fix CI for Paddle#72862

paddlenlp/transformers/convbert/modeling.py

@@ -322,6 +322,7 @@ def __init__(self, hidden_size, hidden_act):
         self.act = get_activation(hidden_act)
 
     def forward(self, discriminator_hidden_states):
+        discriminator_hidden_states = discriminator_hidden_states[0]
         hidden_states = self.dense(discriminator_hidden_states)
         hidden_states = self.act(hidden_states)
         logits = self.dense_prediction(hidden_states).squeeze()

paddlenlp/transformers/rembert/modeling.py

@@ -177,6 +177,7 @@ def __init__(self, config: RemBertConfig):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
 
     def forward(self, hidden_states, input_tensor):
+        hidden_states = hidden_states[0]
         hidden_states = self.dense(hidden_states)
         hidden_states = self.dropout(hidden_states)
         hidden_states = self.layer_norm(hidden_states + input_tensor)

paddlenlp/utils/optimizer.py

@@ -144,7 +144,7 @@ def adamw_python(
         mom1 = beta1 * mom1 + (1.0 - beta1) * grad
         mom2 = beta2 * mom2 + (1.0 - beta2) * (grad * grad).mean()
         denom = mom2.sqrt() / (1.0 - beta2_pow).sqrt() + epsilon
-        p += (moment1 / denom) * (-(lr / (1.0 - beta1_pow)))
+        p += (mom1 / denom) * (-(lr / (1.0 - beta1_pow)))
         if master_weight is not None:
             master_weight[:] = p
             param[:] = p.astype(param.dtype)

requirements.txt

@@ -5,7 +5,8 @@ colorama
 seqeval
 dill<0.3.5
 multiprocess<=0.70.12.2
-datasets >= 2.0.0
+datasets==3.6.0
+pyarrow==20.0.0
 tqdm
 paddlefsl
 sentencepiece
@@ -21,7 +22,7 @@ typer
 rich
 safetensors
 fast_dataindex>=0.1.1 ; platform_system == "Linux"
-aistudio-sdk>=0.1.3
+aistudio-sdk>=0.3.0
 jinja2
 regex
 numpy<=1.26.4
@@ -30,3 +31,4 @@ ml_dtypes
 tokenizers<=0.20.3; python_version<="3.8"
 tokenizers>=0.21,<0.22; python_version>"3.8"
 omegaconf
+einops>=0.7.0

tests/transformers/bloom/test_modeling.py

@@ -505,6 +505,10 @@ def test_inputs_embeds(self):
 
             with paddle.no_grad():
                 embeds_output = model(**inputs)
+            if isinstance(ids_output, tuple):
+                ids_output = ids_output[0]
+            if isinstance(embeds_output, tuple):
+                embeds_output = embeds_output[0]
 
             self.assertTrue(paddle.allclose(ids_output, embeds_output, rtol=1e-4, atol=1e-4))
 

tests/transformers/mbart/test_modeling.py

@@ -258,6 +258,10 @@ def test_inputs_embeds_for_mbart(self):
 
             with paddle.no_grad():
                 embeds_output = model(**inputs)
+            if isinstance(ids_output, tuple):
+                ids_output = ids_output[0]
+            if isinstance(embeds_output, tuple):
+                embeds_output = embeds_output[0]
 
             self.assertTrue(paddle.allclose(ids_output, embeds_output, rtol=1e-4, atol=1e-4))