torch-memory-saver 0.0.8__cp39-abi3-manylinux2014_x86_64.whl → 0.0.9rc1__cp39-abi3-manylinux2014_x86_64.whl

torch_memory_saver/__init__.py

@@ -1,135 +1,5 @@
-import ctypes
-import logging
-import os
-from contextlib import contextmanager
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Optional
+from .entrypoint import TorchMemorySaver
+from .hooks.mode_preload import configure_subprocess
 
-import torch
-
-logger = logging.getLogger(__name__)
-
-
-class TorchMemorySaver:
-    def __init__(self):
-        self._mem_pool = None
-
-    @contextmanager
-    def region(self, tag: str = "default"):
-        """Context manager for memory saving with optional tag"""
-        if _global_info.binary_info.enabled:
-            self._ensure_mem_pool()
-            with torch.cuda.use_mem_pool(self._mem_pool):
-                _global_info.binary_info.cdll.tms_set_current_tag(tag.encode('utf-8'))
-                _global_info.binary_info.cdll.tms_region_enter()
-                try:
-                    yield
-                finally:
-                    _global_info.binary_info.cdll.tms_set_current_tag(b"default")
-                    _global_info.binary_info.cdll.tms_region_leave()
-        else:
-            yield
-
-    def pause(self, tag: Optional[str] = None):
-        """Pause memory for specific tag or all memory if tag is None"""
-        if _global_info.binary_info.enabled:
-            tag_bytes = tag.encode('utf-8') if tag else None
-            _global_info.binary_info.cdll.tms_pause(tag_bytes)
-
-    def resume(self, tag: Optional[str] = None):
-        """Resume memory for specific tag or all memory if tag is None"""
-        if _global_info.binary_info.enabled:
-            tag_bytes = tag.encode('utf-8') if tag else None
-            _global_info.binary_info.cdll.tms_resume(tag_bytes)
-
-    @property
-    def enabled(self):
-        return _global_info.binary_info.enabled
-
-    def _ensure_mem_pool(self):
-        if self._mem_pool is None:
-            self._mem_pool = torch.cuda.MemPool()
-
-@dataclass
-class _BinaryInfo:
-    cdll: Optional[ctypes.CDLL]
-
-    @property
-    def enabled(self):
-        return self.cdll is not None
-
-    @staticmethod
-    def _setup_function_signatures(cdll):
-        """Define function signatures for the C library"""
-        cdll.tms_region_enter.argtypes = []
-        cdll.tms_region_leave.argtypes = []
-        cdll.tms_set_current_tag.argtypes = [ctypes.c_char_p]
-        cdll.tms_pause.argtypes = [ctypes.c_char_p]
-        cdll.tms_resume.argtypes = [ctypes.c_char_p]
-
-    @staticmethod
-    def compute():
-        env_ld_preload = os.environ.get('LD_PRELOAD', '')
-        if 'torch_memory_saver' in env_ld_preload:
-            try:
-                cdll = ctypes.CDLL(env_ld_preload)
-                _BinaryInfo._setup_function_signatures(cdll)
-                return _BinaryInfo(cdll=cdll)
-            except OSError as e:
-                logger.error(f'Failed to load CDLL from {env_ld_preload}: {e}')
-                return _BinaryInfo(cdll=None)
-        else:
-            print(
-                f'TorchMemorySaver is disabled for the current process because invalid LD_PRELOAD. '
-                f'You can use configure_subprocess() utility, '
-                f'or directly specify `LD_PRELOAD=/path/to/torch_memory_saver_cpp.some-postfix.so python your_script.py. '
-                f'(LD_PRELOAD="{env_ld_preload}" process_id={os.getpid()})'
-            )
-            return _BinaryInfo(cdll=None)
-
-
-class _GlobalInfo:
-    def __init__(self):
-        self._binary_info: Optional[_BinaryInfo] = None
-
-    @property
-    def binary_info(self):
-        if self._binary_info is None:
-            self._binary_info = _BinaryInfo.compute()
-        return self._binary_info
-
-
-_global_info = _GlobalInfo()
-
-# Global singleton instance
+# Global singleton
 torch_memory_saver = TorchMemorySaver()
-
-def get_binary_path():
-    dir_package = Path(__file__).parent
-    candidates = [
-        p
-        for d in [dir_package, dir_package.parent]
-        for p in d.glob('torch_memory_saver_cpp.*.so')
-    ]
-    assert len(candidates) == 1, f'Expected exactly one torch_memory_saver_cpp library, found: {candidates}'
-    return candidates[0]
-
-
-@contextmanager
-def configure_subprocess():
-    with change_env('LD_PRELOAD', str(get_binary_path())):
-        yield
-
-
-@contextmanager
-def change_env(key: str, value: str):
-    old_value = os.environ.get(key, '')
-    os.environ[key] = value
-    logger.debug(f'change_env set key={key} value={value}')
-    try:
-        yield
-    finally:
-        assert os.environ[key] == value
-        os.environ[key] = old_value
-        logger.debug(f'change_env restore key={key} value={old_value}')

