add ditorch and op_tool config file (#71)

* add ditorch and op_tool config file

* minor change

* Ignore framework custom operators in the recommended configuration

* set dtype used by cpu for autocompare for torch.exp torch.Tensor.std

* minor change

ditorch/common_adapter.py

@@ -9,7 +9,10 @@ def __init__(self):
             super().__init__()
 
         def __torch_function__(self, func, types, args, kwargs=None):
-            name = resolve_name(func)
+            try:
+                name = resolve_name(func)
+            except Exception:
+                name = None
             result = func(*args, **(kwargs or {}))
             if name == "torch.Tensor.device.__get__":
                 if result.type != "cpu":

ditorch_config.sh

@@ -0,0 +1,59 @@
+# export env_name=${env_name:-env_default_value}
+
+export DITORCH_SHOW_DEVICE_AS_CUDA=${DITORCH_SHOW_DEVICE_AS_CUDA:-1}
+
+
+export OP_TOOLS_PRINT_STACK=${OP_TOOLS_PRINT_STACK:-0}
+
+export OP_TOOLS_MAX_CACHE_SIZE=${OP_TOOLS_MAX_CACHE_SIZE:-1000}
+
+
+export OP_AUTOCOMPARE_DISABLE_LIST=${OP_AUTOCOMPARE_DISABLE_LIST:-"torch.rand,torch.randn,torch_mlu.*,torch_npu.*"}
+export OP_AUTOCOMPARE_LIST=${OP_AUTOCOMPARE_LIST:-".*"}
+
+
+export OP_CAPTURE_DISABLE_LIST=${OP_CAPTURE_DISABLE_LIST:-""}
+export OP_CAPTURE_LIST=${OP_CAPTURE_LIST:-".*"}
+
+export OP_DTYPE_CAST_DISABLE_LIST=${OP_DTYPE_CAST_DISABLE_LIST:-""}
+export OP_DTYPE_CAST_LIST=${OP_DTYPE_CAST_LIST:-".*"}
+
+export OP_FALLBACK_DISABLE_LIST=${OP_FALLBACK_DISABLE_LIST:-""}
+export OP_FALLBACK_LIST=${OP_FALLBACK_LIST:-".*"}
+
+export OP_OBSERVE_DISABLE_LIST=${OP_OBSERVE_DISABLE_LIST:-""}
+export OP_OBSERVE_LIST=${OP_OBSERVE_LIST:-".*"}
+
+
+export OP_OVERFLOW_CHECK_DISABLE_LIST=${OP_OVERFLOW_CHECK_DISABLE_LIST:-""}
+export OP_OVERFLOW_CHECK_LIST=${OP_OVERFLOW_CHECK_LIST:-".*"}
+
+
+export OP_TIME_MEASURE_DISABLE_LIST=${OP_TIME_MEASURE_DISABLE_LIST:-""}
+export OP_TIME_MEASURE_LIST=${OP_TIME_MEASURE_LIST:-".*"}
+
+
+# for autocompare and op_dtype_cast tools
+# Set the dtype used by the CPU for autocompare
+# Set the dtype used by the DEVICE for op_dtype_cast
+# for special op
+export LINEAR_OP_DTYPE_CAST_DICT=${LINEAR_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export EMBEDDING_OP_DTYPE_CAST_DICT=${EMBEDDING_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export NORMALIZE_OP_DTYPE_CAST_DICT=${NORMALIZE_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export NORM_OP_DTYPE_CAST_DICT=${NORM_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export CROSS_ENTROPY_OP_DTYPE_CAST_DICT=${CROSS_ENTROPY_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export MUL_OP_DTYPE_CAST_DICT=${MUL_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export MATMUL_OP_DTYPE_CAST_DICT=${MATMUL_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export STD_OP_DTYPE_CAST_DICT=${STD_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+export EXP_OP_DTYPE_CAST_DICT=${EXP_OP_DTYPE_CAST_DICT:-"torch.float16->torch.float64,torch.bfloat16->torch.float64,torch.float32->torch.float64"}
+# for generally op
+export OP_DTYPE_CAST_DICT=${OP_DTYPE_CAST_DICT:-"torch.float16->torch.float32,torch.bfloat16->torch.float32"}
+
+
+export AUTOCOMPARE_ERROR_TOLERANCE=${AUTOCOMPARE_ERROR_TOLERANCE:-"1e-3,1e-3"}
+export AUTOCOMPARE_ERROR_TOLERANCE_FLOAT16=${AUTOCOMPARE_ERROR_TOLERANCE_FLOAT16:-"1e-4,1e-4"}
+export AUTOCOMPARE_ERROR_TOLERANCE_BFLOAT16=${AUTOCOMPARE_ERROR_TOLERANCE_BFLOAT16:-"1e-3,1e-3"}
+export AUTOCOMPARE_ERROR_TOLERANCE_FLOAT32=${AUTOCOMPARE_ERROR_TOLERANCE_FLOAT32:-"1e-5,1e-5"}
+export AUTOCOMPARE_ERROR_TOLERANCE_FLOAT64=${AUTOCOMPARE_ERROR_TOLERANCE_FLOAT64:-"1e-8,1e-8"}
+
+export LINEAR_AUTOCOMPARE_ERROR_TOLERANCE_BFLOAT16=${LINEAR_AUTOCOMPARE_ERROR_TOLERANCE_BFLOAT16:-"1e-2,1e-2"}

op_tools/op_autocompare_hook.py

@@ -276,7 +276,7 @@ def count_params_with_requires_grad(self):
 
     def compare_input_grad(self):
         self.args_grad = self.grad_inputs_cpu
-        compare_info = compare_result(self.name + " grad", self.args_grad, self.args_cpu_grad)
+        compare_info = compare_result(self.name + " grad", list(self.args_grad), list(self.args_cpu_grad))
         compare_info["forward_id"] = self.forward_op_id
 
         compare_result_cache.append(self.forward_op_id, compare_info)

op_tools/test/test_opname_match.py

@@ -25,6 +25,7 @@ def test_opname_match(self):
         self.assertEqual(is_opname_match("torch.subc", "torch.add,.*"), True)
         self.assertEqual(is_opname_match("torch.subc", None), True)
         self.assertEqual(is_opname_match("torch.subc", ""), False)
+        self.assertEqual(is_opname_match("torch_mlu.fused_l2_norm", "torch_mlu.*"), True)
         self.assertEqual(is_opname_match(None, "torch.add"), False)
         self.assertEqual(is_opname_match(None, None), False)