Replace model_config.yaml (full 98-field dataclass dump) with a minimal config.json (only the fields this model actually sets)
Browse files- config.json +22 -0
config.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"hidden_size": 896,
|
| 3 |
+
"num_hidden_layers": 16,
|
| 4 |
+
"num_attention_heads": 14,
|
| 5 |
+
"num_key_value_heads": 2,
|
| 6 |
+
"intermediate_size": 256,
|
| 7 |
+
"shared_intermediate_size": 3072,
|
| 8 |
+
"vocab_size": 32000,
|
| 9 |
+
"max_position_embeddings": 2048,
|
| 10 |
+
"attention_type": "gqa",
|
| 11 |
+
"mlp_type": "tr_hash_engine",
|
| 12 |
+
"num_experts": 4,
|
| 13 |
+
"top_k": 2,
|
| 14 |
+
"top_k_primary_weight": 0.5,
|
| 15 |
+
"routing_strategy": "token_id_multi_hash",
|
| 16 |
+
"route_hash_count": 2,
|
| 17 |
+
"shared_expert": true,
|
| 18 |
+
"shared_output_scale": 1.0,
|
| 19 |
+
"routed_output_scale": 2.0,
|
| 20 |
+
"use_qk_norm": true,
|
| 21 |
+
"tie_word_embeddings": true
|
| 22 |
+
}
|