torch_memory_saver/binary_wrapper.py

@@ -0,0 +1,31 @@
+import ctypes
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class BinaryWrapper:
+    def __init__(self, path_binary: str):
+        try:
+            self.cdll = ctypes.CDLL(path_binary)
+        except OSError as e:
+            logger.error(f"Failed to load CDLL from {path_binary}: {e}")
+            raise
+
+        _setup_function_signatures(self.cdll)
+
+    def set_config(self, *, tag: str, interesting_region: bool, enable_cpu_backup: bool):
+        self.cdll.tms_set_current_tag(tag.encode("utf-8"))
+        self.cdll.tms_set_interesting_region(interesting_region)
+        self.cdll.tms_set_enable_cpu_backup(enable_cpu_backup)
+
+
+def _setup_function_signatures(cdll):
+    """Define function signatures for the C library"""
+    cdll.tms_set_current_tag.argtypes = [ctypes.c_char_p]
+    cdll.tms_set_interesting_region.argtypes = [ctypes.c_bool]
+    cdll.tms_get_interesting_region.restype = ctypes.c_bool
+    cdll.tms_set_enable_cpu_backup.argtypes = [ctypes.c_bool]
+    cdll.tms_get_enable_cpu_backup.restype = ctypes.c_bool
+    cdll.tms_pause.argtypes = [ctypes.c_char_p]
+    cdll.tms_resume.argtypes = [ctypes.c_char_p]

