From 173606dae722ee8f1d2d1b4c7226d55a60e6e061 Mon Sep 17 00:00:00 2001
From: Kazeia Team <support@kazeia.com>
Date: Thu, 9 Apr 2026 13:42:02 +0200
Subject: [PATCH] Stable: decoder 8T optimization + restore pre-computed embeds
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- BigVGAN: 8 threads (2757→1872ms), pre_conv/pre_transformer: 4 threads
- Restored pre-computed embeds format (codec_sum+text from Python)
- Text-only trailing embeds don't work: model needs codec_sum for EOS

For long phrases, pre-computed embeds must be generated from Python.
RTF 1.26 on short phrase.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 executorch-custom/jni_layer_tts.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/executorch-custom/jni_layer_tts.cpp b/executorch-custom/jni_layer_tts.cpp
index 5105ff2..a6cc15d 100644
--- a/executorch-custom/jni_layer_tts.cpp
+++ b/executorch-custom/jni_layer_tts.cpp
@@ -839,7 +839,8 @@ ExecuTorchJni::runTtsPipelineImpl(
         for(int i=0;i<NUM_CB;i++) allCodes.push_back(codes[i]);
         cb0Hist.push_back(curCb0);
 
-        // Next embed: pre-computed trailing OR codec_sum + eos/pad
+        // Next embed: use pre-computed trailing embeds (codec_sum+text from Python)
+        // then codec_sum + eos/pad after trailing exhausted
         float nextEmb[DIM]={};
         if(trIdx<nTrailing){
             memcpy(nextEmb,trailing.data()+trIdx*DIM,DIM*4);