Upload model_v5.py with huggingface_hub

model_v5.py

@@ -212,8 +212,12 @@ class GPTV5(nn.Module):
         if targets is not None:
             logits = self.lm_head(x)
             if self.cfg.logit_softcap > 0:
+                # In-place softcap — bellek tasarrufu (compile yokken kritik)
+                # cap * tanh(logits / cap) — 3 tensor yerine in-place
                 cap = self.cfg.logit_softcap
-                logits = cap * torch.tanh(logits / cap)
+                logits.div_(cap)
+                logits = torch.tanh(logits)
+                logits.mul_(cap)
             flat_logits = logits.view(-1, logits.size(-1))
             flat_targets = targets.view(-1)
             if self._ce_loss is not None:
@@ -226,7 +230,9 @@ class GPTV5(nn.Module):
             logits = self.lm_head(x[:, [-1], :])
             if self.cfg.logit_softcap > 0:
                 cap = self.cfg.logit_softcap
-                logits = cap * torch.tanh(logits / cap)
+                logits.div_(cap)
+                logits = torch.tanh(logits)
+                logits.mul_(cap)
             return logits, None
 
     @torch.no_grad()