skips due to padding size issue 63 on amd

ROCm · Oct 31, 2024 · a02eb24 · a02eb24
1 parent ede06df
commit a02eb24
Show file tree

Hide file tree

Showing 3 changed files with 6 additions and 0 deletions.
diff --git a/tests/models/jamba/test_modeling_jamba.py b/tests/models/jamba/test_modeling_jamba.py
@@ -29,6 +29,7 @@
     require_torch_gpu,
     slow,
     torch_device,
+    skipIfRocm
 )
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -365,6 +366,7 @@ def test_decoder_model_past_with_large_inputs(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
         self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
 
+    @skipIfRocm
     def test_load_balancing_loss(self):
         r"""
         Let's make sure we can actually compute the loss and do a backward on it.

diff --git a/tests/models/mixtral/test_modeling_mixtral.py b/tests/models/mixtral/test_modeling_mixtral.py
@@ -28,6 +28,7 @@
     require_torch_sdpa,
     slow,
     torch_device,
+    skipIfRocm
 )
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -493,6 +494,7 @@ def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest(reason="Mixtral flash attention does not support right padding")
 
     # Ignore copy
+    @skipIfRocm
     def test_load_balancing_loss(self):
         r"""
         Let's make sure we can actually compute the loss and do a backward on it.

diff --git a/tests/models/qwen2_moe/test_modeling_qwen2_moe.py b/tests/models/qwen2_moe/test_modeling_qwen2_moe.py
@@ -30,6 +30,7 @@
     require_torch_sdpa,
     slow,
     torch_device,
+    skipIfRocm
 )
 
 from ...generation.test_utils import GenerationTesterMixin
@@ -528,6 +529,7 @@ def test_flash_attn_2_inference_equivalence_right_padding(self):
         self.skipTest(reason="Qwen2Moe flash attention does not support right padding")
 
     # Ignore copy
+    @skipIfRocm
     def test_load_balancing_loss(self):
         r"""
         Let's make sure we can actually compute the loss and do a backward on it.