Update PersadianNanoV4Model.py

PersadianNanoV4Model.py

@@ -14,17 +14,17 @@ import torch.nn.functional as F
 class PersadianNanoV4Config:
     def __init__(self):
 
-        # Core architecture
-        self.hidden_size = 768
-        self.intermediate_size = 2048
+        # Core architecture (MATCH CHECKPOINT)
+        self.hidden_size = 512
+        self.intermediate_size = 1024
         self.num_hidden_layers = 12
 
         # Attention
-        self.num_attention_heads = 12
+        self.num_attention_heads = 8
         self.num_key_value_heads = 4
 
         # MoE
-        self.num_experts = 8
+        self.num_experts = 4
         self.num_experts_per_tok = 2
 
         # Progressive experts
@@ -58,8 +58,8 @@ class PersadianNanoV4Config:
         self.attention_dropout = 0.0
         self.use_flash_attention = True
 
-        # Precision
-        self.torch_dtype = torch.float16
+        # Precision (IMPORTANT: keep as string or remove entirely)
+        self.torch_dtype = "float16"
 
 
 # ============================================================