Merge pull request #585 from orbitalquark/fix-qwen3-tts-streaming-memory-leak

lucasnewman · web-flow · commit 8083120b62fa · 2026-03-16T21:52:25.000-07:00
Fix Qwen3-TTS streaming memory leak.
diff --git a/mlx_audio/tts/models/qwen3_tts/qwen3_tts.py b/mlx_audio/tts/models/qwen3_tts/qwen3_tts.py
@@ -2072,6 +2072,7 @@ def _generate_icl(
                     is_final_chunk=True,
                 )
             self.speech_tokenizer.decoder.reset_streaming_state()
+            mx.clear_cache()
             return
 
         if not generated_codes:
@@ -2386,6 +2387,7 @@ def _generate_with_instruct(
                     is_final_chunk=True,
                 )
             self.speech_tokenizer.decoder.reset_streaming_state()
+            mx.clear_cache()
             return
 
         if not generated_codes:

Original file line number	Diff line number	Diff line change
`@@ -2072,6 +2072,7 @@ def _generate_icl(`
`2072`	`2072`	`is_final_chunk=True,`
`2073`	`2073`	`)`
`2074`	`2074`	`self.speech_tokenizer.decoder.reset_streaming_state()`
	`2075`	`+ mx.clear_cache()`
`2075`	`2076`	`return`
`2076`	`2077`
`2077`	`2078`	`if not generated_codes:`
`@@ -2386,6 +2387,7 @@ def _generate_with_instruct(`
`2386`	`2387`	`is_final_chunk=True,`
`2387`	`2388`	`)`
`2388`	`2389`	`self.speech_tokenizer.decoder.reset_streaming_state()`
	`2390`	`+ mx.clear_cache()`
`2389`	`2391`	`return`
`2390`	`2392`
`2391`	`2393`	`if not generated_codes:`