Update app.py
Browse files
app.py
CHANGED
|
@@ -15,42 +15,50 @@ from time_series_gradio import (
|
|
| 15 |
get_model_time_series_dfs,
|
| 16 |
)
|
| 17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
|
| 20 |
-
matplotlib.rcParams['
|
| 21 |
-
matplotlib.rcParams['
|
| 22 |
-
matplotlib.rcParams['
|
| 23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
|
| 26 |
-
# Load data once at startup
|
| 27 |
Ci_results = CIResults()
|
| 28 |
Ci_results.load_data()
|
| 29 |
-
|
| 30 |
if Ci_results.available_dates:
|
| 31 |
-
start_date_val = Ci_results.available_dates[-1]
|
| 32 |
-
end_date_val = Ci_results.available_dates[0]
|
| 33 |
Ci_results.load_historical_data(start_date_val, end_date_val)
|
| 34 |
logger.info(f"Preloaded historical data: {len(Ci_results.historical_df)} records")
|
| 35 |
-
# Start the auto-reload scheduler
|
| 36 |
Ci_results.schedule_data_reload()
|
| 37 |
|
| 38 |
|
| 39 |
-
# Function to check if a model has failures
|
| 40 |
def model_has_failures(model_name):
|
| 41 |
"""Check if a model has any failures (AMD or NVIDIA)."""
|
| 42 |
if Ci_results.df is None or Ci_results.df.empty:
|
| 43 |
return False
|
| 44 |
|
| 45 |
-
# Normalize model name to match DataFrame index
|
| 46 |
model_name_lower = model_name.lower()
|
| 47 |
|
| 48 |
-
# Check if model exists in DataFrame
|
| 49 |
if model_name_lower not in Ci_results.df.index:
|
| 50 |
return False
|
| 51 |
row = Ci_results.df.loc[model_name_lower]
|
| 52 |
|
| 53 |
-
# Check for failures in both AMD and NVIDIA
|
| 54 |
amd_multi_failures = row.get('failed_multi_no_amd', 0)
|
| 55 |
amd_single_failures = row.get('failed_single_no_amd', 0)
|
| 56 |
nvidia_multi_failures = row.get('failed_multi_no_nvidia', 0)
|
|
@@ -63,7 +71,6 @@ def model_has_failures(model_name):
|
|
| 63 |
])
|
| 64 |
|
| 65 |
|
| 66 |
-
# Function to get current description text
|
| 67 |
def get_description_text():
|
| 68 |
"""Get description text with integrated last update time."""
|
| 69 |
msg = [
|
|
@@ -79,7 +86,6 @@ def get_description_text():
|
|
| 79 |
msg.append("*(loading...)*")
|
| 80 |
return "<br>".join(msg)
|
| 81 |
|
| 82 |
-
# Load CSS from external file
|
| 83 |
def load_css():
|
| 84 |
try:
|
| 85 |
with open("styles.css", "r") as f:
|
|
@@ -88,7 +94,7 @@ def load_css():
|
|
| 88 |
return css_content
|
| 89 |
except FileNotFoundError:
|
| 90 |
logger.warning("styles.css not found, using minimal default styles")
|
| 91 |
-
return "body { background:
|
| 92 |
|
| 93 |
js_func = """
|
| 94 |
function refresh() {
|
|
@@ -101,20 +107,16 @@ function refresh() {
|
|
| 101 |
}
|
| 102 |
"""
|
| 103 |
|
| 104 |
-
# Create the Gradio interface with sidebar and dark theme
|
| 105 |
with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func) as demo:
|
| 106 |
|
| 107 |
|
| 108 |
with gr.Row():
|
| 109 |
-
# Sidebar for model selection
|
| 110 |
with gr.Column(scale=1, elem_classes=["sidebar"]):
|
| 111 |
gr.Markdown("# 🤖 TCID", elem_classes=["sidebar-title"])
|
| 112 |
|
| 113 |
-
# Description with integrated last update time
|
| 114 |
description_text = get_description_text()
|
| 115 |
description_display = gr.Markdown(description_text, elem_classes=["sidebar-description"])
|
| 116 |
|
| 117 |
-
# Summary button (for current view)
|
| 118 |
summary_button = gr.Button(
|
| 119 |
"summary\n📊",
|
| 120 |
variant="primary",
|
|
@@ -130,65 +132,58 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 130 |
)
|
| 131 |
|
| 132 |
|
| 133 |
-
# Model selection header (clickable toggle)
|
| 134 |
model_toggle_button = gr.Button(
|
| 135 |
f"► Select model ({len(Ci_results.available_models)})",
|
| 136 |
variant="secondary",
|
| 137 |
elem_classes=["model-header"]
|
| 138 |
)
|
| 139 |
|
| 140 |
-
# Model buttons container (collapsible) - start folded
|
| 141 |
with gr.Column(elem_classes=["model-list", "model-list-hidden"]) as model_list_container:
|
| 142 |
-
# Create individual buttons for each model
|
| 143 |
model_buttons = []
|
| 144 |
model_choices = [model.lower() for model in Ci_results.available_models] if Ci_results.available_models else ["auto", "bert", "clip", "llama"]
|
| 145 |
|
| 146 |
print(f"Creating {len(model_choices)} model buttons: {model_choices}")
|
| 147 |
|
| 148 |
for model_name in model_choices:
|
| 149 |
-
# Check if model has failures to determine styling
|
| 150 |
