From 23c9434c7694667fd3937c5114f27eb9729cc599 Mon Sep 17 00:00:00 2001 From: runzhech Date: Tue, 26 Nov 2024 11:38:29 +0800 Subject: [PATCH] [XPU] set appropriate mask value for xpu --- paddlenlp/transformers/llama/modeling.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/paddlenlp/transformers/llama/modeling.py b/paddlenlp/transformers/llama/modeling.py index 7611fd961ab6..cd69f566ebcd 100755 --- a/paddlenlp/transformers/llama/modeling.py +++ b/paddlenlp/transformers/llama/modeling.py @@ -1545,8 +1545,9 @@ def _prepare_decoder_attention_mask(attention_mask, input_shape, past_key_values expanded_attn_mask = expanded_attn_mask.astype("float32") expanded_attn_mask = paddle.where(expanded_attn_mask, x, y).astype(dtype) elif get_env_device() in ["xpu", "gcu"]: + min_val = paddle.finfo(dtype).min if get_env_device() == "gcu" else -1e37 # mask value for xpu x = paddle.to_tensor(0.0, dtype=dtype) - y = paddle.to_tensor(paddle.finfo(dtype).min, dtype=dtype) + y = paddle.to_tensor(min_val, dtype=dtype) expanded_attn_mask = expanded_attn_mask.astype(dtype) expanded_attn_mask = paddle.where(expanded_attn_mask, x, y).astype(dtype) else: