improvement(*): replace flatten with ravel to avoid unnecessary time …

…cost of data copy
triton-inference-server · Aug 2, 2023 · 89c2ae0 · 89c2ae0
1 parent f7c45d3
commit 89c2ae0
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/src/python/library/tritonclient/utils/cuda_shared_memory/__init__.py b/src/python/library/tritonclient/utils/cuda_shared_memory/__init__.py
@@ -200,7 +200,7 @@ def set_shared_memory_region(cuda_shm_handle, input_values):
 
         offset_current = 0
         for input_value in input_values:
-            input_value = np.ascontiguousarray(input_value).flatten()
+            input_value = np.ascontiguousarray(input_value).ravel()
             # When the input array is in type "BYTES" (np.object_ is
             # the numpy equivalent), expect the array has been serialized
             # via 'tritonclient.utils.serialize_byte_tensor'.

diff --git a/src/python/library/tritonclient/utils/shared_memory/__init__.py b/src/python/library/tritonclient/utils/shared_memory/__init__.py
@@ -153,7 +153,7 @@ def set_shared_memory_region(shm_handle, input_values, offset=0):
 
     offset_current = offset
     for input_value in input_values:
-        input_value = np.ascontiguousarray(input_value).flatten()
+        input_value = np.ascontiguousarray(input_value).ravel()
         if input_value.dtype == np.object_:
             input_value = input_value.item()
             byte_size = np.dtype(np.byte).itemsize * len(input_value)