has_failures = model_has_failures(model_name)
|
| 151 |
button_classes = ["model-button"]
|
|
|
|
|
|
|
| 152 |
if has_failures:
|
| 153 |
button_classes.append("model-button-failed")
|
|
|
|
| 154 |
|
| 155 |
btn = gr.Button(
|
| 156 |
-
|
| 157 |
variant="secondary",
|
| 158 |
size="sm",
|
| 159 |
elem_classes=button_classes
|
| 160 |
)
|
| 161 |
model_buttons.append(btn)
|
| 162 |
|
| 163 |
-
# CI job links at bottom of sidebar
|
| 164 |
ci_links_display = gr.Markdown("🔗 **CI Jobs:** *Loading...*", elem_classes=["sidebar-links"])
|
| 165 |
|
| 166 |
-
# Main content area
|
| 167 |
with gr.Column(scale=4, elem_classes=["main-content"]):
|
| 168 |
-
# Current view components
|
| 169 |
with gr.Column(visible=True, elem_classes=["current-view"]) as current_view:
|
| 170 |
-
# Summary display (default view)
|
| 171 |
summary_display = gr.Plot(
|
| 172 |
-
value=create_summary_page(Ci_results.df, Ci_results.available_models),
|
| 173 |
label="",
|
| 174 |
format="png",
|
| 175 |
elem_classes=["plot-container"],
|
| 176 |
visible=True
|
| 177 |
)
|
| 178 |
|
| 179 |
-
# Detailed view components (hidden by default)
|
| 180 |
with gr.Column(visible=False, elem_classes=["detail-view"]) as detail_view:
|
| 181 |
-
# Create the plot output
|
| 182 |
plot_output = gr.Plot(
|
| 183 |
label="",
|
| 184 |
format="png",
|
| 185 |
elem_classes=["plot-container"]
|
| 186 |
)
|
| 187 |
|
| 188 |
-
# Create two separate failed tests displays in a row layout
|
| 189 |
with gr.Row():
|
| 190 |
with gr.Column(scale=1):
|
| 191 |
amd_failed_tests_output = gr.Textbox(
|
|
|
|
| 192 |
value="",
|
| 193 |
lines=8,
|
| 194 |
max_lines=8,
|
|
@@ -198,6 +193,7 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 198 |
)
|
| 199 |
with gr.Column(scale=1):
|
| 200 |
nvidia_failed_tests_output = gr.Textbox(
|
|
|
|
| 201 |
value="",
|
| 202 |
lines=8,
|
| 203 |
max_lines=8,
|
|
@@ -206,17 +202,15 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 206 |
elem_classes=["failed-tests"]
|
| 207 |
)
|
| 208 |
|
| 209 |
-
# Historical view components (hidden by default)
|
| 210 |
with gr.Column(visible=False, elem_classes=["historical-view"]) as historical_view:
|
| 211 |
|
| 212 |
|
| 213 |
-
# Time-series summary displays (multiple Gradio plots)
|
| 214 |
time_series_failure_rates = gr.LinePlot(
|
| 215 |
label="",
|
| 216 |
x="date",
|
| 217 |
y="failure_rate",
|
| 218 |
color="platform",
|
| 219 |
-
color_map={"AMD":
|
| 220 |
title="Overall Failure Rates Over Time",
|
| 221 |
tooltip=["failure_rate", "date", "change"],
|
| 222 |
height=300,
|
|
@@ -230,7 +224,7 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 230 |
x="date",
|
| 231 |
y="count",
|
| 232 |
color="test_type",
|
| 233 |
-
color_map={"Passed":
|
| 234 |
title="AMD Test Results Over Time",
|
| 235 |
tooltip=["count", "date", "change"],
|
| 236 |
height=300,
|
|
@@ -244,7 +238,7 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 244 |
x="date",
|
| 245 |
y="count",
|
| 246 |
color="test_type",
|
| 247 |
-
color_map={"Passed":
|
| 248 |
title="NVIDIA Test Results Over Time",
|
| 249 |
tooltip=["count", "date", "change"],
|
| 250 |
height=300,
|
|
@@ -253,15 +247,13 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 253 |
elem_classes=["plot-container"]
|
| 254 |
)
|
| 255 |
|
| 256 |
-
# Time-series model view (hidden by default)
|
| 257 |
with gr.Column(visible=False, elem_classes=["time-series-detail-view"]) as time_series_detail_view:
|
| 258 |
-
# Time-series plots for specific model (with spacing)
|
| 259 |
time_series_amd_model_plot = gr.LinePlot(
|
| 260 |
label="",
|
| 261 |
x="date",
|
| 262 |
y="count",
|
| 263 |
color="test_type",
|
| 264 |
-
color_map={"Passed":
|
| 265 |
title="AMD Results Over Time",
|
| 266 |
tooltip=["count", "date", "change"],
|
| 267 |
height=300,
|
|
@@ -275,7 +267,7 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 275 |
x="date",
|
| 276 |
y="count",
|
| 277 |
color="test_type",
|
| 278 |
-
color_map={"Passed":
|
| 279 |
title="NVIDIA Results Over Time",
|
| 280 |
tooltip=["count", "date", "change"],
|
| 281 |
height=300,
|
|
@@ -284,14 +276,12 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 284 |
elem_classes=["plot-container"]
|
| 285 |
)
|
| 286 |
|
| 287 |
-
# Model toggle functionality
|
| 288 |
def toggle_model_list(current_visible):
|
| 289 |
"""Toggle the visibility of the model list."""