torch_memory_saver/entrypoint.py

@@ -0,0 +1,142 @@
+import ctypes
+import logging
+import os
+from contextlib import contextmanager
+from typing import Optional
+import torch
+
+from .binary_wrapper import BinaryWrapper
+from .hooks.base import HookUtilBase, HookMode
+
+logger = logging.getLogger(__name__)
+
+_TAG_DEFAULT = "default"
+
+
+class TorchMemorySaver:
+    def __init__(self):
+        self._impl_ctor_kwargs = {}
+        self._impl: Optional[_TorchMemorySaverImpl] = None
+
+    @contextmanager
+    def region(self, tag: str = _TAG_DEFAULT, enable_cpu_backup: bool = False):
+        """Context manager for memory saving with optional tag"""
+        self._ensure_initialized()
+        with self._impl.region(tag=tag, enable_cpu_backup=enable_cpu_backup):
+            yield
+
+    @contextmanager
+    def cuda_graph(
+            self,
+            cuda_graph, pool=None, stream=None, capture_error_mode='global',
+            tag: str = _TAG_DEFAULT, enable_cpu_backup: bool = False,
+    ):
+        """Similar to `torch.cuda.graph`, but ensures memory in it to be pauseable."""
+        self._ensure_initialized()
+        with self._impl.cuda_graph(
+                cuda_graph=cuda_graph,
+                pool=pool, stream=stream, capture_error_mode=capture_error_mode,
+                tag=tag, enable_cpu_backup=enable_cpu_backup,
+        ):
+            yield
+
+    @contextmanager
+    def disable(self):
+        self._ensure_initialized()
+        with self._impl.disable():
+            yield
+
+    def pause(self, tag: Optional[str] = None):
+        """Pause memory for specific tag or all memory if tag is None"""
+        self._ensure_initialized()
+        self._impl.pause(tag=tag)
+
+    def resume(self, tag: Optional[str] = None):
+        """Resume memory for specific tag or all memory if tag is None"""
+        self._ensure_initialized()
+        self._impl.resume(tag=tag)
+
+    # for compatibility
+    @property
+    def enabled(self):
+        return True
+
+    @property
+    def hook_mode(self):
+        raise AttributeError
+
+    @hook_mode.setter
+    def hook_mode(self, hook_mode: HookMode):
+        assert self._impl_ctor_kwargs is not None, "Cannot configure after initialization"
+        self._impl_ctor_kwargs["hook_mode"] = hook_mode
+
+    def _ensure_initialized(self):
+        if self._impl is not None:
+            return
+        self._impl = _TorchMemorySaverImpl(**self._impl_ctor_kwargs)
+        del self._impl_ctor_kwargs
+
+
+class _TorchMemorySaverImpl:
+    def __init__(self, hook_mode: HookMode = "preload"):
+        self._hook_mode = hook_mode
+        self._hook_util = HookUtilBase.create(hook_mode=hook_mode)
+        self._binary_wrapper = BinaryWrapper(path_binary=self._hook_util.get_path_binary())
+        self._primary_mem_pool = torch.cuda.MemPool(allocator=self._hook_util.get_allocator())
+        _sanity_checks()
+
+    @contextmanager
+    def region(self, tag: str, enable_cpu_backup: bool):
+        with torch.cuda.use_mem_pool(self._primary_mem_pool):
+            with self._with_region_config(tag=tag, enable_cpu_backup=enable_cpu_backup):
+                yield
+
+    @contextmanager
+    def cuda_graph(self, cuda_graph, pool, stream, capture_error_mode, tag: str, enable_cpu_backup: bool):
+        assert self._hook_mode == "preload", "Only hook_mode=preload supports pauseable CUDA Graph currently"
+        with torch.cuda.graph(cuda_graph, pool=pool, stream=stream, capture_error_mode=capture_error_mode):
+            with self._with_region_config(tag=tag, enable_cpu_backup=enable_cpu_backup):
+                yield
+
+    @contextmanager
+    def _with_region_config(self, tag: str, enable_cpu_backup: bool):
+        assert not self._binary_wrapper.cdll.tms_get_interesting_region()
+        original_enable_cpu_backup = self._binary_wrapper.cdll.tms_get_enable_cpu_backup()
+
+        self._binary_wrapper.set_config(tag=tag, interesting_region=True, enable_cpu_backup=enable_cpu_backup)
+        try:
+            yield
+        finally:
+            assert self._binary_wrapper.cdll.tms_get_interesting_region()
+            self._binary_wrapper.set_config(tag=_TAG_DEFAULT, interesting_region=False, enable_cpu_backup=original_enable_cpu_backup)
+
+    @contextmanager
+    def disable(self, dispose_mem_pool_after_use: bool = True):
+        assert dispose_mem_pool_after_use, "Only dispose_mem_pool_after_use=true is supported now"
+        assert self._binary_wrapper.cdll.tms_get_interesting_region(), "disable() should be called only when tms is active"
+
+        self._binary_wrapper.cdll.tms_set_interesting_region(False)
+        try:
+            # We can either reuse the pool or delete it immediately, and we implement the latter currently since Slime uses it.
+            # About why we need a pool: https://github.com/fzyzcjy/torch_memory_saver/pull/20#issuecomment-3047099047
+            pool = torch.cuda.MemPool()
+            with torch.cuda.use_mem_pool(pool):
+                yield
+            del pool
+        finally:
+            self._binary_wrapper.cdll.tms_set_interesting_region(True)
+
+    def pause(self, tag: Optional[str]):
+        tag_bytes = tag.encode("utf-8") if tag else None
+        self._binary_wrapper.cdll.tms_pause(tag_bytes)
+
+    def resume(self, tag: Optional[str]):
+        tag_bytes = tag.encode("utf-8") if tag else None
+        self._binary_wrapper.cdll.tms_resume(tag_bytes)
+
+
+def _sanity_checks():
+    if "expandable_segments:True" in os.environ.get("PYTORCH_CUDA_ALLOC_CONF", ""):
+        raise RuntimeError(
+            "TorchMemorySaver is disabled for the current process because expandable_segments is not supported yet."
+        )

