apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-26 20:41:53 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[CudaGraph] support cudagraph use shared pool (#4199)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details

Browse Source 
* support cudagraph use shared pool

* add envs

* change CUDAGRAPH_POOL_ID to int

* change CUDAGRAPH_POOL_ID to use_memory_pool

* unify use_unique_memory_pool

* fix use_unique_memory_pool
This commit is contained in:
lizhenyun01
2025-09-24 21:32:04 +08:00
committed by GitHub
parent e2b68b33c9
commit bab779011c
2 changed files with 12 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
fastdeploy/config.py
View File

@@ -588,6 +588,9 @@ class GraphOptimizationConfig:
Thus this flag cannot be used together with splitting_ops.""" Thus this flag cannot be used together with splitting_ops."""
self.full_cuda_graph: bool = True self.full_cuda_graph: bool = True
""" Whether to use shared memory pool for multi capture_size """
self.use_unique_memory_pool: bool = False
self.max_capture_size: int = None self.max_capture_size: int = None
self.real_shape_to_captured_size: dict[int, int] = None self.real_shape_to_captured_size: dict[int, int] = None
# CINN Config ... # CINN Config ...

16
fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py
View File

@@ -20,6 +20,7 @@ from typing import Callable, Dict, List, Optional
import paddle.jit.dy2static.utils as jit_utils import paddle.jit.dy2static.utils as jit_utils
import paddle.nn.layer import paddle.nn.layer
from paddle.base.core import CUDAGraph
from paddle.device.cuda import graphs from paddle.device.cuda import graphs
from fastdeploy import envs from fastdeploy import envs
@@ -85,17 +86,14 @@ class Dy2StCudaGraphManager:
class CudaGraphPiecewiseBackend: class CudaGraphPiecewiseBackend:
"""Manage the capture and replay of CUDA graphs at the subgraph level.""" """Manage the capture and replay of CUDA graphs at the subgraph level."""
def __init__( def __init__(self, fd_config: FDConfig, runnable: Callable):
self,
fd_config: FDConfig,
runnable: Callable,
):
self.fd_config = fd_config self.fd_config = fd_config
self.runnable = runnable self.runnable = runnable
self.cudagraph_capture_sizes = fd_config.graph_opt_config.cudagraph_capture_sizes self.cudagraph_capture_sizes = fd_config.graph_opt_config.cudagraph_capture_sizes
self.warm_up_size = fd_config.graph_opt_config.cudagraph_num_of_warmups self.warm_up_size = fd_config.graph_opt_config.cudagraph_num_of_warmups
self.real_shape_to_captured_size = fd_config.graph_opt_config.real_shape_to_captured_size self.real_shape_to_captured_size = fd_config.graph_opt_config.real_shape_to_captured_size
if self.fd_config.graph_opt_config.use_unique_memory_pool:
self.unique_memory_pool_id = CUDAGraph.gen_new_memory_pool_id()
self._create_entry_dict() self._create_entry_dict()
self.cuda_graph_manager = None self.cuda_graph_manager = None
@@ -168,7 +166,11 @@ class CudaGraphPiecewiseBackend:
input_addresses = [x.data_ptr() for (_, x) in kwargs.items() if isinstance(x, paddle.Tensor)] input_addresses = [x.data_ptr() for (_, x) in kwargs.items() if isinstance(x, paddle.Tensor)]
entry.input_addresses = input_addresses entry.input_addresses = input_addresses
new_grpah = graphs.CUDAGraph() new_grpah = (
graphs.CUDAGraph(pool_id=self.unique_memory_pool_id)
if self.fd_config.graph_opt_config.use_unique_memory_pool
else graphs.CUDAGraph()
)
paddle.device.synchronize() paddle.device.synchronize()
# Capture # Capture