[fx] fixed unit tests for torch 1.12 (#1327)

colossalai/fx/tracer/_tracer_utils.py

@@ -22,7 +22,6 @@ def extract_meta(*args, **kwargs):
         if isinstance(val, MetaDeviceAttribute):
             return 'meta'
         elif isinstance(val, ColoProxy):
-            assert val.meta_data is not None
             return val.meta_data
         return val
 

requirements/requirements-test.txt

 pytest
 torchvision
 transformers
+timm
 titans

tests/test_fx/test_pipeline/test_hf_model/test_albert.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 2
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_single_sentence_albert():
     MODEL_LIST = [
         transformers.AlbertModel,

tests/test_fx/test_pipeline/test_hf_model/test_bert.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 2
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_single_sentence_bert():
     MODEL_LIST = [
         transformers.BertModel,

tests/test_fx/test_pipeline/test_hf_model/test_gpt.py

@@ -9,7 +9,6 @@ NUM_EPOCHS = 2
 NUM_CHUNKS = 1
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_gpt():
     MODEL_LIST = [
         transformers.GPT2Model,

tests/test_fx/test_pipeline/test_hf_model/test_opt.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 1
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_opt():
     MODEL_LIST = [
         transformers.OPTModel,

tests/test_fx/test_pipeline/test_hf_model/test_t5.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 1
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_t5():
     MODEL_LIST = [
         transformers.T5Model,

tests/test_fx/test_pipeline/test_timm_model/test_timm.py

@@ -7,7 +7,6 @@ except:
 from timm_utils import split_model_and_compare_output
 
 
-@pytest.mark.skip('skip as timm is required')
 def test_timm_models_without_control_flow():
 
     MODEL_LIST = [
@@ -28,7 +27,6 @@ def test_timm_models_without_control_flow():
         split_model_and_compare_output(model, data)
 
 
-@pytest.mark.skip('skip as timm is required')
 def test_timm_models_with_control_flow():
     torch.backends.cudnn.deterministic = True
 

tests/test_fx/test_pipeline/test_torchvision/test_torchvision.py

@@ -19,7 +19,6 @@ torch.manual_seed(MANUAL_SEED)
 torch.backends.cudnn.deterministic = True
 
 
-@pytest.mark.skip('skip as torchvision is required')
 def test_torchvision_models():
     MODEL_LIST = [
         tm.vgg11, tm.resnet18, tm.densenet121, tm.mobilenet_v3_small, tm.resnext50_32x4d, tm.wide_resnet50_2,

tests/test_fx/test_tracer/test_hf_model/test_hf_albert.py

@@ -34,7 +34,6 @@ def test_single_sentence_albert():
         trace_model_and_compare_output(model, data_gen)
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_multi_sentence_albert():
     config = transformers.AlbertConfig(hidden_size=128,
                                        num_hidden_layers=2,

tests/test_fx/test_tracer/test_hf_model/test_hf_bert.py

@@ -31,7 +31,6 @@ def test_single_sentence_bert():
         trace_model_and_compare_output(model, data_gen)
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_multi_sentence_bert():
     config = transformers.BertConfig(hidden_size=128, num_hidden_layers=2, num_attention_heads=4, intermediate_size=256)
     tokenizer = transformers.BertTokenizer.from_pretrained("bert-base-uncased")

tests/test_fx/test_tracer/test_hf_model/test_hf_gpt.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 1
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_gpt():
     MODEL_LIST = [
         transformers.GPT2Model,

tests/test_fx/test_tracer/test_hf_model/test_hf_opt.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 1
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_opt():
     MODEL_LIST = [
         transformers.OPTModel,

tests/test_fx/test_tracer/test_hf_model/test_hf_t5.py

@@ -7,7 +7,6 @@ BATCH_SIZE = 1
 SEQ_LENGHT = 16
 
 
-@pytest.mark.skip("error with pytorch 1.10")
 def test_t5():
     MODEL_LIST = [
         transformers.T5Model,

tests/test_fx/test_tracer/test_patched_module.py

@@ -40,7 +40,7 @@ def test_embedding():
     _assert_output_shape(data, ln, patched_module.torch_nn_normalize, False, data.shape)
 
     # test group norm
-    gn = torch.nn.GroupNorm(4, num_channels=2)
+    gn = torch.nn.GroupNorm(4, num_channels=8)
     _assert_output_shape(data, gn, patched_module.torch_nn_normalize, False, data.shape)
 
     # test batch norm 1d

tests/test_fx/test_tracer/test_timm_model/test_timm_model.py

@@ -36,7 +36,6 @@ def trace_and_compare(model_cls, tracer, data, meta_args=None):
             fx_out, non_fx_out), f'{model.__class__.__name__} has inconsistent outputs, {fx_out} vs {non_fx_out}'
 
 
-@pytest.mark.skip('skip as timm is required')
 def test_timm_models_without_control_flow():
     torch.backends.cudnn.deterministic = True
 
@@ -58,7 +57,6 @@ def test_timm_models_without_control_flow():
         trace_and_compare(model_cls, tracer, data)
 
 
-@pytest.mark.skip('skip as timm is required')
 def test_timm_models_with_control_flow():
     torch.backends.cudnn.deterministic = True
 

tests/test_fx/test_tracer/test_torchvision_model/test_torchvision_model.py

@@ -8,7 +8,6 @@ from colossalai.fx import ColoTracer
 from torch.fx import GraphModule
 
 
-@pytest.mark.skip('skip as torchvision is required')
 def test_torchvision_models():
     MODEL_LIST = [
         tm.vgg11, tm.resnet18, tm.densenet121, tm.mobilenet_v3_small, tm.resnext50_32x4d, tm.wide_resnet50_2,