Spaces:
Running
Running
File size: 14,884 Bytes
c048b97 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 |
import dash
import dash_bootstrap_components as dbc
from dash import dcc, html, Input, Output, State, callback_context
import plotly.graph_objects as go
from src.execution_model import ScheduleConfig, Schedule
from src.strategies import (
generate_1f1b_schedule,
generate_zero_bubble_1p_schedule,
generate_1f1b_overlap_schedule,
generate_1f1b_interleave_schedule,
generate_1f1b_interleave_overlap_schedule,
generate_dualpipe_schedule
)
from src.visualizer import convert_schedule_to_visualization_format, create_pipeline_figure
STRATEGIES = {
"1f1b": generate_1f1b_schedule,
"zb1p": generate_zero_bubble_1p_schedule,
"1f1b_overlap": generate_1f1b_overlap_schedule,
"1f1b_interleave": generate_1f1b_interleave_schedule,
"1f1b_interleave_overlap": generate_1f1b_interleave_overlap_schedule,
"dualpipe": generate_dualpipe_schedule,
}
app = dash.Dash(__name__, external_stylesheets=[dbc.themes.BOOTSTRAP], suppress_callback_exceptions=True)
app.title = "Pipeline Parallelism Schedule Visualizer"
# Initial default values
default_values = {
"num_devices": 4,
"num_stages": 8,
"num_batches": 16,
"p2p_latency": 0.0,
"op_time_forward": 1.0,
"op_time_backward_d": 1.0,
"op_time_backward_w": 1.0,
"op_time_backward": 2.0,
"strategy": "1f1b_interleave",
"op_time_overlapped_fwd_bwd": None,
}
# Define input groups using dbc components
basic_params_card = dbc.Card(
dbc.CardBody([
html.H5("Basic Parameters", className="card-title"),
html.Div([
dbc.Label("Number of Devices (GPUs):"),
dbc.Input(id='num_devices', type='number', value=default_values["num_devices"], min=1, step=1),
], className="mb-3"),
html.Div([
dbc.Label("Number of Stages (Model Chunks):"),
dbc.Input(id='num_stages', type='number', value=default_values["num_stages"], min=1, step=1),
], className="mb-3"),
html.Div([
dbc.Label("Number of Microbatches:"),
dbc.Input(id='num_batches', type='number', value=default_values["num_batches"], min=1, step=1),
], className="mb-3"),
html.Div([
dbc.Label("P2P Latency (ms):"),
dbc.Input(id='p2p_latency', type='number', value=default_values["p2p_latency"], min=0, step=0.01),
], className="mb-3"),
])
)
scheduling_params_card = dbc.Card(
dbc.CardBody([
html.H5("Scheduling Parameters", className="card-title"),
html.Div([
dbc.Label("Scheduling Strategies:"),
dbc.Checklist(
id='strategy-checklist',
options=[{'label': k, 'value': k} for k in STRATEGIES.keys()],
value=list(STRATEGIES.keys()),
inline=False,
),
], className="mb-3"),
])
)
timing_params_card = dbc.Card(
dbc.CardBody([
html.H5("Operation Timing (ms)", className="card-title"),
html.Div([
dbc.Label("Forward:"),
dbc.Input(id='op_time_forward', type='number', value=default_values["op_time_forward"], min=0.01, step=0.01),
], className="mb-3"),
html.Div([
dbc.Label("Backward (Combined):"),
dbc.Input(id='op_time_backward', type='number', value=default_values["op_time_backward"], min=0.01, step=0.01),
dbc.FormText("Used when strategy does NOT require split backward."),
], className="mb-3"),
html.Div([
dbc.Label("Backward D (Data Grad):"),
dbc.Input(id='op_time_backward_d', type='number', value=default_values["op_time_backward_d"], min=0.01, step=0.01),
dbc.FormText("Used when strategy requires split backward (e.g., ZB-1P, DualPipe)."),
], className="mb-3"),
html.Div([
dbc.Label("Backward W (Weight Grad):"),
dbc.Input(id='op_time_backward_w', type='number', value=default_values["op_time_backward_w"], min=0.01, step=0.01),
dbc.FormText("Used when strategy requires split backward (e.g., ZB-1P, DualPipe)."),
], className="mb-3"),
html.Div([
dbc.Label("Overlapped Forward+Backward:"),
dbc.Input(id='op_time_overlapped_fwd_bwd', type='number', placeholder="Optional: Defaults to Fwd + Bwd times", min=0.01, step=0.01, value=default_values["op_time_overlapped_fwd_bwd"]),
dbc.FormText("Specify a custom duration if Forward and Backward ops overlap completely."),