|
| 290 |
new_visible = not current_visible
|
| 291 |
arrow = "▼" if new_visible else "►"
|
| 292 |
button_text = f"{arrow} Select model ({len(Ci_results.available_models)})"
|
| 293 |
|
| 294 |
-
# Use CSS classes instead of Gradio visibility
|
| 295 |
css_classes = ["model-list"]
|
| 296 |
if new_visible:
|
| 297 |
css_classes.append("model-list-visible")
|
|
@@ -300,11 +290,8 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 300 |
|
| 301 |
return gr.update(value=button_text), gr.update(elem_classes=css_classes), new_visible
|
| 302 |
|
| 303 |
-
# Track model list visibility state
|
| 304 |
model_list_visible = gr.State(False)
|
| 305 |
-
# Track last selected model for mode switches
|
| 306 |
selected_model_state = gr.State(None)
|
| 307 |
-
# Track whether current view is model detail (True) or summary (False)
|
| 308 |
in_model_view_state = gr.State(False)
|
| 309 |
|
| 310 |
model_toggle_button.click(
|
|
@@ -314,7 +301,6 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 314 |
)
|
| 315 |
|
| 316 |
|
| 317 |
-
# Unified summary handler: respects History toggle
|
| 318 |
def handle_summary_click(history_mode: bool):
|
| 319 |
description = get_description_text()
|
| 320 |
links = get_ci_links()
|
|
@@ -323,30 +309,30 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 323 |
return (
|
| 324 |
description,
|
| 325 |
links,
|
| 326 |
-
gr.update(visible=False),
|
| 327 |
-
gr.update(visible=True),
|
| 328 |
-
gr.update(visible=False),
|
| 329 |
-
gr.update(visible=False),
|
| 330 |
fr_plot,
|
| 331 |
amd_plot,
|
| 332 |
nvidia_plot,
|
| 333 |
-
gr.update(visible=False),
|
| 334 |
-
False,
|
| 335 |
)
|
| 336 |
else:
|
| 337 |
-
fig = create_summary_page(Ci_results.df, Ci_results.available_models)
|
| 338 |
return (
|
| 339 |
description,
|
| 340 |
links,
|
| 341 |
-
gr.update(visible=True),
|
| 342 |
-
gr.update(visible=False),
|
| 343 |
-
gr.update(value=fig, visible=True),
|
| 344 |
-
gr.update(visible=False),
|
| 345 |
-
gr.update(visible=False),
|
| 346 |
-
gr.update(visible=False),
|
| 347 |
-
gr.update(visible=False),
|
| 348 |
-
gr.update(visible=False),
|
| 349 |
-
False,
|
| 350 |
)
|
| 351 |
|
| 352 |
summary_button.click(
|
|
@@ -367,15 +353,12 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 367 |
],
|
| 368 |
)
|
| 369 |
|
| 370 |
-
# Function to get CI job links
|
| 371 |
def get_ci_links():
|
| 372 |
"""Get CI job links from the most recent data."""
|
| 373 |
try:
|
| 374 |
-
# Check if df exists and is not empty
|
| 375 |
if Ci_results.df is None or Ci_results.df.empty:
|
| 376 |
return "🔗 **CI Jobs:** *Loading...*"
|
| 377 |
|
| 378 |
-
# Get links from any available model (they should be the same for all models in a run)
|
| 379 |
amd_multi_link = None
|
| 380 |
amd_single_link = None
|
| 381 |
nvidia_multi_link = None
|
|
@@ -384,7 +367,6 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 384 |
for model_name in Ci_results.df.index:
|
| 385 |
row = Ci_results.df.loc[model_name]
|
| 386 |
|
| 387 |
-
# Extract AMD links
|
| 388 |
if pd.notna(row.get('job_link_amd')) and (not amd_multi_link or not amd_single_link):
|
| 389 |
amd_link_raw = row.get('job_link_amd')
|
| 390 |
if isinstance(amd_link_raw, dict):
|
|
@@ -393,7 +375,6 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 393 |
if 'single' in amd_link_raw and not amd_single_link:
|
| 394 |
amd_single_link = amd_link_raw['single']
|
| 395 |
|
| 396 |
-
# Extract NVIDIA links
|
| 397 |
if pd.notna(row.get('job_link_nvidia')) and (not nvidia_multi_link or not nvidia_single_link):
|
| 398 |
nvidia_link_raw = row.get('job_link_nvidia')
|
| 399 |
if isinstance(nvidia_link_raw, dict):
|
|
@@ -402,16 +383,13 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 402 |
if 'single' in nvidia_link_raw and not nvidia_single_link:
|
| 403 |
nvidia_single_link = nvidia_link_raw['single']
|
| 404 |
|
| 405 |
-
# Break if we have all links
|
| 406 |
if amd_multi_link and amd_single_link and nvidia_multi_link and nvidia_single_link:
|
| 407 |
break
|
| 408 |
|
| 409 |
|
| 410 |
-
# Add FAQ link at the bottom
|
| 411 |
links_md = "❓ [**FAQ**](https://huggingface.co/spaces/transformers-community/transformers-ci-dashboard/blob/main/README.md)\n\n"
|
| 412 |
links_md += "🔗 **CI Jobs:**\n\n"
|
| 413 |
|
| 414 |
-
# AMD links
|
| 415 |
if amd_multi_link or amd_single_link:
|
| 416 |
links_md += "**AMD:**\n"
|
| 417 |
if amd_multi_link:
|
|
@@ -420,7 +398,6 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 420 |
links_md += f"• [Single GPU]({amd_single_link})\n"
|
| 421 |
links_md += "\n"
|
| 422 |
|
| 423 |
-
# NVIDIA links
|
| 424 |
if nvidia_multi_link or nvidia_single_link:
|
| 425 |
links_md += "**NVIDIA:**\n"
|
| 426 |
if nvidia_multi_link:
|
|
@@ -437,7 +414,7 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 437 |
return "🔗 **CI Jobs:** *Error loading links*\n\n❓ **[FAQ](README.md)**"
|
| 438 |
|
| 439 |
|
| 440 |
-
|
| 441 |
|
| 442 |
def get_historical_summary_plots():
|
| 443 |
"""Get historical summary plots from preloaded data."""
