mindspore 2.4.10__cp311-cp311-win_amd64.whl → 2.6.0rc1__cp311-cp311-win_amd64.whl

mindspore/runtime/memory.py

@@ -0,0 +1,409 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""Memory interfaces."""
+
+from mindspore._c_expression import RuntimeConf, DeviceManagerConf, _memory_stats, \
+    _reset_max_mem_reserved, _reset_max_mem_allocated, DeviceContextManager
+from mindspore import _checkparam as Validator
+from mindspore._checkparam import args_type_check
+from mindspore import log as logger
+import mindspore as ms
+
+_MEMORY_PATTERN = r'[1-9][0-9]*(\.)?[0-9]*GB|0\.[0-9]*GB'
+_RESERVE_PATTERN = r'[0-9][0-9]*(\.)?[0-9]*GB|0\.[0-9]*GB'
+_device_context_mgr = DeviceContextManager.get_instance()
+
+
+@args_type_check(
+    init_size=str,
+    increase_size=str,
+    max_size=str,
+    optimize_level=str,
+    huge_page_reserve_size=str,
+)
+def set_memory(init_size="2GB", increase_size="2GB", max_size="1024GB", optimize_level="O0",
+               huge_page_reserve_size="0GB"):
+    """
+    Set the memory parameters of runtime device memory management that is implemented using a memory pool.
+
+    The framework will set all the args by default as follows.
+
+    Args:
+        init_size (str): The init size of memory pool. The format is "xxGB". Default: ``2GB`` .
+        increase_size (str): The increase size of memory pool. When the current memory pool has no
+            enough memory, the memory pool will be expanded by this value. The format is "xxGB". Default: ``2GB`` .
+        max_size (str): The maximum memory available for memory pool.
+            The actual used memory size is the minimum of the available memory of the device and max_device_memory.
+            The format is "xxGB". Default is the maximum available memory of the device, expressed as ``1024GB``.
+        optimize_level (str): The memory optimize level. The value must be in ['O0', 'O1']. Default: ``O0`` .
+        huge_page_reserve_size (str): The reserved size of huge page memory. The format is "xxGB". Default: ``0GB``.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> ms.set_device("Ascend", 1)
+        >>> ms.runtime.set_memory("10GB", "2GB", "60GB", "O1", "0GB")
+    """
+    if RuntimeConf.get_instance().is_memory_configured():
+        raise RuntimeError("The 'set_memory' can not be set repeatedly.")
+
+    _check_memory_conf_valid(init_size)
+    _check_memory_conf_valid(increase_size)
+    _check_memory_conf_valid(max_size)
+    Validator.check_str_by_regular(huge_page_reserve_size, _RESERVE_PATTERN)
+    init_value = float(init_size[:-2])
+    increase_value = float(increase_size[:-2])
+    max_value = float(max_size[:-2])
+    huge_page_reserve_value = float(huge_page_reserve_size[:-2])
+
+    memory_optimize_levels = ["O0", "O1"]
+    if optimize_level not in memory_optimize_levels:
+        raise ValueError(f"The optimize_level must be one of "
+                         f"{memory_optimize_levels}, but got {optimize_level}.")
+    optimize_value = 0
+    if optimize_level == "O1":
+        optimize_value = 1
+
+    return RuntimeConf.get_instance().set_memory(
+        init_value,
+        increase_value,
+        max_value,
+        optimize_value,
+        huge_page_reserve_value,
+    )
+
+
+def _check_memory_conf_valid(memory_size):
+    """
+    Check whether the configuration memory value format is "xxGB" and can not be "0GB".
+    """
+    if not Validator.check_str_by_regular(memory_size, _MEMORY_PATTERN):
+        raise ValueError("The memory value should be in correct format!"
+                         "It must be a string ending with 'GB', in addition to that, it must contain "
+                         "only numbers or decimal points, such as \"5GB\" or \"3.5GB\", but got {}."
+                         .format(memory_size))
+    if memory_size == "0GB" or memory_size == "0.0GB":
+        raise ValueError("The memory value should not be \"0GB\".")
+
+def _is_initialized(device_target):
+    """
+    Returns whether specified backend is initialized.
+    """
+    _device_context = _device_context_mgr.get_device_context(device_target)
+    if _device_context is None:
+        return False
+    return _device_context.initialized()
+
+def memory_stats():
+    """
+    Returns status information queried from the memory pool.
+
+    Note:
+        For the `CPU` backend, a dictionary with empty data is always returned.
+
+    Returns:
+        dict, the queried memory information.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.memory_stats())
+        {'total_reserved_memory': 1073741824, 'total_allocated_memory': 1024, 'total_idle_memory': 1073740800,
+        'total_eager_free_memory': 0, 'max_reserved_memory': 1073741824, 'max_allocated_memory': 1536,
+        'common_mem_pool_stats': {'block_unit_size': 1073741824, 'block_counts': 1, 'blocks_info':
+        {<capsule object NULL at 0x7f7e8c27b030>: {'block_stream_id': 0, 'block_memory_size': 1073741824}}},
+        'persistent_mem_pool_stats': {'block_unit_size': 1073741824, 'block_counts': 0, 'blocks_info': {}}}
+    """
+    device_target = ms.context.get_context("device_target")
+    if not _is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return empty dict.")
+        return {}
+    return _memory_stats(device_target)
+
+
+def memory_reserved():
+    """
+    Returns the total amount of memory currently managed by the memory pool.