], className="mb-3"),
])
)
# Updated app layout using dbc components and structure
app.layout = dbc.Container([
html.H1("Pipeline Parallelism Schedule Visualizer", className="my-4 text-center"),
dbc.Row([
dbc.Col(basic_params_card, md=4),
dbc.Col(scheduling_params_card, md=4),
dbc.Col(timing_params_card, md=4),
]),
dbc.Row([
dbc.Col([
dbc.Button('Generate Schedule', id='generate-button', n_clicks=0, color="primary", className="mt-4"),
], className="text-center")
]),
dbc.Row([
dbc.Col([
dcc.Loading(
id="loading-graph-area",
type="circle",
children=html.Div(id='graph-output-container', className="mt-4")
)
])
])
], fluid=True)
@app.callback(
Output('graph-output-container', 'children'),
Input('generate-button', 'n_clicks'),
State('num_devices', 'value'),
State('num_stages', 'value'),
State('num_batches', 'value'),
State('p2p_latency', 'value'),
State('op_time_forward', 'value'),
State('op_time_backward', 'value'),
State('op_time_backward_d', 'value'),
State('op_time_backward_w', 'value'),
State('op_time_overlapped_fwd_bwd', 'value'),
State('strategy-checklist', 'value'),
prevent_initial_call=True
)
def update_graph(n_clicks, num_devices, num_stages, num_batches, p2p_latency,
op_time_forward, op_time_backward, op_time_backward_d, op_time_backward_w,
op_time_overlapped_fwd_bwd,
selected_strategies):
# Define the desired display order for strategies
strategy_display_order = ["1f1b", "1f1b_interleave", "1f1b_overlap", "1f1b_interleave_overlap", "dualpipe", "zb1p"]
output_components = []
valid_results = [] # Store (strategy_name, schedule, vis_data) for valid schedules
error_messages = [] # Store (strategy_name, error_message) for errors
automatic_adjustments = [] # Store messages about automatic parameter adjustments
if not selected_strategies:
return [dbc.Alert("Please select at least one scheduling strategy.", color="warning")]
if not all([num_devices, num_stages, num_batches, op_time_forward]):
return [dbc.Alert("Missing required basic input values (Devices, Stages, Batches, Forward Time).", color="danger")]
for strategy in selected_strategies:
error_message = ""
placement_strategy = ""
# Use local copies of params that might be adjusted for this strategy
current_num_stages = num_stages
current_num_devices = num_devices
# Apply automatic adjustments for dualpipe
if strategy == "dualpipe" and num_stages != num_devices:
current_num_stages = num_devices # Force num_stages = num_devices for dualpipe
automatic_adjustments.append(
f"Strategy '{strategy}': Number of Stages automatically adjusted to {num_devices} to match Number of Devices."
)
# Apply automatic adjustments for strategies that require num_stages == num_devices
if strategy in ["1f1b", "1f1b_overlap", "zb1p"] and num_stages != num_devices:
current_num_stages = num_devices
automatic_adjustments.append(
f"Strategy '{strategy}': Number of Stages automatically adjusted to {num_devices} to match Number of Devices."
)
split_backward = strategy in ["zb1p", "dualpipe"]
if split_backward and not all([op_time_backward_d, op_time_backward_w]):
error_message = f"Strategy '{strategy}': Backward D and Backward W times are required."
elif not split_backward and not op_time_backward:
error_message = f"Strategy '{strategy}': Combined Backward time is required."
if not error_message:
if strategy in ["1f1b", "1f1b_overlap", "zb1p"]:
placement_strategy = "standard"
# No need to check num_stages == num_devices as we've enforced it above
elif strategy in ["1f1b_interleave", "1f1b_interleave_overlap"]:
placement_strategy = "interleave"
if current_num_stages % current_num_devices != 0:
error_message = f"Strategy '{strategy}': Requires Number of Stages to be divisible by Number of Devices."
elif strategy == "dualpipe":
placement_strategy = "dualpipe"
if current_num_stages % 2 != 0:
error_message = f"Strategy '{strategy}' (DualPipe): Requires an even number of stages."