torch_memory_saver/hooks/base.py

@@ -0,0 +1,21 @@
+from abc import ABC
+from typing import Literal
+
+HookMode = Literal["preload", "torch"]
+
+
+class HookUtilBase(ABC):
+    @staticmethod
+    def create(hook_mode: HookMode) -> "HookUtilBase":
+        from torch_memory_saver.hooks.mode_preload import HookUtilModePreload
+        from torch_memory_saver.hooks.mode_torch import HookUtilModeTorch
+        return {
+            "preload": HookUtilModePreload,
+            "torch": HookUtilModeTorch,
+        }[hook_mode]()
+
+    def get_path_binary(self):
+        raise NotImplementedError
+
+    def get_allocator(self):
+        return None

torch_memory_saver/hooks/mode_preload.py

@@ -0,0 +1,26 @@
+import logging
+import os
+from contextlib import contextmanager
+from torch_memory_saver.hooks.base import HookUtilBase
+from torch_memory_saver.utils import get_binary_path_from_package, change_env
+
+logger = logging.getLogger(__name__)
+
+
+class HookUtilModePreload(HookUtilBase):
+    def get_path_binary(self):
+        env_ld_preload = os.environ.get("LD_PRELOAD", "")
+        assert "torch_memory_saver" in env_ld_preload, (
+            f"TorchMemorySaver observes invalid LD_PRELOAD. "
+            f"You can use configure_subprocess() utility, "
+            f"or directly specify `LD_PRELOAD=/path/to/torch_memory_saver_cpp.some-postfix.so python your_script.py. "
+            f'(LD_PRELOAD="{env_ld_preload}" process_id={os.getpid()})'
+        )
+        return env_ld_preload
+
+
+@contextmanager
+def configure_subprocess():
+    """Configure environment variables for subprocesses. Only needed for hook_mode=preload."""
+    with change_env("LD_PRELOAD", str(get_binary_path_from_package("torch_memory_saver_hook_mode_preload"))):
+        yield

torch_memory_saver/hooks/mode_torch.py

@@ -0,0 +1,19 @@
+import logging
+
+from torch_memory_saver.hooks.base import HookUtilBase
+from torch_memory_saver.utils import get_binary_path_from_package
+from torch.cuda.memory import CUDAPluggableAllocator
+
+logger = logging.getLogger(__name__)
+
+
+class HookUtilModeTorch(HookUtilBase):
+    def __init__(self):
+        self.allocator = CUDAPluggableAllocator(self.get_path_binary(), "tms_torch_malloc", "tms_torch_free")
+        logger.debug(f"HookUtilModeTorch {self.allocator=} {self.get_path_binary()=}")
+
+    def get_path_binary(self):
+        return str(get_binary_path_from_package("torch_memory_saver_hook_mode_torch"))
+
+    def get_allocator(self):
+        return self.allocator.allocator()

torch_memory_saver/testing_utils.py

@@ -0,0 +1,10 @@
+"""Not to be used by end users, but only for tests of the package itself."""
+
+import torch
+
+
+def get_and_print_gpu_memory(message, gpu_id=0):
+    """Print GPU memory usage with optional message"""
+    mem = torch.cuda.device_memory_used(gpu_id)
+    print(f"GPU {gpu_id} memory: {mem / 1024 ** 3:.2f} GB ({message})")
+    return mem

torch_memory_saver/utils.py

@@ -0,0 +1,27 @@
+import logging
+import os
+from contextlib import contextmanager
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+
+def get_binary_path_from_package(stem: str):
+    dir_package = Path(__file__).parent
+    candidates = [p for d in [dir_package, dir_package.parent] for p in d.glob(f"{stem}.*.so")]
+    assert len(candidates) == 1, f"Expected exactly one torch_memory_saver_cpp library, found: {candidates}"
+    return candidates[0]
+
+
+# private utils, not to be used by end users
+@contextmanager
+def change_env(key: str, value: str):
+    old_value = os.environ.get(key, "")
+    os.environ[key] = value
+    logger.debug(f"change_env set key={key} value={value}")
+    try:
+        yield
+    finally:
+        assert os.environ[key] == value
+        os.environ[key] = old_value
+        logger.debug(f"change_env restore key={key} value={old_value}")

