PP-schedule-visualizer / pipeline.py
Victarry's picture
Initial commit: 1F1B PP schedule visualization.
ec19476
raw
history blame
16.2 kB
import matplotlib.pyplot as plt
import numpy as np
import argparse
import json
import yaml
import os
from matplotlib.patches import Rectangle
from typing import List, Tuple, Dict, Literal
# Import visualization function from the new module
from visualizer import visualize_pipeline_parallelism
def create_1f1b_schedule(
num_stages: int,
num_batches: int,
forward_times: List[float],
backward_times: List[float],
p2p_time: float = 0.0,
) -> Dict[int, List[Dict]]:
"""
Create a 1F1B (One-Forward-One-Backward) schedule for pipeline parallelism.
This implementation takes a data-centric approach:
1. First determine the operation sequence for each pipeline stage (which microbatch to process when)
2. Then calculate timing based on dependencies between operations
The 1F1B pattern has three phases:
- Warmup: Forward passes for first num_stages microbatches
- Steady state: Alternating between forward and backward passes
- Cooldown: Backward passes for remaining microbatches
Returns:
A dictionary mapping device IDs to lists of tasks.
Each task is a dictionary with keys:
- 'type': 'forward' or 'backward'
- 'batch': batch number
- 'start_time': start time of the task
- 'duration': duration of the task
"""
# Initialize empty schedule
schedule = {stage: [] for stage in range(num_stages)}
# Step 1: Determine operation sequence for each stage
# This will generate the sequence of operations (forward/backward on which microbatch)
# that each stage should perform, without timing information yet
operation_sequence = determine_1f1b_operation_sequence(num_stages, num_batches)
# Step 2: Convert operation sequence to schedule with timing
# Taking into account dependencies between operations
schedule = calculate_operation_timing(
operation_sequence, num_stages, forward_times, backward_times, p2p_time
)
return schedule
def determine_1f1b_operation_sequence(
num_stages: int, num_batches: int
) -> Dict[int, List[Dict]]:
"""
Determine the sequence of operations (forward/backward) for each stage in 1F1B scheduling.
Args:
num_stages: Number of pipeline stages
num_batches: Number of micro-batches
Returns:
Dictionary mapping stage ID to a list of operations in sequence.
Each operation is a dict with keys 'type' ('forward' or 'backward') and 'batch'.
"""
operation_sequence = {i: [] for i in range(num_stages)}
for current_stage in range(num_stages):
warmup_batches = num_stages - current_stage
for j in range(1, warmup_batches + 1):
operation_sequence[current_stage].append({"type": "forward", "batch": j})
steady_batches = num_batches - warmup_batches
for j in range(warmup_batches + 1, warmup_batches + steady_batches + 1):
operation_sequence[current_stage].append(
{"type": "backward", "batch": j - warmup_batches}
)
operation_sequence[current_stage].append({"type": "forward", "batch": j})
for j in range(warmup_batches):
operation_sequence[current_stage].append(
{"type": "backward", "batch": j + steady_batches + 1}
)
return operation_sequence
def calculate_operation_timing(
operation_sequence: Dict[int, List[Dict]],
num_stages: int,
forward_times: List[float],
backward_times: List[float],
p2p_time: float = 0.0,
) -> Dict[int, List[Dict]]:
"""
Recursively calculate the specific timing of each operation in a 1F1B schedule.
When encountering an operation that depends on a previous operation that hasn't been calculated yet,
it will recursively calculate the timing of those operations.
