Create config.json
Browse files- config.json +28 -0
config.json
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"producer": {
|
| 3 |
+
"name": "ammo",
|
| 4 |
+
"version": "0.7.4"
|
| 5 |
+
},
|
| 6 |
+
"architecture": "LlamaForCausalLM",
|
| 7 |
+
"dtype": "float16",
|
| 8 |
+
"num_hidden_layers": 80,
|
| 9 |
+
"num_attention_heads": 64,
|
| 10 |
+
"num_key_value_heads": 8,
|
| 11 |
+
"hidden_size": 8192,
|
| 12 |
+
"norm_epsilon": 1e-05,
|
| 13 |
+
"vocab_size": 128256,
|
| 14 |
+
"max_position_embeddings": 8192,
|
| 15 |
+
"hidden_act": "silu",
|
| 16 |
+
"use_parallel_embedding": true,
|
| 17 |
+
"embedding_sharding_dim": 0,
|
| 18 |
+
"quantization": {
|
| 19 |
+
"quant_algo": "W4A16_AWQ",
|
| 20 |
+
"kv_cache_quant_algo": "INT8",
|
| 21 |
+
"group_size": 64,
|
| 22 |
+
"has_zero_point": false,
|
| 23 |
+
"pre_quant_scale": true,
|
| 24 |
+
"exclude_modules": [
|
| 25 |
+
"lm_head"
|
| 26 |
+
]
|
| 27 |
+
}
|
| 28 |
+
}
|