export: disable non-ONNX group planner

scripts/export_8k.py

@@ -34,7 +34,8 @@ class DecoderHead(torch.nn.Module):
         x = frames + m.frame_proj(frame_meta) + m.local_ctx(local_ctx_raw)
         x = x + m.abs_frame(abs_pos)
         if m.cfg.use_frame_pitch:
-            x = x + m.pitch_proj(pitch_frame)
+            refined = m.refine_frame_pitch(frames, frame_meta, pitch_frame)
+            x = x + m.pitch_proj(refined)
         for blk in m.decoder:
             x = blk(x, frame_mask)
         x = x + m.frame_gru(x)[0]
@@ -73,7 +74,13 @@ def main():
 
     ac = torch.load(args.acoustic_ckpt, map_location=dev, weights_only=False)
     cfg = MicroFastSpeechConfig(**ac["config"])
-    m = MicroFastSpeech(cfg); m.load_state_dict(ac["model"]); m.eval()
+    m = MicroFastSpeech(cfg); m.load_state_dict(ac["model"], strict=False); m.eval()
+    # The group-duration planner uses a non-ONNX-able host loop and only adjusts inference-time
+    # durations (the mel decoder is trained on GT durations). Disable it so the exported encoder's
+    # plain-duration path (which keeps the contextual duration-delta) matches m.infer() for parity.
+    if getattr(m, "group_duration_delta", None) is not None:
+        m.group_duration_delta = None
+        print("note: group_duration_planner disabled at export (host-loop; plain durations used)")
     enc, dec = EncoderHead(m).eval(), DecoderHead(m).eval()
     print(f"acoustic: sr={cfg.sample_rate} vocab={cfg.vocab_size} tone={cfg.tone_size} lang={cfg.lang_size} "
           f"abs_bins={cfg.abs_frame_bins} max_frames={cfg.max_frames}")