Update PersadianNanoV4Config.json

PersadianNanoV4Config.json

@@ -1,49 +1,36 @@
-class PersadianNanoV4Config:
-    def __init__(self):
+{
+  "hidden_size": 768,
+  "intermediate_size": 2048,
+  "num_hidden_layers": 12,
 
-        # Core architecture (MATCH CHECKPOINT)
-        self.hidden_size = 512
-        self.intermediate_size = 1024
-        self.num_hidden_layers = 12
+  "num_attention_heads": 12,
+  "num_key_value_heads": 4,
 
-        # Attention
-        self.num_attention_heads = 8
-        self.num_key_value_heads = 4
+  "num_experts": 8,
+  "num_experts_per_tok": 2,
 
-        # MoE (MATCH CHECKPOINT)
-        self.num_experts = 4
-        self.num_experts_per_tok = 2
+  "progressive_experts": true,
+  "min_experts": 1,
+  "max_experts": 2,
 
-        # Progressive experts
-        self.progressive_experts = True
-        self.min_experts = 1
-        self.max_experts = 2
+  "use_hyper_connection": true,
 
-        # Hyper connections
-        self.use_hyper_connection = True
+  "use_compressed_attention": true,
+  "compress_ratio": 4,
 
-        # Compression
-        self.use_compressed_attention = True
-        self.compress_ratio = 4
+  "max_position_embeddings": 2048,
 
-        # Context
-        self.max_position_embeddings = 2048
+  "vocab_size": 50257,
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
 
-        # Tokenizer
-        self.vocab_size = 50257
-        self.bos_token_id = 50256
-        self.eos_token_id = 50256
+  "rope_theta": 10000.0,
 
-        # RoPE
-        self.rope_theta = 10000.0
+  "dropout": 0.1,
+  "layer_norm_eps": 1e-5,
 
-        # Regularization
-        self.dropout = 0.1
-        self.layer_norm_eps = 1e-5
+  "attention_dropout": 0.0,
+  "use_flash_attention": true,
 
-        # Attention
-        self.attention_dropout = 0.0
-        self.use_flash_attention = True
-
-        # Precision (IMPORTANT: keep as attribute, not torch dtype object)
-        self.torch_dtype = "float16"
+  "torch_dtype": "float16"
+}