[NPU] Fix baichuan2-13b-chat infer

PaddlePaddle · Sep 2, 2024 · cffca66 · cffca66
1 parent e204b6d
commit cffca66
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 3 deletions.
diff --git a/paddlenlp/experimental/transformers/fused_transformer_layers.py b/paddlenlp/experimental/transformers/fused_transformer_layers.py
@@ -38,8 +38,7 @@
         "The paddlenlp_ops package is not installed. you can read the docs and install it by hand, "
         "you can refer to: https://github.com/PaddlePaddle/PaddleNLP/blob/develop/csrc/README.md"
     )
-if core.is_compiled_with_xpu() or core.is_compiled_with_cuda():
-    from paddlenlp_ops import rebuild_padding_v2
+from paddlenlp_ops import rebuild_padding_v2
 
 if core.is_compiled_with_cuda():
     from paddlenlp_ops import (

diff --git a/paddlenlp/utils/llm_utils.py b/paddlenlp/utils/llm_utils.py
@@ -461,7 +461,7 @@ def get_alibi_slopes(num_heads):
         extra_base = 2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3)))
         num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2)
         extra_powers = np.arange(1, 1 + 2 * num_remaining_heads, 2)
-        slopes = np.concatante([slopes, np.power(extra_base, extra_powers)], axis=0)
+        slopes = np.concatenate([slopes, np.power(extra_base, extra_powers)], axis=0)
 
     return slopes.astype("float32")