Run-all-models: run concurrently and stream rows as each finishes
Browse filesThe first version ran the six models sequentially, so with the reasoning models
inheriting the Config default max_tokens (9192) a comparison could block for a
minute-plus with a blank table. Run them in a thread pool instead and make the
handler a generator that yields the table after each model returns: every model
shows "running" immediately and its verdict fills in as it lands, so total time is
about the slowest single model rather than the sum, and a stalled endpoint no
longer holds up the rest.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
app.py
CHANGED
|
@@ -167,15 +167,44 @@ def _comparison_note(model_info_dict, parsed, result) -> str:
|
|
| 167 |
return ""
|
| 168 |
|
| 169 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
def handle_run_all_models(test_input, current_policy, structured_rules, reasoning_effort,
|
| 171 |
max_tokens, temperature, top_p, system_prompt_val, response_format_val,
|
| 172 |
oauth_token: gr.OAuthToken | None = None):
|
| 173 |
-
"""Run the current content + policy against every model
|
| 174 |
|
| 175 |
This is the workshop's core move — the same input judged by each model side by side.
|
| 176 |
-
|
| 177 |
-
|
|
|
|
| 178 |
"""
|
|
|
|
|
|
|
| 179 |
if not test_input or not test_input.strip():
|
| 180 |
raise gr.Error("Please enter test content before running a comparison.")
|
| 181 |
if not current_policy or current_policy == "*No policy loaded*":
|
|
@@ -185,32 +214,30 @@ def handle_run_all_models(test_input, current_policy, structured_rules, reasonin
|
|
| 185 |
if hf_token is None:
|
| 186 |
raise gr.Error("Please log in or set tokens to use Inference Providers. Check the sidebar.")
|
| 187 |
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
verdict, note =
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
return rows, gr.update(visible=True)
|
| 214 |
|
| 215 |
|
| 216 |
# ============================================================================
|
|
|
|
| 167 |
return ""
|
| 168 |
|
| 169 |
|
| 170 |
+
def _run_one_model(model, test_input, current_policy, structured_rules, reasoning_effort,
|
| 171 |
+
max_tokens, temperature, top_p, system_prompt_val, response_format_val, hf_token):
|
| 172 |
+
"""Run a single model and return its [name, verdict, note] row. Never raises."""
|
| 173 |
+
try:
|
| 174 |
+
result = run_test(
|
| 175 |
+
model_id=model["id"], test_input=test_input, policy=current_policy, hf_token=hf_token,
|
| 176 |
+
reasoning_effort=reasoning_effort, max_tokens=int(max_tokens),
|
| 177 |
+
temperature=float(temperature), top_p=float(top_p),
|
| 178 |
+
system_prompt=system_prompt_val, response_format=response_format_val,
|
| 179 |
+
structured_rules=structured_rules,
|
| 180 |
+
)
|
| 181 |
+
parsed = parse_json_response(result.get("content", ""))
|
| 182 |
+
label = parsed.get("label", -1)
|
| 183 |
+
verdict = (
|
| 184 |
+
"❌ Violation" if label == 1
|
| 185 |
+
else "✅ No violation" if label == 0
|
| 186 |
+
else "⚠️ Unclear"
|
| 187 |
+
)
|
| 188 |
+
note = _comparison_note(model, parsed, result)
|
| 189 |
+
except gr.Error as e:
|
| 190 |
+
verdict, note = "⚠️ Unavailable", str(e)[:200]
|
| 191 |
+
except Exception as e:
|
| 192 |
+
verdict, note = "⚠️ Error", str(e)[:200]
|
| 193 |
+
return [model["name"], verdict, note]
|
| 194 |
+
|
| 195 |
+
|
| 196 |
def handle_run_all_models(test_input, current_policy, structured_rules, reasoning_effort,
|
| 197 |
max_tokens, temperature, top_p, system_prompt_val, response_format_val,
|
| 198 |
oauth_token: gr.OAuthToken | None = None):
|
| 199 |
+
"""Run the current content + policy against every model, streaming the comparison table.
|
| 200 |
|
| 201 |
This is the workshop's core move — the same input judged by each model side by side.
|
| 202 |
+
Models run concurrently (reasoning models are slow, so sequential runs blocked for
|
| 203 |
+
minutes), and the table is yielded as each one finishes, so rows fill in live and a
|
| 204 |
+
model whose endpoint is down shows as unavailable rather than stalling the rest.
|
| 205 |
"""
|
| 206 |
+
from concurrent.futures import ThreadPoolExecutor
|
| 207 |
+
|
| 208 |
if not test_input or not test_input.strip():
|
| 209 |
raise gr.Error("Please enter test content before running a comparison.")
|
| 210 |
if not current_policy or current_policy == "*No policy loaded*":
|
|
|
|
| 214 |
if hf_token is None:
|
| 215 |
raise gr.Error("Please log in or set tokens to use Inference Providers. Check the sidebar.")
|
| 216 |
|
| 217 |
+
# Start every model at once; show them all as "running" immediately so the table
|
| 218 |
+
# isn't blank while the slow reasoning models work.
|
| 219 |
+
status = {m["name"]: "⏳ running…" for m in MODELS}
|
| 220 |
+
notes = {m["name"]: "" for m in MODELS}
|
| 221 |
+
order = [m["name"] for m in MODELS]
|
| 222 |
+
|
| 223 |
+
def table():
|
| 224 |
+
return [[name, status[name], notes[name]] for name in order]
|
| 225 |
+
|
| 226 |
+
yield table(), gr.update(visible=True)
|
| 227 |
+
|
| 228 |
+
with ThreadPoolExecutor(max_workers=len(MODELS)) as executor:
|
| 229 |
+
futures = {
|
| 230 |
+
executor.submit(
|
| 231 |
+
_run_one_model, m, test_input, current_policy, structured_rules, reasoning_effort,
|
| 232 |
+
max_tokens, temperature, top_p, system_prompt_val, response_format_val, hf_token,
|
| 233 |
+
): m for m in MODELS
|
| 234 |
+
}
|
| 235 |
+
from concurrent.futures import as_completed
|
| 236 |
+
for fut in as_completed(futures):
|
| 237 |
+
name, verdict, note = fut.result()
|
| 238 |
+
status[name] = verdict
|
| 239 |
+
notes[name] = note
|
| 240 |
+
yield table(), gr.update(visible=True)
|
|
|
|
|
|
|
| 241 |
|
| 242 |
|
| 243 |
# ============================================================================
|