fix sem_conn_all undefined-name bug

diag.py

@@ -4,7 +4,7 @@ Variants (head-only overfits, one fixed batch, per-frame-mean loss, lr 2e-4):
   B cond = frozen connect features of the target frames
   C cond = LLM hidden states (real pathway)
   D cond = hidden states with voice slots acoustic-only (inference-style _process_speech_inputs)
-Plus: initial per-element loss bucketed by timestep range; then a 30-step full overfit (C).
+Plus: initial per-element loss bucketed by timestep range.
 """
 import io, math, random, sys
 sys.path.insert(0, "/VibeVoice-classic")
@@ -43,9 +43,11 @@ def load_dialogues(shard_idx, min_turns=4, max_turns=10, min_sec=20.0, max_sec=6
             try:
                 wav, sr = sf.read(io.BytesIO(auds[i]["bytes"]))
             except Exception:
-                ok = False; break
+                ok = False
+                break
             if sr != SR or wav.ndim > 1:
-                ok = False; break
+                ok = False
+                break
             turns.append((1 + int(spks[i]), texts[i], wav.astype(np.float32)))
         if not ok:
             continue
@@ -76,7 +78,7 @@ def build_sample(model, proc, tok, dialogue, refs, turn_sec_cap, gap_sec=0.45):
     voice_masks = out["speech_masks"].cpu().numpy().astype(bool)
     n_voice = voice_np.shape[0]
     with torch.no_grad():
-        _enc = model.model.acoustic_tokenizer.encode(torch.from_numpy(target := np.concatenate(parts)).float()[None, None, :])
+        _enc = model.model.acoustic_tokenizer.encode(torch.from_numpy(target).float()[None, None, :])
         t_lat, _ = _enc.sample("fix")
     T_t = t_lat.shape[1]
     lat_segs = []
@@ -125,7 +127,8 @@ def build_sample(model, proc, tok, dialogue, refs, turn_sec_cap, gap_sec=0.45):
         "sems": torch.from_numpy(sems),
     }
 
-random.seed(0); torch.manual_seed(0)
+random.seed(0)
+torch.manual_seed(0)
 model = VibeVoiceForConditionalGeneration.from_pretrained(
     MODEL, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).to(device).eval()
 proc = VibeVoiceProcessor.from_pretrained(MODEL)
@@ -193,16 +196,15 @@ def overfit(name, cond, steps=10):
     model.model.prediction_head.load_state_dict(head0)
 
 overfit("A cond = clean latent x0 (cheat)", x0r)
-conn_all = conn[smask] + sem_conn_all if (sem_conn := model.model.semantic_connector(b["sems"].to(device).type_as(x))) is not None else conn[smask]
+conn_all = conn[smask] + sem_conn[smask]
 conn_t = conn_all[-N:]
 overfit("B cond = target connect feats", conn_t.repeat_interleave(mul, 0))
 overfit("C cond = LLM hidden states", condr)
 # D: inference-style — voice slots get acoustic-only connect features
 with torch.no_grad():
     xe_d = x.clone()
-    xe_d[b["in_mask"].unsqueeze(0).to(device)] = torch.cat([conn[smask], sem_conn[smask]], 0) * 0
-    xe_d[b["in_mask"].unsqueeze(0).to(device)] = torch.cat([
-        conn[smask], conn[smask][-N:] + sem_conn[smask][-N:]], 0)
+    voice_rows = b["in_mask"].sum() - N
+    xe_d[b["in_mask"].unsqueeze(0).to(device)] = torch.cat([conn[smask][:-N], conn_all[-N:]], 0)
     hidden_d = model.model.language_model(
         inputs_embeds=xe_d, attention_mask=torch.ones(1, L, device=device), use_cache=False).last_hidden_state
     cond_d = hidden_d[b["lo_mask"].unsqueeze(0).to(device)].repeat_interleave(mul, 0)