# Create adjusted operation times based on placement strategy
if not error_message:
try:
# Calculate number of stages per device for time adjustment
stages_per_device = current_num_stages // current_num_devices
# Calculate scaling factor - this normalizes operation time by stages per device
# For standard placement (1:1 stage:device mapping), this remains 1.0
# For interleaved, this scales down the time proportionally
time_scale_factor = 1.0 / stages_per_device if stages_per_device > 0 else 1.0
if stages_per_device > 1:
automatic_adjustments.append(
f"Strategy '{strategy}': Operation times scaled by 1/{stages_per_device} to account for {stages_per_device} stages per device."
)
# Apply scaling to operation times
op_times = {
"forward": float(op_time_forward) * time_scale_factor
}
if split_backward:
op_times["backward_D"] = float(op_time_backward_d) * time_scale_factor
op_times["backward_W"] = float(op_time_backward_w) * time_scale_factor
# Keep combined for compatibility
op_times["backward"] = (float(op_time_backward_d) + float(op_time_backward_w)) * time_scale_factor
else:
op_times["backward"] = float(op_time_backward) * time_scale_factor
if op_time_overlapped_fwd_bwd is not None:
try:
overlapped_val = float(op_time_overlapped_fwd_bwd)
if overlapped_val > 0:
# Scale overlapped time too
op_times["overlapped_forward_backward"] = overlapped_val * time_scale_factor
except (ValueError, TypeError):
pass
config = ScheduleConfig(
num_devices=int(current_num_devices),
num_stages=int(current_num_stages), # Use adjusted value
num_batches=int(num_batches),
p2p_latency=float(p2p_latency),
placement_strategy=placement_strategy,
split_backward=split_backward,
op_times=op_times,
)
schedule_func = STRATEGIES.get(strategy)
if not schedule_func:
raise ValueError(f"Invalid strategy function for: {strategy}")
schedule = schedule_func(config)
schedule.execute()
# Store valid results instead of creating figure immediately
vis_data = convert_schedule_to_visualization_format(schedule)
valid_results.append((strategy, schedule, vis_data))
except (AssertionError, ValueError, TypeError) as e:
error_message = f"Error generating schedule for '{strategy}': {e}"
import traceback
traceback.print_exc()
except Exception as e:
error_message = f"An unexpected error occurred for '{strategy}': {e}"
import traceback
traceback.print_exc()
if error_message:
error_messages.append((strategy, error_message))
# Add alerts for any automatic parameter adjustments
for adjustment in automatic_adjustments:
output_components.append(
dbc.Alert(adjustment, color="info", dismissable=True)
)
# If we have valid results, calculate the maximum execution time across all schedules
if valid_results:
# Find global maximum execution time
max_execution_time = max(schedule.get_total_execution_time() for _, schedule, _ in valid_results)
# Sort valid results according to the display order
sorted_valid_results = []
# First add strategies in the predefined order
for strategy_name in strategy_display_order:
for result in valid_results:
if result[0] == strategy_name:
sorted_valid_results.append(result)
# Then add any remaining strategies that might not be in the predefined order
for result in valid_results:
if result[0] not in strategy_display_order:
sorted_valid_results.append(result)
# Create figures with aligned x-axis, using the sorted results
for strategy, _, vis_data in sorted_valid_results:
fig = create_pipeline_figure(vis_data, max_time=max_execution_time, show_progress=False)
# Force the x-axis range to be the same for all figures
# Add a small margin (5%) for better visualization
margin = max_execution_time * 0.05
fig.update_layout(
xaxis=dict(
range=[0, max_execution_time + margin]
)
)
output_components.append(html.Div([
html.H4(f"Schedule: {strategy}", className="text-center mt-3 mb-2"),
dcc.Graph(figure=fig)
]))
# Add error messages to output
for strategy, msg in error_messages:
output_components.append(
dbc.Alert(msg, color="danger", className="mt-3")
)
return output_components
# For Hugging Face Spaces deployment
server = app.server
if __name__ == '__main__':
app.run_server(debug=False, host='0.0.0.0', port=7860) |