frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 12:54:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Add unified memory APIs for torch.accelerator (#152932)"

Browse Source 
This reverts commit 15f1173e5d72d6d45faba4cecd135e0160f06c6f.

Reverted https://github.com/pytorch/pytorch/pull/152932 on behalf of https://github.com/jithunnair-amd due to Broke ROCm periodic runs on MI300 e.g. https://github.com/pytorch/pytorch/actions/runs/16764977800/job/47470050573 ([comment](https://github.com/pytorch/pytorch/pull/138222#issuecomment-3164941815))
This commit is contained in:
PyTorch MergeBot
2025-08-07 16:34:36 +00:00
parent c4e64467b5
commit 74da2604c9
7 changed files with 2 additions and 335 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
aten/src/ATen/DeviceAccelerator.h
View File

@ -1,6 +1,5 @@
#pragma once
#include <c10/core/CachingDeviceAllocator.h>
#include <c10/core/DeviceType.h>
#include <c10/macros/Macros.h>
@ -73,27 +72,6 @@ TORCH_API c10::DeviceIndex exchangeDevice(c10::DeviceIndex device_index);
// original device index that was active before the change.
TORCH_API c10::DeviceIndex maybeExchangeDevice(c10::DeviceIndex device_index);
TORCH_API inline void emptyCache() {
const auto device_type = getAccelerator(true).value();
at::getDeviceAllocator(device_type)->emptyCache();
}
TORCH_API inline at::CachingDeviceAllocator::DeviceStats getDeviceStats(
c10::DeviceIndex device_index) {
const auto device_type = getAccelerator(true).value();
return at::getDeviceAllocator(device_type)->getDeviceStats(device_index);
}
TORCH_API inline void resetAccumulatedStats(c10::DeviceIndex device_index) {
const auto device_type = getAccelerator(true).value();
at::getDeviceAllocator(device_type)->resetAccumulatedStats(device_index);
}
TORCH_API inline void resetPeakStats(c10::DeviceIndex device_index) {
const auto device_type = getAccelerator(true).value();
at::getDeviceAllocator(device_type)->resetPeakStats(device_index);
}
} // namespace at::accelerator
namespace at {

