update test

tests/test_torch_compile.py

@@ -12,5 +12,4 @@ def test_torch_compile_forward():
 def test_torch_compile_backward():
     model = SimpleNN()
     compiled_model = torch.compile(model, backend="aot_eager")
-
     check_torch_compile_backward(model, compiled_model)

tests/test_torch_compile_aot_backward.py

@@ -6,7 +6,6 @@
 from tests.module_pool.simple_nn import SimpleNN
 
 
-@make_boxed_compiler
 def custom_compiler(gm: torch.fx.GraphModule, example_inputs):
     print("Using custom compiler!")
     gm.graph.print_tabular()
@@ -15,11 +14,13 @@ def custom_compiler(gm: torch.fx.GraphModule, example_inputs):
     return gm.forward
 
 
-def test_torch_dump_compile_backward():
+def test_torch_dump_compile():
     # 初始化模型、损失函数和优化器
     model = SimpleNN()
-    aot_backend = aot_autograd(fw_compiler=custom_compiler, bw_compiler=custom_compiler)  # 在backward时也使用自定义编译器
+    boxed_compiler = make_boxed_compiler(custom_compiler)  # 使用boxed_compiler包装自定义编译器, 解决aot_autograd的内存释放问题.
+    aot_backend = aot_autograd(fw_compiler=boxed_compiler, bw_compiler=boxed_compiler)  # 在backward时也使用自定义编译器
     compiled_model = torch.compile(model, backend=aot_backend)
+
     criterion = nn.MSELoss()
     optimizer = optim.SGD(compiled_model.parameters(), lr=0.01)
 

tests/test_torch_compile_mutiple_output.py

@@ -1,16 +1,19 @@
 import torch
 from torch._dynamo.backends.common import aot_autograd
-from torch._functorch._aot_autograd.utils import make_boxed_compiler
+from torch._functorch.aot_autograd import make_boxed_compiler
 
-from plai.pl_torch_compiler import dummy_compiler, dump_compiler, plnn_compiler
-from tests.module_pool.simple_nn import SimpleNN, check_torch_compile_forward, check_torch_compile_backward
+from plai.pl_torch_compiler import plnn_compiler
 
 
-def test_torch_plnn_compile_mutiple_output():
-    model = lambda x: torch.max(x, dim=0)[0]
+def test_torch_plnn_compile_multiple_output():
+    model = (lambda x: torch.max(x, dim=0))
     custom_compiler = plnn_compiler.CustomCompiler()
-    aot_backend = aot_autograd(fw_compiler=make_boxed_compiler(custom_compiler), bw_compiler=None)
+    aot_backend = aot_autograd(fw_compiler=make_boxed_compiler(custom_compiler))
     compiled_model = torch.compile(model, backend=aot_backend)
-    check_torch_compile_forward(model, compiled_model)
+    input_data = torch.randn(1, 10)
+    expected_output = model(input_data)
+    actual_output = compiled_model(input_data)
+    assert torch.allclose(expected_output.values, actual_output.values)
+    assert torch.allclose(expected_output.indices, actual_output.indices)
     print('dump compile forward:')
     print(custom_compiler.graph)