, **kwargs)
-        from triton.backends.nvidia.driver import CudaUtils
 
-        self._utils: CudaUtils = CudaUtils()
-
-    def load_binary(self, *args, **kwargs):
-        return self._utils.load_binary(*args, **kwargs)
+    def load_binary(
+        self, name: str, data: bytes, shared: int, device: int
+    ) -> Tuple[c_void_p, c_void_p, int, int, int]:
+        return load_binary(name, data, shared, device)
 
     def get_device_properties(self, device_id: int) -> Mapping[str, int]:
         return get_device_properties(device_id)
@@ -34,13 +34,19 @@ class TVMFFIUtils(object):
         )
 
     def fill_tma_descriptor(self, *args, **kwargs):
-        return self._utils.fill_tma_descriptor(*args, **kwargs)
+        raise NotImplementedError(
+            '"fill_tma_descriptor" hasn\'t been supported for Hopper'
+        )
 
     def launch(self, *args, **kwargs):
-        return self._utils.launch(*args, **kwargs)
+        raise NotImplementedError(
+            '"launch" is introduced in triton after commit d2b3925410689155e0f6028e8554bba972989348, which is still not supported yed'
+        )
 
     def build_signature_metadata(self, *args, **kwargs):
-        return self._utils.build_signature_metadata(*args, **kwargs)
+        raise NotImplementedError(
+            '"launch" is introduced in triton after commit d2b3925410689155e0f6028e8554bba972989348, which is still not supported yed'
+        )
 
 
 class TVMFFIDriver(CudaDriver):

python/triton_tvm_ffi/utils/_ffi_api.py

@@ -17,6 +17,7 @@ LIB = _FFI_LOAD_LIB("triton_tvm_ffi", "utils")
 _FFI_INIT_FUNC("triton_tvm_ffi.utils", __name__)
 if TYPE_CHECKING:
     def get_device_properties(_0: int, /) -> Mapping[str, int]: ...
+    def load_binary(_0: str, _1: bytes, _2: int, _3: int, /) -> tuple[int, int, int, int, int]: ...
 # fmt: on
 # tvm-ffi-stubgen(end)
 
@@ -24,5 +25,6 @@ __all__ = [
     # tvm-ffi-stubgen(begin): __all__
     "LIB",
     "get_device_properties",
+    "load_binary",
     # tvm-ffi-stubgen(end)
 ]