23
docs/source/accelerator.md
View File

@ -25,26 +25,3 @@
synchronize
device_index
```
```{eval-rst}
.. automodule:: torch.accelerator.memory
```
```{eval-rst}
.. currentmodule:: torch.accelerator.memory
```
## Memory management
```{eval-rst}
.. autosummary::
:toctree: generated
:nosignatures:
empty_cache
max_memory_allocated
max_memory_reserved
memory_allocated
memory_reserved
memory_stats
reset_accumulated_memory_stats
reset_peak_memory_stats
```

5
torch/_C/__init__.pyi.in
View File

@ -2435,11 +2435,6 @@ def _accelerator_synchronizeDevice(device_index: _int) -> None: ...
def _accelerator_exchangeDevice(device_index: _int) -> _int: ...
def _accelerator_maybeExchangeDevice(device_index: _int) -> _int: ...
def _accelerator_setAllocatorSettings(env: str) -> None: ...
def _accelerator_isAllocatorInitialized() -> _bool: ...
def _accelerator_emptyCache() -> None: ...
def _accelerator_getDeviceStats(device_index: _int) -> dict[str, Any]: ...
def _accelerator_resetAccumulatedStats(device_index: _int) -> None: ...
def _accelerator_resetPeakStats(device_index: _int) -> None: ...
# Defined in torch/csrc/jit/python/python_tracer.cpp
class TracingState:

18
torch/accelerator/__init__.py
View File

@ -8,16 +8,6 @@ from typing_extensions import deprecated
import torch
from ._utils import _device_t, _get_device_index
from .memory import (
empty_cache,
max_memory_allocated,
max_memory_reserved,
memory_allocated,
memory_reserved,
memory_stats,
reset_accumulated_memory_stats,
reset_peak_memory_stats,
)
__all__ = [
@ -25,17 +15,9 @@ __all__ = [
"current_device_idx", # deprecated
"current_device_index",
"current_stream",
"empty_cache",
"device_count",
"device_index",
"is_available",
"max_memory_allocated",
"max_memory_reserved",
"memory_allocated",
"memory_reserved",
"memory_stats",
"reset_accumulated_memory_stats",
"reset_peak_memory_stats",
"set_device_idx", # deprecated
"set_device_index",
"set_stream",

201
torch/accelerator/memory.py
View File

@ -1,201 +0,0 @@
from collections import OrderedDict
from typing import Any
import torch
from ._utils import _device_t, _get_device_index
__all__ = [
"empty_cache",
"max_memory_allocated",
"max_memory_reserved",
"memory_allocated",
"memory_reserved",
"memory_stats",
"reset_accumulated_memory_stats",
"reset_peak_memory_stats",
]
def empty_cache() -> None:
r"""Release all unoccupied cached memory currently held by the caching
allocator so that those can be used in other application.
.. note:: This function is a no-op if the memory allocator for the current
:ref:`accelerator <accelerators>` has not been initialized.
"""
if not torch._C._accelerator_isAllocatorInitialized():
return
torch._C._accelerator_emptyCache()
def memory_stats(device_index: _device_t = None, /) -> OrderedDict[str, Any]:
r"""Return a dictionary of accelerator device memory allocator statistics for a given device index.
The return value of this function is a dictionary of statistics, each of
which is a non-negative integer.
Core statistics:
- ``"allocated.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
number of allocation requests received by the memory allocator.
- ``"allocated_bytes.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
amount of allocated memory.
- ``"segment.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
number of reserved segments from device memory allocation.
- ``"reserved_bytes.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
amount of reserved memory.
- ``"active.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
number of active memory blocks.
- ``"active_bytes.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
amount of active memory.
- ``"inactive_split.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
number of inactive, non-releasable memory blocks.
- ``"inactive_split_bytes.{all,large_pool,small_pool}.{current,peak,allocated,freed}"``:
amount of inactive, non-releasable memory.
For these core statistics, values are broken down as follows.
Pool type:
- ``all``: combined statistics across all memory pools.
- ``large_pool``: statistics for the large allocation pool
(as of June 2025, for size >= 1MB allocations).
- ``small_pool``: statistics for the small allocation pool
(as of June 2025, for size < 1MB allocations).
Metric type:
- ``current``: current value of this metric.
- ``peak``: maximum value of this metric.
- ``allocated``: historical total increase in this metric.
- ``freed``: historical total decrease in this metric.
In addition to the core statistics, we also provide some simple event
counters:
- ``"num_alloc_retries"``: number of failed device memory allocation calls that
result in a cache flush and retry.
- ``"num_ooms"``: number of out-of-memory errors thrown.
- ``"num_sync_all_streams"``: number of ``synchronize_and_free_events`` calls.
- ``"num_device_alloc"``: number of device memory allocation calls.
- ``"num_device_free"``: number of device memory free calls.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
"""
if not torch._C._accelerator_isAllocatorInitialized():
return OrderedDict()
device_index = _get_device_index(device_index, optional=True)
stats = torch._C._accelerator_getDeviceStats(device_index)
flat_stats = []
def flatten(prefix: str, value: Any) -> None:
if isinstance(value, dict):
for k, v in value.items():
nested_prefix = f"{prefix}.{k}" if prefix else k
flatten(nested_prefix, v)
else:
flat_stats.append((prefix, value))
flatten("", stats)
flat_stats.sort()
return OrderedDict(flat_stats)
def memory_allocated(device_index: _device_t = None, /) -> int:
r"""Return the current :ref:`accelerator<accelerators>` device memory occupied by tensors
in bytes for a given device index.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
"""
return memory_stats(device_index).get("allocated_bytes.all.current", 0)
def max_memory_allocated(device_index: _device_t = None, /) -> int:
