Fix precison precision (#61033)

co63oc · web-flow · commit 38b9793e51f7 · 2024-01-23T14:41:27.000+08:00
* Fix

* Fix
diff --git a/paddle/fluid/platform/device/ipu/popart_canonicalization/tensor_ops.cc b/paddle/fluid/platform/device/ipu/popart_canonicalization/tensor_ops.cc
@@ -213,7 +213,7 @@ Node *cast_handler(Graph *graph, Node *node) {
                              node->inputs,
                              node->outputs,
                              static_cast<VarType::Type>(otype));
-  // Cast op created in mixed-precison has no pipline attrs
+  // Cast op created in mixed-precision has no pipline attrs
   auto &prev_nodes = node->inputs.front()->inputs;
   if (!prev_nodes.empty()) {
     auto *prev_op = prev_nodes.front()->Op();
diff --git a/paddle/phi/kernels/funcs/cufft_util.h b/paddle/phi/kernels/funcs/cufft_util.h
@@ -84,20 +84,20 @@ class FFTConfig {
   // sizes are full signal, including batch size and always two-sided
   FFTConfig(const std::vector<int64_t>& sizes,
             FFTTransformType fft_type,
-            DataType precison)
-      : fft_type_(fft_type), precision_(precison) {
+            DataType precision)
+      : fft_type_(fft_type), precision_(precision) {
     const auto batch_size = static_cast<plan_size_type>(sizes[0]);
     std::vector<plan_size_type> signal_sizes(sizes.cbegin() + 1, sizes.cend());
     const int signal_ndim = sizes.size() - 1;
 
     cudaDataType itype, otype, exec_type;
     const bool complex_input = has_complex_input(fft_type);
     const bool complex_output = has_complex_output(fft_type);
-    if (precison == DataType::FLOAT32) {
+    if (precision == DataType::FLOAT32) {
       itype = complex_input ? CUDA_C_32F : CUDA_R_32F;
       otype = complex_output ? CUDA_C_32F : CUDA_R_32F;
       exec_type = CUDA_C_32F;
-    } else if (precison == DataType::FLOAT64) {
+    } else if (precision == DataType::FLOAT64) {
       itype = complex_input ? CUDA_C_64F : CUDA_R_64F;
       otype = complex_output ? CUDA_C_64F : CUDA_R_64F;
       exec_type = CUDA_C_64F;
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
@@ -1527,7 +1527,7 @@ set -x
         # set trt_convert ut to run 15% cases.
         export TEST_NUM_PERCENT_CASES=0.15
         export FLAGS_trt_ibuilder_cache=1
-        precison_cases=""
+        precision_cases=""
         bash $PADDLE_ROOT/tools/check_added_ut.sh
         if [ ${PRECISION_TEST:-OFF} == "ON" ]; then
             python $PADDLE_ROOT/tools/get_pr_ut.py
@@ -2529,7 +2529,7 @@ set -x
         # set trt_convert ut to run 15% cases.
         export TEST_NUM_PERCENT_CASES=0.15
         export FLAGS_trt_ibuilder_cache=1
-        precison_cases=""
+        precision_cases=""
         bash $PADDLE_ROOT/tools/check_added_ut.sh
         #check change of pr_unittests and dev_unittests
         check_approvals_of_unittest 2
diff --git a/python/paddle/amp/auto_cast.py b/python/paddle/amp/auto_cast.py
@@ -138,7 +138,7 @@ def need_keep_fp32(layer, dtype):
     need_keep_fp32 = False
     # Highest prority. Because all the layers except BN will use bfloat16 params in bfoat16 training,
     # here we provide a option to keep fp32 param.
-    if not layer._cast_to_low_precison:
+    if not layer._cast_to_low_precision:
         need_keep_fp32 = True
     # The BN layers will keep fp32
     elif isinstance(
@@ -197,12 +197,12 @@ def set_excluded_layers(models, excluded_layers):
         for layer in excluded_layers_instances[idx].sublayers(
             include_self=True
         ):
-            layer._cast_to_low_precison = False
+            layer._cast_to_low_precision = False
     excluded_layers_types = tuple(excluded_layers_types)
     for idx in range(len(models)):
         for layer in models[idx].sublayers(include_self=True):
             if isinstance(layer, excluded_layers_types):
-                layer._cast_to_low_precison = False
+                layer._cast_to_low_precision = False
 
 
 @dygraph_only
diff --git a/python/paddle/nn/layer/layers.py b/python/paddle/nn/layer/layers.py
@@ -412,7 +412,7 @@ def __init__(self, name_scope=None, dtype="float32"):
         self._forward_post_hooks = collections.OrderedDict()
 
         # only used in AMP Training
-        self._cast_to_low_precison = True
+        self._cast_to_low_precision = True
 
         self._state_dict_hooks = collections.OrderedDict()
         # Records orignal functions after @to_static to support to rollback
diff --git a/python/paddle/static/amp/fp16_utils.py b/python/paddle/static/amp/fp16_utils.py
@@ -422,7 +422,7 @@ def fp16_guard():
 def set_var_dst_dtype(
     op, var_names, block, global_block, dtype, need_set_dtype
 ):
-    low_precison_var_names = set()
+    low_precision_var_names = set()
     for var_name in var_names:
         var = None
         try:
@@ -439,7 +439,7 @@ def set_var_dst_dtype(
             continue
 
         if var.dtype in FLOAT_TYPES:
-            low_precison_var_names.add(var_name)
+            low_precision_var_names.add(var_name)
             if need_set_dtype:
                 var.desc.set_dtype(dtype)
 
@@ -449,7 +449,7 @@ def set_var_dst_dtype(
             )
         )
 
-    return low_precison_var_names
+    return low_precision_var_names
 
 
 def set_param_dtype(program, dtype, amp_lists, use_fp16_guard, level):
@@ -581,8 +581,8 @@ def get_amp_dst_dtype(
 
 
 def process_op_input_and_outputs(op, block, global_block, dtype):
-    low_precison_var_names = set()
-    # Get the FP16 input because the low_precison_var_names is required for the parameter casting.
+    low_precision_var_names = set()
+    # Get the FP16 input because the low_precision_var_names is required for the parameter casting.
     # The dtype of the input is not set to fp16, because it is done in the step 3 of cast_model_to_fp16.
     for in_name in op.input_names:
         # for ipu, all inputs must be converted to fp16
@@ -596,7 +596,7 @@ def process_op_input_and_outputs(op, block, global_block, dtype):
             dtype,
             need_set_dtype=False,
         )
-        low_precison_var_names = low_precison_var_names.union(in_vars)
+        low_precision_var_names = low_precision_var_names.union(in_vars)
     # Set the output to FP16 because its consumer OP needs to determine if the dtype needs
     # to be promoted.
     for out_name in op.output_names:
@@ -611,7 +611,7 @@ def process_op_input_and_outputs(op, block, global_block, dtype):
             dtype,
             need_set_dtype=True,
         )
-    return low_precison_var_names
+    return low_precision_var_names
 
 
 def map_block(block, fn, parent_op=None):