+
+    Note:
+        - For the `CPU` backend, 0 is always returned.
+
+    Returns:
+        int, in Byte.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.memory_reserved())
+        1073741824
+    """
+    device_target = ms.context.get_context("device_target")
+    if not _is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return 0.")
+        return 0
+    return _memory_stats(device_target).get("total_reserved_memory", 0)
+
+
+def max_memory_reserved():
+    """
+    Returns the peak value of the total memory managed by the memory pool since the process was started.
+
+    Note:
+        - For the `CPU` backend, 0 is always returned.
+
+    Returns:
+        int, in Byte.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.max_memory_reserved())
+        1073741824
+    """
+    device_target = ms.context.get_context("device_target")
+    if not _is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return 0.")
+        return 0
+    return _memory_stats(device_target).get("max_reserved_memory", 0)
+
+
+def empty_cache():
+    """
+    Release all memory fragments in the memory pool, so that memory arrangement
+    will be optimized.
+
+    Note:
+        Currently, the MindSpore memory pool does not have the function of releasing memory fragments.
+        This interface is reserved but implemented as an empty method and prompted in log mode when using.
+    """
+    logger.warning(f"The empty_cache operation is currently not supported.")
+
+
+def reset_peak_memory_stats():
+    """
+    Reset the "peak" stats tracked by memory manager.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.max_memory_reserved())
+        1073741824
+        >>> print(ms.runtime.max_memory_allocated())
+        1536
+        >>> ms.runtime.reset_peak_memory_stats()
+        >>> print(ms.runtime.max_memory_reserved())
+        0
+        >>> print(ms.runtime.max_memory_allocated())
+        0
+    """
+    device_target = ms.context.get_context("device_target")
+    _reset_max_mem_reserved(device_target)
+    _reset_max_mem_allocated(device_target)
+
+
+def memory_summary():
+    """
+    Returns readable memory pool status information.
+
+    Returns:
+        str, readable memory pool status information in tabular form.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+    """
+    device_target = DeviceManagerConf.get_instance().get_device_target()
+    stats = _memory_stats(device_target)
+
+    def _format_size(sz, pref_sz):
+        prefixes = ["B  ", "KB", "MB", "GB", "TB", "PB"]
+        prefix = prefixes[0]
+        for new_prefix in prefixes[1:]:
+            if pref_sz < 768 * 1024:
+                break
+            prefix = new_prefix
+            sz //= 1024
+            pref_sz /= 1024
+        return f"{sz:6d} {prefix}"
+
+    metrics_to_display = [
+        ("total_reserved_memory", "Reserved memory", _format_size),
+        ("total_allocated_memory", "Allocated memory", _format_size),
+        ("total_idle_memory", "Idle memory", _format_size),
+        ("total_eager_free_memory", "Eager free memory", _format_size),
+        ("max_reserved_memory", "Max reserved memory", _format_size),
+        ("max_allocated_memory", "Max allocated memory", _format_size),
+    ]
+
+    lines = []
+    lines.append("=" * 45)
+    lines.append(" {:^43} ".format("Memory summary"))
+    lines.append("=" * 45)
+    lines.append(" {:<20} | {:<20} ".format("Metric", "Data"))
+
+    for metric_key, metric_name, formatter in metrics_to_display:
+        lines.append("-" * 45)
+        data = stats[metric_key]
+        lines.append(" {:<20} | {:<20} ".format(metric_name, formatter(data, data)))
+
+    lines.append("=" * 45)
+
+    return "|" + "|\n|".join(lines) + "|\n"
+
+
+def memory_allocated():
+    """
+    Returns the actual memory size currently occupied by Tensor.
+
+    Note:
+        - For the `CPU` backend, 0 is always returned.
+
+    Returns:
+        int, in Byte.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.memory_allocated())
+        1024
+    """
+    device_target = ms.context.get_context("device_target")
+    if not _is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return 0.")
+        return 0
+    return _memory_stats(device_target).get("total_allocated_memory", 0)
+
+
+def max_memory_allocated():
+    """
+    Returns the peak memory size of the memory pool actually occupied by Tensor since the process was started.
+
+    Note:
+        - For the `CPU` backend, 0 is always returned.
+
+    Returns:
+        int, in Byte.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.max_memory_allocated())
+        1536
+    """
+    device_target = ms.context.get_context("device_target")
+    if not _is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return 0.")
+        return 0
+    return _memory_stats(device_target).get("max_allocated_memory", 0)
+
+
+def reset_max_memory_reserved():
+    """
+    Reset the peak memory size managed by the memory pool.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.max_memory_reserved())
+        1073741824
+        >>> ms.runtime.reset_max_memory_reserved()
+        >>> print(ms.runtime.max_memory_reserved())
+        0
+    """
+    device_target = ms.context.get_context("device_target")
+    _reset_max_mem_reserved(device_target)
+
+
+def reset_max_memory_allocated():
+    """
+    Reset the peak memory size of the memory pool actually occupied by Tensor.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> ms.set_device("Ascend", 0)
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.runtime.max_memory_allocated())
+        1536
+        >>> ms.runtime.reset_max_memory_allocated()
+        >>> print(ms.runtime.max_memory_allocated())
+        0
+    """
+    device_target = ms.context.get_context("device_target")
+    _reset_max_mem_allocated(device_target)