Args:
operation_sequence: Operation sequence for each stage
num_stages: Number of pipeline stages
forward_times: Forward propagation time for each stage
backward_times: Backward propagation time for each stage
p2p_time: Point-to-point communication time between stages
Returns:
Complete schedule with timing information, each operation includes start_time and duration
"""
# Initialize schedule with timing information
schedule = {i: [] for i in range(num_stages)}
# For recording already computed operation end times
# Format: {(stage, batch, op_type): (start_time, end_time)}
computed_ops = {}
# For recording the end time of the last operation for each stage
stage_last_end_time = [0.0] * num_stages
# Helper function: recursively calculate the time for an operation
def compute_op_time(stage, batch, op_type):
# Check if this operation has already been calculated
key = (stage, batch, op_type)
if key in computed_ops:
return computed_ops[key]
# Get operation duration
duration = (
forward_times[stage] if op_type == "forward" else backward_times[stage]
)
# Determine start time (dependent on other operations)
# 1. Consider sequential dependencies on the stage (must wait for previous operation to complete)
start_time = stage_last_end_time[stage]
# 2. Forward pass also depends on forward pass of previous stage (if not the first stage)
if op_type == "forward" and stage > 0:
# Recursively calculate the time for the forward pass of the previous stage (if not calculated yet)
prev_stage_key = (stage - 1, batch, "forward")
if prev_stage_key not in computed_ops:
prev_start, prev_end = compute_op_time(stage - 1, batch, "forward")
else:
_, prev_end = computed_ops[prev_stage_key]
# Update start time
start_time = max(start_time, prev_end + p2p_time)
# 3. Backward pass depends on:
elif op_type == "backward":
# a. Forward pass of the same stage
same_stage_forward_key = (stage, batch, "forward")
if same_stage_forward_key not in computed_ops:
_, forward_end = compute_op_time(stage, batch, "forward")
else:
_, forward_end = computed_ops[same_stage_forward_key]
start_time = max(start_time, forward_end)
# b. Backward pass of the next stage (if not the last stage)
if stage < num_stages - 1:
next_stage_backward_key = (stage + 1, batch, "backward")
if next_stage_backward_key not in computed_ops:
_, next_backward_end = compute_op_time(stage + 1, batch, "backward")
else:
_, next_backward_end = computed_ops[next_stage_backward_key]
start_time = max(start_time, next_backward_end + p2p_time)
# Calculate end time
end_time = start_time + duration
# Store calculation results
computed_ops[key] = (start_time, end_time)
# Update the end time of the last operation for this stage
stage_last_end_time[stage] = end_time
return start_time, end_time
# Calculate time for each operation in the operation_sequence
for i in range(len(operation_sequence[0])):
for stage in range(num_stages):
batch = operation_sequence[stage][i]["batch"]
op_type = operation_sequence[stage][i]["type"]
# Recursively calculate the time for this operation
start_time, _ = compute_op_time(stage, batch, op_type)
# Fill in scheduling information
op_with_timing = operation_sequence[stage][i].copy()
op_with_timing["start_time"] = start_time
op_with_timing["duration"] = (
forward_times[stage] if op_type == "forward" else backward_times[stage]
)
schedule[stage].append(op_with_timing)
return schedule
def get_bubble_rate(schedule: Dict[int, List[Dict]]):
num_stages = len(schedule)
max_time = 0
for device in schedule:
for task in schedule[device]:
end_time = task["start_time"] + task["duration"]
if end_time > max_time:
max_time = end_time
total_execution_time = max_time * num_stages
total_computation_time = 0
device_computation_times = {}
for device in schedule:
device_computation_time = 0
for task in schedule[device]:
device_computation_time += task["duration"]
device_computation_times[device] = device_computation_time
total_computation_time += device_computation_time
bubble_rate = (
total_execution_time - total_computation_time
) / total_computation_time
return bubble_rate
def read_config_file(config_path):
"""
Read configuration from a JSON or YAML file.
Args:
config_path: Path to the config file (JSON or YAML)
Returns:
Dictionary containing configuration parameters
"""
if not os.path.exists(config_path):
raise FileNotFoundError(f"Config file not found: {config_path}")
file_ext = os.path.splitext(config_path)[1].lower()
try:
with open(config_path, "r") as f:
if file_ext == ".json":
config = json.load(f)
elif file_ext in (".yaml", ".yml"):
config = yaml.safe_load(f)
else:
raise ValueError(
f"Unsupported config file format: {file_ext}. Use .json, .yaml, or .yml"
)
return config
except Exception as e:
raise ValueError(f"Error reading config file: {str(e)}")
def parse_args():
"""
Parse command-line arguments for the pipeline parallelism tool.
Returns:
Parsed arguments namespace
"""
parser = argparse.ArgumentParser(
description="Pipeline Parallelism Scheduler and Visualizer",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
)
# Config file option
parser.add_argument(
"--config", "-c", type=str, help="Path to config file (JSON or YAML)"
)
# Main parameters
parser.add_argument(
"--num-stages",
"-s",
type=int,
default=4,
help="Number of pipeline stages (devices)",
)
parser.add_argument(
"--num-batches", "-b", type=int, default=10, help="Number of micro-batches"
)
# Forward and backward times
parser.add_argument(
"--forward-times",
"-f",
type=float,
nargs="+",
help="Time for forward pass at each stage (space-separated list)",
)
parser.add_argument(
"--backward-times",
"-bw",
type=float,
nargs="+",
help="Time for backward pass at each stage (space-separated list)",
)
# Output options
parser.add_argument(
"--output",
"-o",
type=str,
default="pipeline_1f1b.png",
help="Output file path for visualization",
)
parser.add_argument(
"--no-visualization", action="store_true", help="Skip visualization generation"
)
parser.add_argument(
"--p2p-time",
type=float,
default=0.0,
help="Time for point-to-point communication between stages",
)
return parser.parse_args()
def example_usage():
"""Example usage of the visualization function and testing the scheduling algorithms."""