|
|
@@ -450,80 +427,77 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 450 |
|
| 451 |
def handle_history_toggle(history_mode, last_selected_model, in_model_view):
|
| 452 |
if history_mode:
|
| 453 |
-
# If currently in model view and valid model, show historical model detail
|
| 454 |
if in_model_view and last_selected_model:
|
| 455 |
amd_ts, nvidia_ts = show_time_series_model(last_selected_model)
|
| 456 |
return (
|
| 457 |
-
gr.update(visible=False),
|
| 458 |
-
gr.update(visible=True),
|
| 459 |
-
gr.update(visible=False),
|
| 460 |
-
gr.update(visible=False),
|
| 461 |
-
gr.update(visible=False),
|
| 462 |
-
gr.update(visible=False),
|
| 463 |
-
gr.update(visible=False),
|
| 464 |
-
amd_ts,
|
| 465 |
-
nvidia_ts,
|
| 466 |
-
gr.update(visible=True),
|
| 467 |
-
gr.update(),
|
| 468 |
-
gr.update(),
|
| 469 |
-
gr.update(),
|
| 470 |
-
True,
|
| 471 |
)
|
| 472 |
-
# Otherwise show historical summary
|
| 473 |
fr_plot, amd_plot, nvidia_plot = get_historical_summary_plots()
|
| 474 |
return (
|
| 475 |
-
gr.update(visible=False),
|
| 476 |
-
gr.update(visible=True),
|
| 477 |
-
gr.update(visible=False),
|
| 478 |
-
gr.update(visible=False),
|
| 479 |
-
fr_plot,
|
| 480 |
-
amd_plot,
|
| 481 |
-
nvidia_plot,
|
| 482 |
-
gr.update(),
|
| 483 |
-
gr.update(),
|
| 484 |
-
gr.update(visible=False),
|
| 485 |
-
gr.update(),
|
| 486 |
-
gr.update(),
|
| 487 |
-
gr.update(),
|
| 488 |
-
False,
|
| 489 |
)
|
| 490 |
else:
|
| 491 |
-
# Switch to current mode: show model if selected; otherwise summary
|
| 492 |
if last_selected_model and Ci_results.df is not None and not Ci_results.df.empty and last_selected_model in Ci_results.df.index:
|
| 493 |
-
fig, amd_txt, nvidia_txt = plot_model_stats(Ci_results.df, last_selected_model)
|
| 494 |
return (
|
| 495 |
-
gr.update(visible=True),
|
| 496 |
-
gr.update(visible=False),
|
| 497 |
-
gr.update(visible=False),
|
| 498 |
-
gr.update(visible=True),
|
| 499 |
-
gr.update(visible=False),
|
| 500 |
-
gr.update(visible=False),
|
| 501 |
-
gr.update(visible=False),
|
| 502 |
-
gr.update(),
|
| 503 |
-
gr.update(),
|
| 504 |
-
gr.update(visible=False),
|
| 505 |
-
fig,
|
| 506 |
-
amd_txt,
|
| 507 |
-
nvidia_txt,
|
| 508 |
-
True,
|
| 509 |
)
|
| 510 |
else:
|
| 511 |
-
fig = create_summary_page(Ci_results.df, Ci_results.available_models)
|
| 512 |
return (
|
| 513 |
-
gr.update(visible=True),
|
| 514 |
-
gr.update(visible=False),
|
| 515 |
-
gr.update(value=fig, visible=True),
|
| 516 |
-
gr.update(visible=False),
|
| 517 |
-
gr.update(visible=False),
|
| 518 |
-
gr.update(visible=False),
|
| 519 |
-
gr.update(visible=False),
|
| 520 |
-
gr.update(),
|
| 521 |
-
gr.update(),
|
| 522 |
-
gr.update(visible=False),
|
| 523 |
-
gr.update(),
|
| 524 |
-
gr.update(),
|
| 525 |
-
gr.update(),
|
| 526 |
-
False,
|
| 527 |
)
|
| 528 |
|
| 529 |
history_view_button.change(
|
|
@@ -548,7 +522,6 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 548 |
)
|
| 549 |
|
| 550 |
|
| 551 |
-
# Time-series model selection functionality
|
| 552 |
def show_time_series_model(selected_model):
|
| 553 |
"""Show time-series view for a specific model."""
|
| 554 |
dfs = get_model_time_series_dfs(Ci_results.historical_df, selected_model)
|
|
@@ -557,42 +530,41 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 557 |
gr.update(value=dfs['nvidia_df'], visible=True, title=f"{selected_model.upper()} - NVIDIA Results Over Time"),
|
| 558 |
)
|
| 559 |
|
| 560 |
-
# Unified model click handler: respects History toggle
|
| 561 |
def handle_model_click(selected_model: str, history_mode: bool):
|
| 562 |
if history_mode:
|
| 563 |
amd_ts, nvidia_ts = show_time_series_model(selected_model)
|
| 564 |
return (
|
| 565 |
-
gr.update(),
|
| 566 |
-
gr.update(),
|
| 567 |
-
gr.update(),
|
| 568 |
-
gr.update(visible=False),
|
| 569 |
-
gr.update(visible=True),
|
| 570 |
-
gr.update(visible=False),
|
| 571 |
-
gr.update(visible=False),
|
| 572 |
-
gr.update(visible=False),
|
| 573 |
-
gr.update(visible=False),
|
| 574 |
-
gr.update(visible=False),
|
| 575 |
-
amd_ts,
|
| 576 |
-
nvidia_ts,
|
| 577 |
-
gr.update(visible=True),
|
| 578 |
-
selected_model, True)
|
| 579 |
else:
|
| 580 |
-
fig, amd_txt, nvidia_txt = plot_model_stats(Ci_results.df, selected_model)
|
| 581 |
return (
|
| 582 |
fig,
|
| 583 |
amd_txt,
|
| 584 |
nvidia_txt,
|
| 585 |
-
