Add meta.json
Browse files
meta.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_id": "deepseek-ai/DeepSeek-V4-Pro",
|
| 3 |
+
"moe_config": {
|
| 4 |
+
"model_type": "deepseek_v4",
|
| 5 |
+
"num_layers": 61,
|
| 6 |
+
"hidden_size": 7168,
|
| 7 |
+
"intermediate_size": 18432,
|
| 8 |
+
"moe_intermediate_size": 3072,
|
| 9 |
+
"n_routed_experts": 384,
|
| 10 |
+
"n_shared_experts": 1,
|
| 11 |
+
"num_experts_per_tok": 6,
|
| 12 |
+
"first_k_dense_replace": 0,
|
| 13 |
+
"torch_dtype": "bfloat16",
|
| 14 |
+
"quant_method": "fp8",
|
| 15 |
+
"templates": {
|
| 16 |
+
"expert_gate_proj": "layers.{layer}.ffn.experts.{e}.w1.weight",
|
| 17 |
+
"expert_up_proj": "layers.{layer}.ffn.experts.{e}.w3.weight",
|
| 18 |
+
"expert_down_proj": "layers.{layer}.ffn.experts.{e}.w2.weight",
|
| 19 |
+
"fused_gate_proj": "layers.{layer}.ffn.experts.w1",
|
| 20 |
+
"fused_down_proj": "layers.{layer}.ffn.experts.w2",
|
| 21 |
+
"shared_down_proj": [
|
| 22 |
+
"layers.{layer}.ffn.shared_experts.w2.weight"
|
| 23 |
+
],
|
| 24 |
+
"router": [
|
| 25 |
+
"layers.{layer}.ffn.gate.weight",
|
| 26 |
+
"layers.{layer}.ffn.router.weight"
|
| 27 |
+
],
|
| 28 |
+
"dense_down_proj": [
|
| 29 |
+
"layers.{layer}.ffn.w2.weight"
|
| 30 |
+
]
|
| 31 |
+
}
|
| 32 |
+
},
|
| 33 |
+
"num_feats": 64,
|
| 34 |
+
"provenance": {
|
| 35 |
+
"aggregator_type": "moe",
|
| 36 |
+
"quant_format": "fp8",
|
| 37 |
+
"probe_mode": "weight_svd_per_expert",
|
| 38 |
+
"fp8_handling": "cast_to_bfloat16_for_svd"
|
| 39 |
+
}
|
| 40 |
+
}
|