apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

polish code with new pre-commit rule (#2923)

Browse Source
This commit is contained in:
Zero Rains
2025-07-19 23:19:27 +08:00
committed by GitHub
parent b8676d71a8
commit 25698d56d1
424 changed files with 14307 additions and 13518 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fastdeploy/distributed/custom_all_reduce/__init__.py
View File

@@ -14,4 +14,4 @@
from .custom_all_reduce import CustomAllreduce
__all__ = ["CustomAllreduce"]
__all__ = ["CustomAllreduce"]

40
fastdeploy/distributed/custom_all_reduce/cuda_wrapper.py
View File

@@ -41,7 +41,7 @@ def find_loaded_library(lib_name) -> Optional[str]:
the file `/proc/self/maps` contains the memory maps of the process, which includes the
shared libraries loaded by the process. We can use this file to find the path of the
a loaded library.
""" # noqa
"""
found = False
with open("/proc/self/maps") as f:
for line in f:
@@ -73,18 +73,40 @@ class CudaRTLibrary:
# const char* cudaGetErrorString ( cudaError_t error )
Function("cudaGetErrorString", ctypes.c_char_p, [cudaError_t]),
# ​cudaError_t cudaMalloc ( void** devPtr, size_t size )
Function("cudaMalloc", cudaError_t, [ctypes.POINTER(ctypes.c_void_p), ctypes.c_size_t]),
Function(
"cudaMalloc",
cudaError_t,
[ctypes.POINTER(ctypes.c_void_p), ctypes.c_size_t],
),
# ​cudaError_t cudaFree ( void* devPtr )
Function("cudaFree", cudaError_t, [ctypes.c_void_p]),
# ​cudaError_t cudaMemset ( void* devPtr, int value, size_t count )
Function("cudaMemset", cudaError_t, [ctypes.c_void_p, ctypes.c_int, ctypes.c_size_t]),
# ​cudaError_t cudaMemcpy ( void* dst, const void* src, size_t count, cudaMemcpyKind kind ) # noqa
Function("cudaMemcpy", cudaError_t, [ctypes.c_void_p, ctypes.c_void_p, ctypes.c_size_t, cudaMemcpyKind]),
# cudaError_t cudaIpcGetMemHandle ( cudaIpcMemHandle_t* handle, void* devPtr ) # noqa
Function("cudaIpcGetMemHandle", cudaError_t, [ctypes.POINTER(cudaIpcMemHandle_t), ctypes.c_void_p]),
# ​cudaError_t cudaIpcOpenMemHandle ( void** devPtr, cudaIpcMemHandle_t handle, unsigned int flags ) # noqa
Function(
"cudaIpcOpenMemHandle", cudaError_t, [ctypes.POINTER(ctypes.c_void_p), cudaIpcMemHandle_t, ctypes.c_uint]
"cudaMemset",
cudaError_t,
[ctypes.c_void_p, ctypes.c_int, ctypes.c_size_t],
),
# ​cudaError_t cudaMemcpy ( void* dst, const void* src, size_t count, cudaMemcpyKind kind )
Function(
"cudaMemcpy",
cudaError_t,
[ctypes.c_void_p, ctypes.c_void_p, ctypes.c_size_t, cudaMemcpyKind],
),
# cudaError_t cudaIpcGetMemHandle ( cudaIpcMemHandle_t* handle, void* devPtr )
Function(
"cudaIpcGetMemHandle",
cudaError_t,
[ctypes.POINTER(cudaIpcMemHandle_t), ctypes.c_void_p],
),
# ​cudaError_t cudaIpcOpenMemHandle ( void** devPtr, cudaIpcMemHandle_t handle, unsigned int flags )
Function(
"cudaIpcOpenMemHandle",
cudaError_t,
[
ctypes.POINTER(ctypes.c_void_p),
cudaIpcMemHandle_t,
ctypes.c_uint,
],
),
]

29
fastdeploy/distributed/custom_all_reduce/custom_all_reduce.py
View File

@@ -13,26 +13,26 @@
# limitations under the License.
from contextlib import contextmanager
import atexit
import ctypes
from contextlib import contextmanager
from typing import List, Optional
import paddle
import paddle.distributed as dist
from paddle.distributed.communication.group import Group
from fastdeploy.distributed.custom_all_reduce import cuda_wrapper
from fastdeploy.model_executor.ops.gpu import (
all_reduce,
dispose,
get_graph_buffer_ipc_meta,
init_custom_all_reduce,
meta_size,
register_buffer,
get_graph_buffer_ipc_meta,
register_graph_buffers,
)
from fastdeploy.distributed.custom_all_reduce import cuda_wrapper
try:
meta_size()
custom_ar = True
@@ -47,7 +47,7 @@ class CustomAllreduce:
_SUPPORTED_WORLD_SIZES = [2, 4, 6, 8]
# max_size: max supported allreduce size
def __init__(self, group: Group, max_size: int=8192 * 1024) -> None:
def __init__(self, group: Group, max_size: int = 8192 * 1024) -> None:
"""
Args:
device: the device to bind the CustomAllreduce to. If None,
@@ -147,7 +147,12 @@ class CustomAllreduce:
return inp_size < self.max_size
return False
def all_reduce(self, inp: paddle.Tensor, out: paddle.Tensor = None, registered: bool = False):
def all_reduce(
self,
inp: paddle.Tensor,
out: paddle.Tensor = None,
registered: bool = False,
):
"""Performs an out-of-place all reduce.
If registered is True, this assumes inp's pointer is already
@@ -165,7 +170,7 @@ class CustomAllreduce:
@contextmanager
def capture(self):
"""
The main responsibility of this context manager is the
The main responsibility of this context manager is the
`register_graph_buffers` call at the end of the context.
It records all the buffer addresses used in the CUDA graph.
"""
@@ -179,22 +184,18 @@ class CustomAllreduce:
def register_graph_buffers(self):
handle, offset = get_graph_buffer_ipc_meta(self._ptr)
all_data = [[None, None]
for _ in range(dist.get_world_size(group=self.group))]
all_data = [[None, None] for _ in range(dist.get_world_size(group=self.group))]
all_data[self.rank] = [handle, offset]
ranks = sorted(dist.get_process_group_ranks(group=self.group))
for i, rank in enumerate(ranks):
dist.broadcast_object_list(all_data[i],
src=rank,
group=self.group,
device="cpu")
dist.broadcast_object_list(all_data[i], src=rank, group=self.group, device="cpu")
# Unpack list of tuples to tuple of lists.
handles = [d[0] for d in all_data] # type: ignore
offsets = [d[1] for d in all_data] # type: ignore
register_graph_buffers(self._ptr, handles, offsets)
def custom_all_reduce(self, input: paddle.Tensor) -> Optional[paddle.Tensor]:
"""The main allreduce API that provides support for cuda graph."""
# When custom allreduce is disabled, this will be None.