gr.update(visible=True),
|
| 586 |
-
gr.update(visible=False),
|
| 587 |
-
gr.update(visible=False),
|
| 588 |
-
gr.update(visible=True),
|
| 589 |
-
gr.update(),
|
| 590 |
-
gr.update(),
|
| 591 |
-
gr.update(),
|
| 592 |
-
gr.update(),
|
| 593 |
-
gr.update(),
|
| 594 |
-
gr.update(visible=False),
|
| 595 |
-
selected_model, True)
|
| 596 |
|
| 597 |
for i, btn in enumerate(model_buttons):
|
| 598 |
model_name = model_choices[i]
|
|
@@ -618,13 +590,11 @@ with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func)
|
|
| 618 |
],
|
| 619 |
)
|
| 620 |
|
| 621 |
-
# Auto-update CI links when the interface loads
|
| 622 |
demo.load(
|
| 623 |
fn=get_ci_links,
|
| 624 |
outputs=[ci_links_display]
|
| 625 |
)
|
| 626 |
|
| 627 |
|
| 628 |
-
# Gradio entrypoint
|
| 629 |
if __name__ == "__main__":
|
| 630 |
-
demo.launch()
|
|
|
|
| 15 |
get_model_time_series_dfs,
|
| 16 |
)
|
| 17 |
|
| 18 |
+
COLORS = {
|
| 19 |
+
'passed': '#4CAF50',
|
| 20 |
+
'failed': '#E53E3E',
|
| 21 |
+
'skipped': '#FFD54F',
|
| 22 |
+
'error': '#8B0000',
|
| 23 |
+
'amd': '#ED1C24',
|
| 24 |
+
'nvidia': '#76B900',
|
| 25 |
+
'background': '#0B0F19',
|
| 26 |
+
'text': '#FFFFFF'
|
| 27 |
+
}
|
| 28 |
|
| 29 |
+
matplotlib.rcParams['figure.facecolor'] = COLORS['background']
|
| 30 |
+
matplotlib.rcParams['axes.facecolor'] = COLORS['background']
|
| 31 |
+
matplotlib.rcParams['savefig.facecolor'] = COLORS['background']
|
| 32 |
+
matplotlib.rcParams['text.color'] = COLORS['text']
|
| 33 |
+
matplotlib.rcParams['axes.labelcolor'] = COLORS['text']
|
| 34 |
+
matplotlib.rcParams['xtick.color'] = COLORS['text']
|
| 35 |
+
matplotlib.rcParams['ytick.color'] = COLORS['text']
|
| 36 |
+
matplotlib.rcParams['axes.edgecolor'] = COLORS['text']
|
| 37 |
+
plt.ioff()
|
| 38 |
|
| 39 |
|
|
|
|
| 40 |
Ci_results = CIResults()
|
| 41 |
Ci_results.load_data()
|
| 42 |
+
|
| 43 |
if Ci_results.available_dates:
|
| 44 |
+
start_date_val = Ci_results.available_dates[-1]
|
| 45 |
+
end_date_val = Ci_results.available_dates[0]
|
| 46 |
Ci_results.load_historical_data(start_date_val, end_date_val)
|
| 47 |
logger.info(f"Preloaded historical data: {len(Ci_results.historical_df)} records")
|
|
|
|
| 48 |
Ci_results.schedule_data_reload()
|
| 49 |
|
| 50 |
|
|
|
|
| 51 |
def model_has_failures(model_name):
|
| 52 |
"""Check if a model has any failures (AMD or NVIDIA)."""
|
| 53 |
if Ci_results.df is None or Ci_results.df.empty:
|
| 54 |
return False
|
| 55 |
|
|
|
|
| 56 |
model_name_lower = model_name.lower()
|
| 57 |
|
|
|
|
| 58 |
if model_name_lower not in Ci_results.df.index:
|
| 59 |
return False
|
| 60 |
row = Ci_results.df.loc[model_name_lower]
|
| 61 |
|
|
|
|
| 62 |
amd_multi_failures = row.get('failed_multi_no_amd', 0)
|
| 63 |
amd_single_failures = row.get('failed_single_no_amd', 0)
|
| 64 |
nvidia_multi_failures = row.get('failed_multi_no_nvidia', 0)
|
|
|
|
| 71 |
])
|
| 72 |
|
| 73 |
|
|
|
|
| 74 |
def get_description_text():
|
| 75 |
"""Get description text with integrated last update time."""
|
| 76 |
msg = [
|
|
|
|
| 86 |
msg.append("*(loading...)*")
|
| 87 |
return "<br>".join(msg)
|
| 88 |
|
|
|
|
| 89 |
def load_css():
|
| 90 |
try:
|
| 91 |
with open("styles.css", "r") as f:
|
|
|
|
| 94 |
return css_content
|
| 95 |
except FileNotFoundError:
|
| 96 |
logger.warning("styles.css not found, using minimal default styles")
|
| 97 |
+
return f"body {{ background: {COLORS['background']}; color: {COLORS['text']}; }}"
|
| 98 |
|
| 99 |
js_func = """
|
| 100 |
function refresh() {
|
|
|
|
| 107 |
}
|
| 108 |
"""
|
| 109 |
|
|
|
|
| 110 |
with gr.Blocks(title="Model Test Results Dashboard", css=load_css(), js=js_func) as demo:
|
| 111 |
|
| 112 |
|
| 113 |
with gr.Row():
|
|
|
|
| 114 |
with gr.Column(scale=1, elem_classes=["sidebar"]):
|
| 115 |
gr.Markdown("# 🤖 TCID", elem_classes=["sidebar-title"])
|
| 116 |
|
|
|
|
| 117 |
description_text = get_description_text()
|
| 118 |
description_display = gr.Markdown(description_text, elem_classes=["sidebar-description"])
|
| 119 |
|
|
|
|
| 120 |
summary_button = gr.Button(
|
| 121 |
"summary\n📊",
|
| 122 |
variant="primary",
|
|
|
|
| 132 |
)
|
| 133 |
|
| 134 |
|
|
|
|
| 135 |
model_toggle_button = gr.Button(
|
| 136 |
f"► Select model ({len(Ci_results.available_models)})",
|
| 137 |
variant="secondary",
|
| 138 |
elem_classes=["model-header"]
|
| 139 |
)
|
