modeling: use_cache=True icin GDN durum-devretme

modeling_erk_linear.py

@@ -21,15 +21,58 @@ REPO_ID = "ecloudtech/Erk-Linear"
 GDN_LAYERS = [1, 3, 5, 7, 10, 36, 38, 39]   # %20 lineer, yayilmis yerlesim
 
 
+class _GDNStateCache:
+    """GatedDeltaNet'in get/update_layer_cache arayuzunun bekledigi minimal katman-durum tutucu.
+
+    FLA'nin recurrent_state + conv_state'ini tek katman icin saklar; boylece cache'li uretim
+    sirasinda GDN gecmis durumu adimlar arasi devreder.
+    """
+    def __init__(self):
+        self._layers = []
+
+    def __len__(self):
+        return len(self._layers)
+
+    def __getitem__(self, idx):
+        return self._layers[idx]
+
+    def update(self, layer_idx=0, recurrent_state=None, conv_state=None, **kwargs):
+        while len(self._layers) <= layer_idx:
+            self._layers.append({"recurrent_state": None, "conv_state": None})
+        if recurrent_state is not None:
+            self._layers[layer_idx]["recurrent_state"] = recurrent_state
+        if conv_state is not None:
+            self._layers[layer_idx]["conv_state"] = conv_state
+        return self
+
+
 class _GDNAttention(nn.Module):
-    """Qwen3 self_attn cagri imzasiyla uyumlu Gated DeltaNet sarmalayici."""
+    """Qwen3 self_attn cagri imzasiyla uyumlu Gated DeltaNet sarmalayici.
+
+    Cache'li uretim (use_cache=True) sirasinda GDN'nin recurrent + convolution durumunu
+    adimlar arasi devreder; boylece model.generate() ciktisi, tam-yeniden-hesaplama
+    (use_cache=False) ile sayisal gurultuye kadar ayni olur. Referans amacli tek-dizi
+    kullanim icindir (es zamanli/batch-paylasimli servis icin ayri durum yonetimi gerekir).
+    """
     def __init__(self, gdn):
         super().__init__()
+        gdn.layer_idx = 0
         self.gdn = gdn
+        self._state = None
 
     def forward(self, hidden_states, *args, **kwargs):
-        out = self.gdn(hidden_states)
+        cache_position = kwargs.get("cache_position", None)
+        seq_len = hidden_states.shape[1]
+        new_sequence = (
+            (cache_position is None and seq_len > 1)
+            or (cache_position is not None and int(cache_position.reshape(-1)[0]) == 0)
+        )
+        if new_sequence or self._state is None:
+            self._state = _GDNStateCache()
+        out = self.gdn(hidden_states, use_cache=True, past_key_values=self._state)
         y = out[0] if isinstance(out, tuple) else out
+        if isinstance(out, tuple) and len(out) >= 3 and out[2] is not None:
+            self._state = out[2]
         return (y, None)