# Example parameters
num_stages = 4 # Number of pipeline stages (devices)
num_batches = 10 # Number of micro-batches
# Example times for forward and backward passes for each stage
forward_times = [1.0, 1.0, 1.0, 1.0] # Time for forward pass at each stage
backward_times = [2.0, 2.0, 2.0, 2.0] # Time for backward pass at each stage
# Create 1F1B schedule
schedule = create_1f1b_schedule(
num_stages=num_stages,
num_batches=num_batches,
forward_times=forward_times,
backward_times=backward_times,
)
# Create visualization with the schedule
visualize_pipeline_parallelism(
schedule=schedule, schedule_type="1f1b", output_file="pipeline_1f1b.png"
)
# Analyze the schedule
bubble_rate = get_bubble_rate(schedule)
print(f"Bubble rate: {bubble_rate:.4f}")
def main():
"""
Main function that parses arguments and runs the pipeline parallelism analysis.
"""
args = parse_args()
# Initialize with default values
num_stages = 4
num_batches = 10
forward_times = None
backward_times = None
output_file = "pipeline_1f1b.png"
p2p_time = 0.0
# Read from config file if provided
if args.config:
try:
print(f"Reading configuration from {args.config}")
config = read_config_file(args.config)
# Update parameters from config
num_stages = config.get("num_stages", num_stages)
num_batches = config.get("num_batches", num_batches)
forward_times = config.get("forward_times")
backward_times = config.get("backward_times")
output_file = config.get("output_file", output_file)
p2p_time = config.get("p2p_time", 0.0)
except Exception as e:
print(f"Error reading config file: {str(e)}")
print("Falling back to command line arguments or defaults")
# Command line arguments override config file
if args.num_stages:
num_stages = args.num_stages
if args.num_batches:
num_batches = args.num_batches
if args.forward_times:
forward_times = args.forward_times
if args.backward_times:
backward_times = args.backward_times
if args.output:
output_file = args.output
if args.p2p_time:
p2p_time = args.p2p_time
# Validate inputs
if forward_times is None:
forward_times = [1.0] * num_stages
elif len(forward_times) != num_stages:
print(
f"Warning: forward_times length ({len(forward_times)}) doesn't match num_stages ({num_stages})"
)
if len(forward_times) < num_stages:
# Extend with repeats of the last value
forward_times = list(forward_times) + [forward_times[-1]] * (
num_stages - len(forward_times)
)
else:
# Truncate
forward_times = forward_times[:num_stages]
print(f"Adjusted forward_times: {forward_times}")
if backward_times is None:
backward_times = [2.0] * num_stages
elif len(backward_times) != num_stages:
print(
f"Warning: backward_times length ({len(backward_times)}) doesn't match num_stages ({num_stages})"
)
if len(backward_times) < num_stages:
# Extend with repeats of the last value
backward_times = list(backward_times) + [backward_times[-1]] * (
num_stages - len(backward_times)
)
else:
# Truncate
backward_times = backward_times[:num_stages]
print(f"Adjusted backward_times: {backward_times}")
print(f"Running with parameters:")
print(f" num_stages: {num_stages}")
print(f" num_batches: {num_batches}")
print(f" forward_times: {forward_times}")
print(f" backward_times: {backward_times}")
print(f" output_file: {output_file}")
# Create 1F1B schedule
schedule = create_1f1b_schedule(
num_stages=num_stages,
num_batches=num_batches,
forward_times=forward_times,
backward_times=backward_times,
p2p_time=p2p_time,
)
# Create visualization unless --no-visualization is specified
if not args.no_visualization:
visualize_pipeline_parallelism(
schedule=schedule, schedule_type="1f1b", output_file=output_file
)
# Analyze the schedule
bubble_rate = get_bubble_rate(schedule)
print(f"Bubble rate: {bubble_rate:.4f}")
return {
"schedule": schedule,
"bubble_rate": bubble_rate,
"num_stages": num_stages,
"num_batches": num_batches,
}
if __name__ == "__main__":
main()