| 140 |
|
|
|
|
| 141 |
with gr.Column(elem_classes=["model-list", "model-list-hidden"]) as model_list_container:
|
|
|
|
| 142 |
model_buttons = []
|
| 143 |
model_choices = [model.lower() for model in Ci_results.available_models] if Ci_results.available_models else ["auto", "bert", "clip", "llama"]
|
| 144 |
|
| 145 |
print(f"Creating {len(model_choices)} model buttons: {model_choices}")
|
| 146 |
|
| 147 |
for model_name in model_choices:
|
|
|
|
| 148 |
has_failures = model_has_failures(model_name)
|
| 149 |
button_classes = ["model-button"]
|
| 150 |
+
display_name = model_name
|
| 151 |
+
|
| 152 |
if has_failures:
|
| 153 |
button_classes.append("model-button-failed")
|
| 154 |
+
display_name = f"🚨 {model_name}"
|
| 155 |
|
| 156 |
btn = gr.Button(
|
| 157 |
+
display_name,
|
| 158 |
variant="secondary",
|
| 159 |
size="sm",
|
| 160 |
elem_classes=button_classes
|
| 161 |
)
|
| 162 |
model_buttons.append(btn)
|
| 163 |
|
|
|
|
| 164 |
ci_links_display = gr.Markdown("🔗 **CI Jobs:** *Loading...*", elem_classes=["sidebar-links"])
|
| 165 |
|
|
|
|
| 166 |
with gr.Column(scale=4, elem_classes=["main-content"]):
|
|
|
|
| 167 |
with gr.Column(visible=True, elem_classes=["current-view"]) as current_view:
|
|
|
|
| 168 |
summary_display = gr.Plot(
|
| 169 |
+
value=create_summary_page(Ci_results.df, Ci_results.available_models, COLORS=COLORS),
|
| 170 |
label="",
|
| 171 |
format="png",
|
| 172 |
elem_classes=["plot-container"],
|
| 173 |
visible=True
|
| 174 |
)
|
| 175 |
|
|
|
|
| 176 |
with gr.Column(visible=False, elem_classes=["detail-view"]) as detail_view:
|
|
|
|
| 177 |
plot_output = gr.Plot(
|
| 178 |
label="",
|
| 179 |
format="png",
|
| 180 |
elem_classes=["plot-container"]
|
| 181 |
)
|
| 182 |
|
|
|
|
| 183 |
with gr.Row():
|
| 184 |
with gr.Column(scale=1):
|
| 185 |
amd_failed_tests_output = gr.Textbox(
|
| 186 |
+
label="AMD Failed Tests",
|
| 187 |
value="",
|
| 188 |
lines=8,
|
| 189 |
max_lines=8,
|
|
|
|
| 193 |
)
|
| 194 |
with gr.Column(scale=1):
|
| 195 |
nvidia_failed_tests_output = gr.Textbox(
|
| 196 |
+
label="NVIDIA Failed Tests",
|
| 197 |
value="",
|
| 198 |
lines=8,
|
| 199 |
max_lines=8,
|
|
|
|
| 202 |
elem_classes=["failed-tests"]
|
| 203 |
)
|
| 204 |
|
|
|
|
| 205 |
with gr.Column(visible=False, elem_classes=["historical-view"]) as historical_view:
|
| 206 |
|
| 207 |
|
|
|
|
| 208 |
time_series_failure_rates = gr.LinePlot(
|
| 209 |
label="",
|
| 210 |
x="date",
|
| 211 |
y="failure_rate",
|
| 212 |
color="platform",
|
| 213 |
+
color_map={"AMD": COLORS['amd'], "NVIDIA": COLORS['nvidia']},
|
| 214 |
title="Overall Failure Rates Over Time",
|
| 215 |
tooltip=["failure_rate", "date", "change"],
|
| 216 |
height=300,
|
|
|
|
| 224 |
x="date",
|
| 225 |
y="count",
|
| 226 |
color="test_type",
|
| 227 |
+
color_map={"Passed": COLORS['passed'], "Failed": COLORS['failed'], "Skipped": COLORS['skipped']},
|
| 228 |
title="AMD Test Results Over Time",
|
| 229 |
tooltip=["count", "date", "change"],
|
| 230 |
height=300,
|
|
|
|
| 238 |
x="date",
|
| 239 |
y="count",
|
| 240 |
color="test_type",
|
| 241 |
+
color_map={"Passed": COLORS['passed'], "Failed": COLORS['failed'], "Skipped": COLORS['skipped']},
|
| 242 |
title="NVIDIA Test Results Over Time",
|
| 243 |
tooltip=["count", "date", "change"],
|
| 244 |
height=300,
|
|
|
|
| 247 |
elem_classes=["plot-container"]
|
| 248 |
)
|
| 249 |
|
|
|
|
| 250 |
with gr.Column(visible=False, elem_classes=["time-series-detail-view"]) as time_series_detail_view:
|
|
|
|
| 251 |
time_series_amd_model_plot = gr.LinePlot(
|
| 252 |
label="",
|
| 253 |
x="date",
|
| 254 |
y="count",
|
| 255 |
color="test_type",
|
| 256 |
+
color_map={"Passed": COLORS['passed'], "Failed": COLORS['failed'], "Skipped": COLORS['skipped']},
|
| 257 |
title="AMD Results Over Time",
|
| 258 |
tooltip=["count", "date", "change"],
|
| 259 |
height=300,
|
|
|
|
| 267 |
x="date",
|
| 268 |
y="count",
|
| 269 |
color="test_type",
|
| 270 |
+
color_map={"Passed": COLORS['passed'], "Failed": COLORS['failed'], "Skipped": COLORS['skipped']},
|
| 271 |
title="NVIDIA Results Over Time",
|
| 272 |
tooltip=["count", "date", "change"],
|
| 273 |
height=300,
|
|
|
|
| 276 |
elem_classes=["plot-container"]
|
| 277 |
)
|
| 278 |
|
|
|
|
| 279 |
def toggle_model_list(current_visible):
|
| 280 |
"""Toggle the visibility of the model list."""
