Add editable LLM penalties and preserve sampling through backup restore

test_endpoint.py

@@ -209,6 +209,17 @@ class EndpointTests(unittest.TestCase):
    self.assertEqual(status,200);self.assertEqual(data['model'],name)
   self.assertEqual(self.worker.stopped,['alpha']);self.assertEqual(self.worker.requests[-1]['temperature'],.2)
   self.assertEqual(self.ep.status()['counts']['llm']['enabled'],2)
+ def test_profile_penalties_and_explicit_client_override(self):
+  self.enable('alpha')
+  self.rows[0]['parameters'].update(repeat_penalty=1.15,presence_penalty=.5,frequency_penalty=.2)
+  request=dict(model='alpha',messages=[dict(role='user',content='synthetic')])
+  self.assertEqual(self.request('/v1/chat/completions',request)[0],200)
+  sent=self.worker.requests[-1]
+  self.assertEqual([sent[k] for k in ('repeat_penalty','presence_penalty','frequency_penalty')],[1.15,.5,.2])
+  self.assertEqual(self.request('/v1/chat/completions',dict(request,presence_penalty=0,repeat_penalty=1))[0],200)
+  self.assertEqual(self.worker.requests[-1]['presence_penalty'],0)
+  self.assertEqual(self.worker.requests[-1]['repeat_penalty'],1)
+  self.assertEqual(self.request('/v1/chat/completions',dict(request,repeat_penalty=-1))[0],400)
  def test_unknown_or_unsupported_request_does_not_load(self):
   self.enable('alpha')
   for req in [dict(model='unknown',messages=[{}]),dict(model='alpha',messages=[dict(content=[dict(type='image_url')])]),dict(model='alpha',messages=[{}],cache_file='/tmp/foo')]: