Standart yukleme: auto_map + ErkLinearForCausalLM

configuration_erk_linear.py

@@ -0,0 +1,45 @@
+"""Erk-Linear yapilandirmasi — %20-lineer hibrit (8/40 dikkat katmani Gated DeltaNet)."""
+from transformers import PretrainedConfig
+
+
+class ErkLinearConfig(PretrainedConfig):
+    """Hibridin kendisi bir govde tasimaz; govde `base_model`'den yuklenir.
+
+    Bu yapilandirma yalnizca hangi katmanlarin lineerlestirildigini ve Gated DeltaNet
+    modullerinin nasil kurulacagini tanimlar. Agirliklar iki kaynaktan gelir:
+      - govde (32 softmax katmani + gomme/LM basi) : `base_model` deposundan
+      - 8 GDN katmani                              : bu deponun gdn_weights.safetensors
+    """
+
+    model_type = "erk_linear"
+
+    def __init__(
+        self,
+        base_model: str = "ecloudtech/Erk-14B",
+        gdn_layers=None,
+        hidden_size: int = 5120,
+        num_hidden_layers: int = 40,
+        gdn_head_dim: int = 128,
+        gdn_num_heads: int = 40,
+        gdn_use_gate: bool = True,
+        gdn_use_short_conv: bool = True,
+        gdn_mode: str = "chunk",
+        gdn_weights_file: str = "gdn_weights.safetensors",
+        **kwargs,
+    ):
+        self.base_model = base_model
+        self.gdn_layers = list(gdn_layers) if gdn_layers is not None else [1, 3, 5, 7, 10, 36, 38, 39]
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.gdn_head_dim = gdn_head_dim
+        self.gdn_num_heads = gdn_num_heads
+        self.gdn_use_gate = gdn_use_gate
+        self.gdn_use_short_conv = gdn_use_short_conv
+        self.gdn_mode = gdn_mode
+        self.gdn_weights_file = gdn_weights_file
+        super().__init__(**kwargs)
+
+    @property
+    def linear_ratio(self) -> float:
+        """Lineerlestirilen dikkat katmanlarinin orani (8/40 = 0.20)."""
+        return len(self.gdn_layers) / float(self.num_hidden_layers)