{torch_memory_saver-0.0.8.dist-info → torch_memory_saver-0.0.9rc1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: torch_memory_saver
-Version: 0.0.8
+Version: 0.0.9rc1
 Requires-Python: >=3.9
 License-File: LICENSE
 

torch_memory_saver-0.0.9rc1.dist-info/RECORD

@@ -0,0 +1,16 @@
+torch_memory_saver_hook_mode_preload.abi3.so,sha256=dQ6PvuqKNXTsbAPSWxP86YkQL82Vo-ngJ9WqB4Zbiss,777208
+torch_memory_saver_hook_mode_torch.abi3.so,sha256=-NE71rUpuYC3Sh8EqaMWbdMbm9C4h6MxeC30gNoqGys,781032
+torch_memory_saver/__init__.py,sha256=9iU_QlTe6OxMR5_OtSRUmvr6ltzk149GjojYvG74sag,154
+torch_memory_saver/binary_wrapper.py,sha256=MeQlPHIuFycamcWp3kOXjVZMiEK8HONuSx4l92J4k_Q,1133
+torch_memory_saver/entrypoint.py,sha256=aFkgqnWRI8vF8EeAL4FvIY33dNVtIbUMk1eM3_xH-fs,5538
+torch_memory_saver/testing_utils.py,sha256=vd9jhMgBLbeEy3vdvbuCjjtO-lRSX-RVB_Dg-wSHVQM,332
+torch_memory_saver/utils.py,sha256=LhtiocZTpMyDEjSexXaGglQtOJeJB7AaH5s43PZX5yo,856
+torch_memory_saver/hooks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+torch_memory_saver/hooks/base.py,sha256=f8Rv_XxNupU80dKWUgE-Ea5pu1qaoXcnzltZrDy90hY,579
+torch_memory_saver/hooks/mode_preload.py,sha256=ELaVloI7T-rjssxn6lujaknNfujxFBxA2oc0SOsiUfk,1041
+torch_memory_saver/hooks/mode_torch.py,sha256=yxGyA8AYrKX7hr3Bawr_MH2AMwSgXxLN76GZaAbQLGU,681
+torch_memory_saver-0.0.9rc1.dist-info/LICENSE,sha256=i806R5xShJFB4k9yNQJ2GYCcSBlu1frTx2vH_nWdWE8,1064
+torch_memory_saver-0.0.9rc1.dist-info/METADATA,sha256=PAMArqZ3_juC25gvir7W_--VffaTK9tO1-422cWOfnA,111
+torch_memory_saver-0.0.9rc1.dist-info/WHEEL,sha256=HUPiMa7ZA9BvJ9gdJRYwZIjK2rWbCcrqYvJ4Onw0owE,102
+torch_memory_saver-0.0.9rc1.dist-info/top_level.txt,sha256=Fdob5gbD3sjPAe3kNfDokaN1sL43cMvwKRLKuR8oitw,91
+torch_memory_saver-0.0.9rc1.dist-info/RECORD,,

torch_memory_saver-0.0.9rc1.dist-info/top_level.txt

@@ -0,0 +1,3 @@
+torch_memory_saver
+torch_memory_saver_hook_mode_preload
+torch_memory_saver_hook_mode_torch

torch_memory_saver-0.0.8.dist-info/RECORD

@@ -1,7 +0,0 @@
-torch_memory_saver_cpp.abi3.so,sha256=kI54XLX1E_R-PgmfwrZoM2-emKOWQXRoabNEMZS0dCE,391672
-torch_memory_saver/__init__.py,sha256=MfMSSGSNhP7xBo4jq6RPug6svnMVPtmua8-l0yWlkTg,4403
-torch_memory_saver-0.0.8.dist-info/LICENSE,sha256=i806R5xShJFB4k9yNQJ2GYCcSBlu1frTx2vH_nWdWE8,1064
-torch_memory_saver-0.0.8.dist-info/METADATA,sha256=tUjtugtoDTFWeIct5kTvWRbGUUTlJlG85h_zSClzUdA,108
-torch_memory_saver-0.0.8.dist-info/WHEEL,sha256=HUPiMa7ZA9BvJ9gdJRYwZIjK2rWbCcrqYvJ4Onw0owE,102
-torch_memory_saver-0.0.8.dist-info/top_level.txt,sha256=uJ27-bVSKHxdcfHRcakvEr_KQxnUlMia6v19fHbfHxA,42
-torch_memory_saver-0.0.8.dist-info/RECORD,,

torch_memory_saver-0.0.8.dist-info/top_level.txt

@@ -1,2 +0,0 @@
-torch_memory_saver
-torch_memory_saver_cpp