Add meta.json
Browse files
meta.json
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_id": "moonshotai/Kimi-K2-Instruct",
|
| 3 |
+
"moe_config": {
|
| 4 |
+
"model_type": "kimi_k2",
|
| 5 |
+
"num_layers": 61,
|
| 6 |
+
"hidden_size": 7168,
|
| 7 |
+
"intermediate_size": 18432,
|
| 8 |
+
"moe_intermediate_size": 2048,
|
| 9 |
+
"n_routed_experts": 384,
|
| 10 |
+
"n_shared_experts": 1,
|
| 11 |
+
"num_experts_per_tok": 8,
|
| 12 |
+
"first_k_dense_replace": 1,
|
| 13 |
+
"torch_dtype": "bfloat16",
|
| 14 |
+
"quant_method": "fp8"
|
| 15 |
+
},
|
| 16 |
+
"num_feats": 64,
|
| 17 |
+
"provenance": {
|
| 18 |
+
"aggregator_type": "moe",
|
| 19 |
+
"quant_format": "fp8",
|
| 20 |
+
"probe_mode": "weight_svd_per_expert",
|
| 21 |
+
"fp8_handling": "cast_to_bfloat16_for_svd"
|
| 22 |
+
}
|
| 23 |
+
}
|