|
| 281 |
new_visible = not current_visible
|
| 282 |
arrow = "▼" if new_visible else "►"
|
| 283 |
button_text = f"{arrow} Select model ({len(Ci_results.available_models)})"
|
| 284 |
|
|
|
|
| 285 |
css_classes = ["model-list"]
|
| 286 |
if new_visible:
|
| 287 |
css_classes.append("model-list-visible")
|
|
|
|
| 290 |
|
| 291 |
return gr.update(value=button_text), gr.update(elem_classes=css_classes), new_visible
|
| 292 |
|
|
|
|
| 293 |
model_list_visible = gr.State(False)
|
|
|
|
| 294 |
selected_model_state = gr.State(None)
|
|
|
|
| 295 |
in_model_view_state = gr.State(False)
|
| 296 |
|
| 297 |
model_toggle_button.click(
|
|
|
|
| 301 |
)
|
| 302 |
|
| 303 |
|
|
|
|
| 304 |
def handle_summary_click(history_mode: bool):
|
| 305 |
description = get_description_text()
|
| 306 |
links = get_ci_links()
|
|
|
|
| 309 |
return (
|
| 310 |
description,
|
| 311 |
links,
|
| 312 |
+
gr.update(visible=False),
|
| 313 |
+
gr.update(visible=True),
|
| 314 |
+
gr.update(visible=False),
|
| 315 |
+
gr.update(visible=False),
|
| 316 |
fr_plot,
|
| 317 |
amd_plot,
|
| 318 |
nvidia_plot,
|
| 319 |
+
gr.update(visible=False),
|
| 320 |
+
False,
|
| 321 |
)
|
| 322 |
else:
|
| 323 |
+
fig = create_summary_page(Ci_results.df, Ci_results.available_models, COLORS=COLORS)
|
| 324 |
return (
|
| 325 |
description,
|
| 326 |
links,
|
| 327 |
+
gr.update(visible=True),
|
| 328 |
+
gr.update(visible=False),
|
| 329 |
+
gr.update(value=fig, visible=True),
|
| 330 |
+
gr.update(visible=False),
|
| 331 |
+
gr.update(visible=False),
|
| 332 |
+
gr.update(visible=False),
|
| 333 |
+
gr.update(visible=False),
|
| 334 |
+
gr.update(visible=False),
|
| 335 |
+
False,
|
| 336 |
)
|
| 337 |
|
| 338 |
summary_button.click(
|
|
|
|
| 353 |
],
|
| 354 |
)
|
| 355 |
|
|
|
|
| 356 |
def get_ci_links():
|
| 357 |
"""Get CI job links from the most recent data."""
|
| 358 |
try:
|
|
|
|
| 359 |
if Ci_results.df is None or Ci_results.df.empty:
|
| 360 |
return "🔗 **CI Jobs:** *Loading...*"
|
| 361 |
|
|
|
|
| 362 |
amd_multi_link = None
|
| 363 |
amd_single_link = None
|
| 364 |
nvidia_multi_link = None
|
|
|
|
| 367 |
for model_name in Ci_results.df.index:
|
| 368 |
row = Ci_results.df.loc[model_name]
|
| 369 |
|
|
|
|
| 370 |
if pd.notna(row.get('job_link_amd')) and (not amd_multi_link or not amd_single_link):
|
| 371 |
amd_link_raw = row.get('job_link_amd')
|
| 372 |
if isinstance(amd_link_raw, dict):
|
|
|
|
| 375 |
if 'single' in amd_link_raw and not amd_single_link:
|
| 376 |
amd_single_link = amd_link_raw['single']
|
| 377 |
|
|
|
|
| 378 |
if pd.notna(row.get('job_link_nvidia')) and (not nvidia_multi_link or not nvidia_single_link):
|
| 379 |
nvidia_link_raw = row.get('job_link_nvidia')
|
| 380 |
if isinstance(nvidia_link_raw, dict):
|
|
|
|
| 383 |
if 'single' in nvidia_link_raw and not nvidia_single_link:
|
| 384 |
nvidia_single_link = nvidia_link_raw['single']
|
| 385 |
|
|
|
|
| 386 |
if amd_multi_link and amd_single_link and nvidia_multi_link and nvidia_single_link:
|
| 387 |
break
|
| 388 |
|
| 389 |
|
|
|
|
| 390 |
links_md = "❓ [**FAQ**](https://huggingface.co/spaces/transformers-community/transformers-ci-dashboard/blob/main/README.md)\n\n"
|
| 391 |
links_md += "🔗 **CI Jobs:**\n\n"
|
| 392 |
|
|
|
|
| 393 |
if amd_multi_link or amd_single_link:
|
| 394 |
links_md += "**AMD:**\n"
|
| 395 |
if amd_multi_link:
|
|
|
|
| 398 |
links_md += f"• [Single GPU]({amd_single_link})\n"
|
| 399 |
links_md += "\n"
|
| 400 |
|
|
|
|
| 401 |
if nvidia_multi_link or nvidia_single_link:
|
| 402 |
links_md += "**NVIDIA:**\n"
|
| 403 |
if nvidia_multi_link:
|
|
|
|
| 414 |
return "🔗 **CI Jobs:** *Error loading links*\n\n❓ **[FAQ](README.md)**"
|
| 415 |
|
| 416 |
|
| 417 |
+
|
| 418 |
|
| 419 |
def get_historical_summary_plots():
|
| 420 |
"""Get historical summary plots from preloaded data."""
