remove some flags and add comments to some flags, test=develop (#19813)

paddle/fluid/memory/detail/buddy_allocator.cc

@@ -19,10 +19,6 @@ limitations under the License. */
 
 #include "glog/logging.h"
 
-DEFINE_bool(free_idle_memory, false,
-            "If it is true, Paddle will try to free idle memory trunks during "
-            "running time.");
-
 #ifdef PADDLE_WITH_CUDA
 DECLARE_uint64(reallocate_gpu_memory_in_mb);
 #endif

paddle/fluid/platform/flags.cc

@@ -104,7 +104,7 @@ DEFINE_bool(cudnn_deterministic, false,
  * CUDNN related FLAG
  * Name: FLAGS_conv_workspace_size_limit
  * Since Version: 0.13.0
- * Value Range: uint64, default=4096 (MB)
+ * Value Range: uint64, default=512 (MB)
  * Example:
  * Note: The internal function of cuDNN obtains the fastest matching algorithm
  *       within this memory limit. Usually, faster algorithms can be chosen in
@@ -316,10 +316,15 @@ DEFINE_string(allocator_strategy, "naive_best_fit",
 /**
  * Memory related FLAG
  * Name: FLAGS_fraction_of_cpu_memory_to_use
- * Since Version:
- * Value Range:
+ * Since Version: 0.12.0
+ * Value Range: double, [0.0, 1.0], default=1
  * Example:
- * Note:
+ * Note: Represents the proportion of allocated CPU memory blocks
+ *       to the total memory size of the CPU. Future CPU memory usage
+ *       will be allocated from this memory block. If the memory block does
+ *       not have enough CUDA pinned memory, new memory blocks of the same
+ *       size as the memory block will be allocated from the CUDA pinned
+ *       request util the CPU does not have enough memory.
  */
 DEFINE_double(fraction_of_cpu_memory_to_use, 1,
               "Default use 100% of CPU memory for PaddlePaddle,"
@@ -343,10 +348,15 @@ DEFINE_uint64(initial_cpu_memory_in_mb, 500ul,
 /**
  * Memory related FLAG
  * Name: FLAGS_fraction_of_cuda_pinned_memory_to_use
- * Since Version:
- * Value Range:
+ * Since Version: 0.12.0
+ * Value Range: double, [0.0, 1.0], default=0.5
  * Example:
- * Note:
+ * Note: Represents the proportion of allocated CUDA pinned memory blocks
+ *       to the total memory size of the CPU. Future CUDA pinned memory usage
+ *       will be allocated from this memory block. If the memory block does
+ *       not have enough CPU memory, new memory blocks of the same
+ *       size as the memory block will be allocated from the CPU
+ *       request util the CPU does not have enough memory.
  */
 DEFINE_double(
     fraction_of_cuda_pinned_memory_to_use, 0.5,

python/paddle/fluid/__init__.py

@@ -154,13 +154,13 @@ def __bootstrap__():
     read_env_flags = [
         'check_nan_inf', 'fast_check_nan_inf', 'benchmark',
         'eager_delete_scope', 'initial_cpu_memory_in_mb', 'init_allocated_mem',
-        'free_idle_memory', 'paddle_num_threads', "dist_threadpool_size",
-        'eager_delete_tensor_gb', 'fast_eager_deletion_mode',
-        'memory_fraction_of_eager_deletion', 'allocator_strategy',
-        'reader_queue_speed_test_mode', 'print_sub_graph_dir',
-        'pe_profile_fname', 'inner_op_parallelism', 'enable_parallel_graph',
-        'fuse_parameter_groups_size', 'multiple_of_cupti_buffer_size',
-        'fuse_parameter_memory_size', 'tracer_profile_fname', 'dygraph_debug'
+        'paddle_num_threads', 'dist_threadpool_size', 'eager_delete_tensor_gb',
+        'fast_eager_deletion_mode', 'memory_fraction_of_eager_deletion',
+        'allocator_strategy', 'reader_queue_speed_test_mode',
+        'print_sub_graph_dir', 'pe_profile_fname', 'inner_op_parallelism',
+        'enable_parallel_graph', 'fuse_parameter_groups_size',
+        'multiple_of_cupti_buffer_size', 'fuse_parameter_memory_size',
+        'tracer_profile_fname', 'dygraph_debug'
     ]
     if 'Darwin' not in sysstr:
         read_env_flags.append('use_pinned_memory')