r"""Return the current :ref:`accelerator<accelerators>` maximum device memory occupied by tensors
in bytes for a given device index.
By default, this returns the peak allocated memory since the beginning of
this program. :func:`~torch.accelerator.reset_peak_memory_stats` can be used to
reset the starting point in tracking this metric.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
"""
return memory_stats(device_index).get("allocated_bytes.all.peak", 0)
def memory_reserved(device_index: _device_t = None, /) -> int:
r"""Return the current :ref:`accelerator<accelerators>` device memory managed by the caching allocator
in bytes for a given device index.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
"""
return memory_stats(device_index).get("reserved_bytes.all.current", 0)
def max_memory_reserved(device_index: _device_t = None, /) -> int:
r"""Return the current :ref:`accelerator<accelerators>` maximum device memory managed by the caching allocator
in bytes for a given device index.
By default, this returns the peak cached memory since the beginning of this
program. :func:`~torch.accelerator.reset_peak_memory_stats` can be used to reset
the starting point in tracking this metric.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
"""
return memory_stats(device_index).get("reserved_bytes.all.peak", 0)
def reset_accumulated_memory_stats(device_index: _device_t = None, /) -> None:
r"""Reset the "accumulated" (historical) stats tracked by the current :ref:`accelerator<accelerators>`
memory allocator for a given device index.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
.. note:: This function is a no-op if the memory allocator for the current
:ref:`accelerator <accelerators>` has not been initialized.
"""
device_index = _get_device_index(device_index, optional=True)
return torch._C._accelerator_resetAccumulatedStats(device_index)
def reset_peak_memory_stats(device_index: _device_t = None, /) -> None:
r"""Reset the "peak" stats tracked by the current :ref:`accelerator<accelerators>`
memory allocator for a given device index.
Args:
device_index (:class:`torch.device`, str, int, optional): the index of the device to target.
If not given, use :func:`torch.accelerator.current_device_index` by default.
If a :class:`torch.device` or str is provided, its type must match the current
:ref:`accelerator<accelerators>` device type.
.. note:: This function is a no-op if the memory allocator for the current
:ref:`accelerator <accelerators>` has not been initialized.
"""
device_index = _get_device_index(device_index, optional=True)
return torch._C._accelerator_resetPeakStats(device_index)

64
torch/csrc/DeviceAccelerator.cpp
View File

@ -77,70 +77,6 @@ void initModule(PyObject* module) {
m.def("_accelerator_setAllocatorSettings", [](std::string env) {
c10::CachingAllocator::setAllocatorSettings(env);
});
m.def("_accelerator_isAllocatorInitialized", []() {
const auto device_type = at::accelerator::getAccelerator(true).value();
return at::getDeviceAllocator(device_type)->initialized();
});
m.def("_accelerator_emptyCache", []() { at::accelerator::emptyCache(); });
m.def("_accelerator_getDeviceStats", [](c10::DeviceIndex device_index) {
using c10::CachingAllocator::Stat;
using c10::CachingAllocator::StatArray;
using c10::CachingAllocator::StatType;
using c10::CachingDeviceAllocator::DeviceStats;
const auto stats = at::accelerator::getDeviceStats(device_index);
const auto stat_to_dict = [](const Stat& stat) -> py::dict {
py::dict dict;
dict["current"] = stat.current;
dict["peak"] = stat.peak;
dict["allocated"] = stat.allocated;
dict["freed"] = stat.freed;
return dict;
};
const auto stat_array_to_dict = [=](const StatArray& stats) -> py::dict {
const std::array<const char*, static_cast<size_t>(StatType::NUM_TYPES)>
kStatTypeNames = {"all", "small_pool", "large_pool"};
py::dict dict;
for (const auto i : c10::irange(kStatTypeNames.size())) {
dict[kStatTypeNames[i]] = stat_to_dict(stats[i]);
}
return dict;
};
py::dict result;
result["num_alloc_retries"] = stats.num_alloc_retries;
result["num_ooms"] = stats.num_ooms;
result["max_split_size"] = stats.max_split_size;
result["num_sync_all_streams"] = stats.num_sync_all_streams;
result["num_device_alloc"] = stats.num_device_alloc;
result["num_device_free"] = stats.num_device_free;
result["allocated_bytes"] = stat_array_to_dict(stats.allocated_bytes);
result["reserved_bytes"] = stat_array_to_dict(stats.reserved_bytes);
result["active_bytes"] = stat_array_to_dict(stats.active_bytes);
result["requested_bytes"] = stat_array_to_dict(stats.requested_bytes);
result["allocation"] = stat_array_to_dict(stats.allocation);
result["segment"] = stat_array_to_dict(stats.segment);
result["active"] = stat_array_to_dict(stats.active);
result["inactive_split"] = stat_array_to_dict(stats.inactive_split);
result["inactive_split_bytes"] =
stat_array_to_dict(stats.inactive_split_bytes);
result["oversize_allocations"] = stat_to_dict(stats.oversize_allocations);
result["oversize_segments"] = stat_to_dict(stats.oversize_segments);
return result;
});
m.def(
"_accelerator_resetAccumulatedStats", [](c10::DeviceIndex device_index) {
at::accelerator::resetAccumulatedStats(device_index);
});
m.def("_accelerator_resetPeakStats", [](c10::DeviceIndex device_index) {
at::accelerator::resetPeakStats(device_index);
});
}
} // namespace torch::accelerator

4
torch/cuda/memory.py
View File

@ -255,9 +255,9 @@ def memory_stats(device: "Device" = None) -> dict[str, Any]:
- ``all``: combined statistics across all memory pools.
- ``large_pool``: statistics for the large allocation pool
(as of June 2025, for size >= 1MB allocations).
(as of October 2019, for size >= 1MB allocations).
- ``small_pool``: statistics for the small allocation pool
(as of June 2025, for size < 1MB allocations).
(as of October 2019, for size < 1MB allocations).
Metric type: