Add Asteria config
Browse files- config.json +27 -0
config.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_type": "asteria_moe",
|
| 3 |
+
"architectures": [
|
| 4 |
+
"AsteriaModel"
|
| 5 |
+
],
|
| 6 |
+
"vocab_size": 163840,
|
| 7 |
+
"hidden_size": 7168,
|
| 8 |
+
"num_layers": 61,
|
| 9 |
+
"num_experts": 384,
|
| 10 |
+
"num_experts_per_tok": 8,
|
| 11 |
+
"num_shared_experts": 1,
|
| 12 |
+
"expert_ffn_dim": 2048,
|
| 13 |
+
"dense_ffn_dim": 18432,
|
| 14 |
+
"first_k_dense_replace": 1,
|
| 15 |
+
"num_attention_heads": 64,
|
| 16 |
+
"kv_lora_rank": 512,
|
| 17 |
+
"q_lora_rank": 1536,
|
| 18 |
+
"qk_nope_head_dim": 128,
|
| 19 |
+
"qk_rope_head_dim": 64,
|
| 20 |
+
"v_head_dim": 128,
|
| 21 |
+
"hidden_act": "silu",
|
| 22 |
+
"routed_scaling_factor": 2.827,
|
| 23 |
+
"dtype": "bfloat16",
|
| 24 |
+
"source_model": "moonshotai/Kimi-K2.7-Code",
|
| 25 |
+
"transfer_date": "2026-07-11 12:43:00",
|
| 26 |
+
"description": "Asteria custom model with Kimi K2.7 brain weights. Supports sparse loading and expert training."
|
| 27 |
+
}
|