fix: inv_freq calcolato runtime, non buffer (evita corruzione meta-device)

modeling_quark.py

@@ -26,18 +26,21 @@ class RMSNorm(nn.Module):
 class RotaryEmbedding(nn.Module):
     def __init__(self, head_dim, max_seq_len, theta=10_000.0):
         super().__init__()
-        inv_freq = 1.0 / (theta ** (torch.arange(0, head_dim, 2).float() / head_dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        # head_dim/theta come Python float, NON tensori gestiti da HF —
+        # evita corruzione da meta-device init durante from_pretrained()
+        self.head_dim    = head_dim
+        self.theta       = theta
         self.max_seq_len = max_seq_len
-        self._max = 0   # forza build al primo forward, mai nell'__init__
-        self.cos_cache = None
-        self.sin_cache = None
+        self._max        = 0
+        self.cos_cache   = None
+        self.sin_cache   = None
 
     def _build_cache(self, seq_len, device, dtype):
-        # Ricalcola sempre da inv_freq corrente (mai cache stantia da meta-device)
-        t     = torch.arange(seq_len, device=device, dtype=torch.float32)
-        freqs = torch.outer(t, self.inv_freq.to(device=device, dtype=torch.float32))
-        emb   = torch.cat([freqs, freqs], dim=-1)
+        # Ricalcola inv_freq da zero ogni volta — niente stato persistito
+        inv_freq = 1.0 / (self.theta ** (torch.arange(0, self.head_dim, 2, device=device).float() / self.head_dim))
+        t        = torch.arange(seq_len, device=device, dtype=torch.float32)
+        freqs    = torch.outer(t, inv_freq)
+        emb      = torch.cat([freqs, freqs], dim=-1)
         self.cos_cache = emb.cos()[None, None].to(dtype)
         self.sin_cache = emb.sin()[None, None].to(dtype)
         self._max = seq_len