model loading/unloading works

models/tacotron2/vocodes_common.py

@@ -160,9 +160,15 @@ class TacotronWaveglowPipeline:
         print('Loading tacotron model into memory: {}'.format(checkpoint_path), flush=True)
         model = load_tacotron_model(checkpoint_path)
         self.tacotron_model_cache[checkpoint_path] = model
+        print('Now {} synthesizer models in memory'.format(len(self.tacotron_model_cache)))
 
     def uncache_tacotron_model(self, checkpoint_path):
+        print('Currently {} synthesizer models in memory'.format(
+            len(self.tacotron_model_cache)))
+        print('Clearing synthesizer from memory: {}'.format(checkpoint_path))
         self.tacotron_model_cache.pop(checkpoint_path, None)
+        print('Now reduced to {} synthesizer models in memory'.format(
+            len(self.tacotron_model_cache)))
 
     def infer(self, args):
         assert('synthesizer_checkpoint_path' in args)

models/tacotron2/vocodes_server.py

@@ -229,7 +229,6 @@ class ApiHandler():
             maybe_clear_synthesizer_checkpoint_path))
 
         if maybe_clear_synthesizer_checkpoint_path:
-            print('Clearing synthesizer: {}'.format(maybe_clear_synthesizer_checkpoint_path))
             pipeline.uncache_tacotron_model(maybe_clear_synthesizer_checkpoint_path)
 
         pipeline.maybe_load_waveglow_model(vocoder_checkpoint_path)

src/clients/tts_inference_sidecar_client.rs

@@ -20,7 +20,9 @@ struct InferenceRequest {
   pub output_audio_filename : String,
   pub output_spectrogram_filename : String,
   pub output_metadata_filename : String,
-  pub maybe_unload_model_path: Option<String>,
+
+  /// To instruct the sidecar to unload the model from memory
+  pub maybe_clear_synthesizer_checkpoint_path: Option<String>,
 }
 
 impl TtsInferenceSidecarClient {
@@ -82,7 +84,7 @@ impl TtsInferenceSidecarClient {
       output_audio_filename,
       output_spectrogram_filename,
       output_metadata_filename,
-      maybe_unload_model_path,
+      maybe_clear_synthesizer_checkpoint_path: maybe_unload_model_path,
     };
 
     let url = format!("http://{}/infer", self.hostname);

src/tts_inference_job.rs

@@ -532,7 +532,7 @@ async fn process_job(
   info!("Expected output audio filename: {:?}", &output_audio_fs_path);
   info!("Expected output spectrogram filename: {:?}", &output_spectrogram_fs_path);
   info!("Expected output metadata filename: {:?}", &output_metadata_fs_path);
-  info!("Maybe unload model: {:?}", &maybe_unload_model_path);
+  warn!("Maybe unload model: {:?}", &maybe_unload_model_path);
 
   //inferencer.tts_inference_command.execute(
   //  &tts_synthesizer_fs_path,