|
|
|
|
| 427 |
|
| 428 |
def handle_history_toggle(history_mode, last_selected_model, in_model_view):
|
| 429 |
if history_mode:
|
|
|
|
| 430 |
if in_model_view and last_selected_model:
|
| 431 |
amd_ts, nvidia_ts = show_time_series_model(last_selected_model)
|
| 432 |
return (
|
| 433 |
+
gr.update(visible=False),
|
| 434 |
+
gr.update(visible=True),
|
| 435 |
+
gr.update(visible=False),
|
| 436 |
+
gr.update(visible=False),
|
| 437 |
+
gr.update(visible=False),
|
| 438 |
+
gr.update(visible=False),
|
| 439 |
+
gr.update(visible=False),
|
| 440 |
+
amd_ts,
|
| 441 |
+
nvidia_ts,
|
| 442 |
+
gr.update(visible=True),
|
| 443 |
+
gr.update(),
|
| 444 |
+
gr.update(),
|
| 445 |
+
gr.update(),
|
| 446 |
+
True,
|
| 447 |
)
|
|
|
|
| 448 |
fr_plot, amd_plot, nvidia_plot = get_historical_summary_plots()
|
| 449 |
return (
|
| 450 |
+
gr.update(visible=False),
|
| 451 |
+
gr.update(visible=True),
|
| 452 |
+
gr.update(visible=False),
|
| 453 |
+
gr.update(visible=False),
|
| 454 |
+
fr_plot,
|
| 455 |
+
amd_plot,
|
| 456 |
+
nvidia_plot,
|
| 457 |
+
gr.update(),
|
| 458 |
+
gr.update(),
|
| 459 |
+
gr.update(visible=False),
|
| 460 |
+
gr.update(),
|
| 461 |
+
gr.update(),
|
| 462 |
+
gr.update(),
|
| 463 |
+
False,
|
| 464 |
)
|
| 465 |
else:
|
|
|
|
| 466 |
if last_selected_model and Ci_results.df is not None and not Ci_results.df.empty and last_selected_model in Ci_results.df.index:
|
| 467 |
+
fig, amd_txt, nvidia_txt = plot_model_stats(Ci_results.df, last_selected_model, COLORS=COLORS)
|
| 468 |
return (
|
| 469 |
+
gr.update(visible=True),
|
| 470 |
+
gr.update(visible=False),
|
| 471 |
+
gr.update(visible=False),
|
| 472 |
+
gr.update(visible=True),
|
| 473 |
+
gr.update(visible=False),
|
| 474 |
+
gr.update(visible=False),
|
| 475 |
+
gr.update(visible=False),
|
| 476 |
+
gr.update(),
|
| 477 |
+
gr.update(),
|
| 478 |
+
gr.update(visible=False),
|
| 479 |
+
fig,
|
| 480 |
+
amd_txt,
|
| 481 |
+
nvidia_txt,
|
| 482 |
+
True,
|
| 483 |
)
|
| 484 |
else:
|
| 485 |
+
fig = create_summary_page(Ci_results.df, Ci_results.available_models, COLORS=COLORS)
|
| 486 |
return (
|
| 487 |
+
gr.update(visible=True),
|
| 488 |
+
gr.update(visible=False),
|
| 489 |
+
gr.update(value=fig, visible=True),
|
| 490 |
+
gr.update(visible=False),
|
| 491 |
+
gr.update(visible=False),
|
| 492 |
+
gr.update(visible=False),
|
| 493 |
+
gr.update(visible=False),
|
| 494 |
+
gr.update(),
|
| 495 |
+
gr.update(),
|
| 496 |
+
gr.update(visible=False),
|
| 497 |
+
gr.update(),
|
| 498 |
+
gr.update(),
|
| 499 |
+
gr.update(),
|
| 500 |
+
False,
|
| 501 |
)
|
| 502 |
|
| 503 |
history_view_button.change(
|
|
|
|
| 522 |
)
|
| 523 |
|
| 524 |
|
|
|
|
| 525 |
def show_time_series_model(selected_model):
|
| 526 |
"""Show time-series view for a specific model."""
|
| 527 |
dfs = get_model_time_series_dfs(Ci_results.historical_df, selected_model)
|
|
|
|
| 530 |
gr.update(value=dfs['nvidia_df'], visible=True, title=f"{selected_model.upper()} - NVIDIA Results Over Time"),
|
| 531 |
)
|
| 532 |
|
|
|
|
| 533 |
def handle_model_click(selected_model: str, history_mode: bool):
|
| 534 |
if history_mode:
|
| 535 |
amd_ts, nvidia_ts = show_time_series_model(selected_model)
|
| 536 |
return (
|
| 537 |
+
gr.update(),
|
| 538 |
+
gr.update(),
|
| 539 |
+
gr.update(),
|
| 540 |
+
gr.update(visible=False),
|
| 541 |
+
gr.update(visible=True),
|
| 542 |
+
gr.update(visible=False),
|
| 543 |
+
gr.update(visible=False),
|
| 544 |
+
gr.update(visible=False),
|
| 545 |
+
gr.update(visible=False),
|
| 546 |
+
gr.update(visible=False),
|
| 547 |
+
amd_ts,
|
| 548 |
+
nvidia_ts,
|
| 549 |
+
gr.update(visible=True),
|
| 550 |
+
selected_model, True)
|
| 551 |
else:
|
| 552 |
+
fig, amd_txt, nvidia_txt = plot_model_stats(Ci_results.df, selected_model, COLORS=COLORS)
|
| 553 |
return (
|
| 554 |
fig,
|
| 555 |
amd_txt,
|
| 556 |
nvidia_txt,
|
| 557 |
+
gr.update(visible=True),
|
| 558 |
+
gr.update(visible=False),
|
| 559 |
+
gr.update(visible=False),
|
| 560 |
+
gr.update(visible=True),
|
| 561 |
+
gr.update(),
|
| 562 |
+
gr.update(),
|
| 563 |
+
gr.update(),
|
| 564 |
+
gr.update(),
|
| 565 |
+
gr.update(),
|
| 566 |
+
gr.update(visible=False),
|
| 567 |
+
selected_model, True)
|
| 568 |
|
| 569 |
for i, btn in enumerate(model_buttons):
|
| 570 |
model_name = model_choices[i]
|
|
|
|
| 590 |
],
|
| 591 |
)
|
| 592 |
|
|
|
|
| 593 |
demo.load(
|
| 594 |
fn=get_ci_links,
|
| 595 |
outputs=[ci_links_display]
|
| 596 |
)
|
| 597 |
|
| 598 |
|
|
|
|
| 599 |
if __name__ == "__main__":
|
| 600 |
+
demo.launch()
|