apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Graph Optimization] Refactor default capture list (#4617)

Browse Source 
* fix bug and refine code

* add debug count

* refine code
This commit is contained in:
RAM
2025-10-28 21:31:02 +08:00
committed by GitHub
parent 0a0c74e717
commit fff5fb5e39
3 changed files with 36 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
fastdeploy/model_executor/graph_optimization/cudagraph_piecewise_backend.py
View File

@@ -171,7 +171,7 @@ class CudaGraphPiecewiseBackend:
for n in range(entry.num_finished_warmup, self.warm_up_size):
entry.num_finished_warmup += 1
entry.runnable(**kwargs)
logger.debug(
logger.info(
f"[CUDA GRAPH][ID:{id(self)}] Warm up for real shape {padding_real_shape}, "
f"finished ({n + 1}/{entry.num_finished_warmup}) times"
)
@@ -207,7 +207,7 @@ class CudaGraphPiecewiseBackend:
# For CUDAGraph debug
# self._save_cudagrpah_dot_files(entry)
logger.debug(f"[CUDA GRAPH][ID:{id(self)}] CUDAGraph captured for real shape {padding_real_shape}")
logger.info(f"[CUDA GRAPH][ID:{id(self)}] CUDAGraph captured for real shape {padding_real_shape}")
# Replay
entry.cuda_graph.replay()
@@ -224,7 +224,7 @@ class CudaGraphPiecewiseBackend:
for shape in self.cudagraph_capture_sizes:
self.concrete_size_entries[shape] = ConcreteSizeEntry(real_shape=shape)
logger.debug(
logger.info(
f"[CUDA GRAPH][ID:{id(self)}] CUDAGraph capture list {self.cudagraph_capture_sizes}, "
"Created all real shape entry."
)
@@ -254,3 +254,9 @@ class CudaGraphPiecewiseBackend:
f"{log_dir}/GraphDotFiles/backend{id(self)}_shape{entry.real_shape}",
1 << 0,
)
def check_capture_successful(self):
"""Check whether the shapes are captured or not"""
for shape, entry in self.concrete_size_entries.items():
if not entry.captured:
raise ValueError(f"[CUDA GRAPH][ID:{id(self)}] Shape {shape} capture failed.")

12
fastdeploy/model_executor/graph_optimization/graph_optimization_backend.py
View File

@@ -34,6 +34,10 @@ from fastdeploy.model_executor.graph_optimization.utils import in_profile_run_mo
from fastdeploy.model_executor.graph_optimization.utils import (
in_sot_warmup_mode as in_warmup_mode,
)
from fastdeploy.utils import get_logger
logger = get_logger("cudagrpah_piecewise_backend", "cudagraph_piecewise_backend.log")
P = ParamSpec("P")
T = TypeVar("T")
@@ -105,6 +109,9 @@ class GraphOptBackend:
self.dy_runnable = self.runnable
self.fd_config = fd_config
self.max_captre_size = fd_config.graph_opt_config.cudagraph_capture_sizes[0]
self._debug_count_cudagraph_replay = 0
self._debug_count_total_step = 0
if self.fd_config.graph_opt_config.graph_opt_level > 0:
# 1. Prepare cuda graph input buffers (contain output of subgraphs)
@@ -123,6 +130,7 @@ class GraphOptBackend:
)
def __call__(self, **kwargs):
self._debug_count_total_step += 1
if not self.fd_config.graph_opt_config.use_cudagraph:
return self.runnable(**kwargs)
if self.cudagraph_piecewise_backend is None:
@@ -136,6 +144,10 @@ class GraphOptBackend:
if (not kwargs["forward_meta"].step_use_cudagraph) or (real_shape > self.cudagraph_switch_threshold):
return self.dy_runnable(**kwargs)
else:
self._debug_count_cudagraph_replay += 1
logger.debug(
f"[CUDA GRAPH][ID:{id(self.cudagraph_piecewise_backend)}] Total step count: {self._debug_count_total_step}, CUDAGraph replay count: {self._debug_count_cudagraph_replay}"
)
return self.cudagraph_piecewise_backend.__call__(**kwargs)
def clear_cudagraph_piecewise_backend(self):