Files
2026-07-13 12:40:42 +08:00

283 lines
8.9 KiB
Python

# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import io
import os
import re
import sys
import tempfile
from contextlib import contextmanager
import paddle
from paddle.base.wrapped_decorator import signature_safe_contextmanager
def _parse_tensors(input_str):
input_tuples = re.findall(
r"\(\s*(\w+)\s*,\s*(\{.*?\}|\[.*?\])\s*\)", input_str, re.DOTALL
)
tensors_list = []
for var_name, value_str in input_tuples:
# Only one Tensor
if value_str.startswith('{'):
tensor_info = _parse_tensor_info(value_str)
tensors_list.append((var_name, [tensor_info]))
# Tensor list
elif value_str.startswith('['):
list_content = value_str.strip('[]')
dict_strs = _split_list_elements(list_content)
tensor_list = []
for dict_str in dict_strs:
tensor_info = _parse_tensor_info(dict_str)
if tensor_info:
tensor_list.append(tensor_info)
tensors_list.append((var_name, tensor_list))
return tensors_list
def _parse_api_name(debug_str):
api_name_match = re.search(r"API_Name:\s*(\w+)", debug_str)
return api_name_match.group(1) if api_name_match else None
def _parse_input_tensors(debug_str):
input_match = re.search(r"Input:\s*(\[.*?\] )", debug_str, re.DOTALL)
return _parse_tensors(input_match.group(1)) if input_match else []
def _parse_output_tensors(debug_str):
output_match = re.search(r"Output:\s*(\[.*?\] )", debug_str, re.DOTALL)
return _parse_tensors(output_match.group(1)) if output_match else []
def _parse_tensor_info(dict_str):
result = {}
lines = dict_str.strip().split('\n')
for line in lines:
line = line.strip()
if line and ':' in line:
key, value = line.split(':', 1)
key = key.strip()
value = value.strip()
if key == "Place":
place_match = re.search(r"Place\((\w+:\d+)\)", value)
if place_match:
value = place_match.group(1)
elif key == "Shape":
value = [int(x.strip()) for x in value.split(',') if x.strip()]
elif value == "None":
value = None
result[key] = value
return result
def _split_list_elements(list_str):
elements = []
current = []
brace_count = 0
for char in list_str:
if char == '{':
brace_count += 1
elif char == '}':
brace_count -= 1
if char == ',' and brace_count == 0:
elements.append(''.join(current).strip())
current = []
else:
current.append(char)
if current:
elements.append(''.join(current).strip())
return elements
def parse_debug_info(debug_str):
result = {"API_Name": None, "Input": [], "Output": []}
result["API_Name"] = _parse_api_name(debug_str)
result["Input"] = _parse_input_tensors(debug_str)
result["Output"] = _parse_output_tensors(debug_str)
return result
class Edge:
def __init__(self, tensor_info: dict, source=""):
self.name = tensor_info["Name"]
self.shape = tensor_info["Shape"]
self.dtype = tensor_info["Dtype"]
self.source = source
def get_name(self) -> str:
return self.name
def get_source(self) -> str:
return self.source
def set_source(self, source):
self.source = source
def __str__(self) -> str:
return f"Edge(name='{self.name}', source='{self.source}',shape='{self.shape}',dtype='{self.dtype}')"
def get_edge_info(self):
return f"{self.name}\nshape:{self.shape}\ndtype:{self.dtype}"
class Graph:
def __init__(self):
from graphviz import Digraph
self.dot = Digraph()
self.orange_box_attrs = {
'style': 'rounded,filled,bold',
'shape': 'box',
'color': '#FFE4B5',
'fillcolor': '#FFE4B5',
'fontcolor': '#ffffff',
'width': '1.3',
'height': '0.84',
'fontname': 'Arial',
}
self.grey_box_attrs = {
'style': 'rounded,filled,bold',
'shape': 'box',
'color': '#999999',
'fillcolor': '#999999',
'fontcolor': '#ffffff',
'width': '1.3',
'height': '0.84',
'fontname': 'Arial',
}
self.edges = {}
self.nodes = []
def add_node(self, name: str, node_attr=""):
if not node_attr:
node_attr = self.grey_box_attrs
self.dot.node(name, name, **node_attr)
self.nodes.append(name)
# Link the src and dst node by edge
def add_edge(self, dst: str, edge: Edge):
edge_name = edge.get_name()
if edge_name not in self.edges:
# The Edge is not stored in the graph,
# so the Edge's source node maybe not in the graph
if not edge.get_source():
src_name = edge.get_name() + "_SourceNode"
edge.set_source(src_name)
self.add_node(src_name, self.orange_box_attrs)
else:
edge = self.edges[edge_name]
src = edge.get_source()
self.dot.edge(src, dst, label=edge.get_edge_info())
# Store an edge, not link the src and dst node
def store_edge(self, edge: Edge):
edge_name = edge.get_name()
self.edges[edge_name] = edge
def render(self, file_path):
self.dot.render(file_path, format='svg')
class GraphBuilder:
def __init__(self):
self.graph = Graph()
def build_graph(self, forward_debug_infos: list):
for info in forward_debug_infos:
debug_info = parse_debug_info(info)
api_name = debug_info['API_Name']
# Add a node for the API
self.graph.add_node(api_name)
# Store the Edge
for out_param in debug_info["Output"]:
var_name = out_param[0]
tensors = out_param[1]
for tensor_info in tensors:
# When we do not know the edge's dst, we should store it to the Graph
edge = Edge(tensor_info, api_name)
self.graph.store_edge(edge)
# Link the Edge
for input_param in debug_info["Input"]:
var_name = input_param[0]
tensors = input_param[1]
for tensor_info in tensors:
edge = Edge(tensor_info)
self.graph.add_edge(dst=api_name, edge=edge)
def save_graph(self, file_path):
self.graph.render(file_path)
@contextmanager
def capture_stderr():
with tempfile.TemporaryFile(mode='w+b') as temp_file:
original_stderr_fd = sys.stderr.fileno()
saved_stderr_fd = os.dup(original_stderr_fd)
stderr_output = io.StringIO()
try:
os.dup2(temp_file.fileno(), original_stderr_fd)
sys.stderr.flush()
yield stderr_output
finally:
sys.stderr.flush()
os.dup2(saved_stderr_fd, original_stderr_fd)
os.close(saved_stderr_fd)
temp_file.seek(0)
stderr_output.write(temp_file.read().decode())
@signature_safe_contextmanager
def capture_forward_subgraph_guard(file_path: str):
log = ""
stderr_buffer = io.StringIO()
origin_enable_unique_name_status = paddle.framework.get_flags(
"FLAGS_enable_unique_name"
)["FLAGS_enable_unique_name"]
try:
paddle.set_flags({"FLAGS_enable_unique_name": True})
# Redirect the stderr to the buffer,because the glog info will be printed to the stderr
with (
capture_stderr() as stderr_buffer,
paddle.base.framework.vlog_guard(3),
):
yield
finally:
paddle.set_flags(
{"FLAGS_enable_unique_name": origin_enable_unique_name_status}
)
log = stderr_buffer.getvalue()
builder = GraphBuilder()
def get_first_indent(s):
match = re.match(r'^\t*', s)
return match.group(0)
# Find the parts describing the input and output of the API in the massive logs
indent = get_first_indent(log.lstrip('\n'))
pattern_str = r'\n' + indent + r'Forward Debug Info \{.*? ] } '
pattern = re.compile(pattern_str, re.DOTALL)
matches = pattern.findall(log)
# Build the forward graph
builder.build_graph(matches)
builder.save_graph(file_path)