Archive viu_1_5b_moe_config.yaml to deprecated folder
Browse files
model/configs/deprecated/viu_1_5b_moe_config.yaml
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# [DEPRECATED / LEGACY ARCHIVE] — Active 32L Dense model uses: viu1_dense_config.yaml
|
| 2 |
+
# VIU-1.5B Frontier MoE Architecture (Archived)
|
| 3 |
+
# Deep Hierarchical Reasoning (Gemma-2 Style 42 Layers x 28 Experts = 1,176 Micro-Experts)
|
| 4 |
+
|
| 5 |
+
dim: 768
|
| 6 |
+
n_layers: 42
|
| 7 |
+
n_heads: 12
|
| 8 |
+
n_kv_heads: 2
|
| 9 |
+
vocab_size: 48000
|
| 10 |
+
multiple_of: 64
|
| 11 |
+
ffn_dim_multiplier: 2.7
|
| 12 |
+
norm_eps: 1.0e-05
|
| 13 |
+
rope_theta: 500000.0
|
| 14 |
+
max_seq_len: 2048
|
| 15 |
+
sliding_window: 512
|
| 16 |
+
hybrid_ratio: 3
|
| 17 |
+
dropout: 0.0
|
| 18 |
+
bias: false
|
| 19 |
+
tie_embeddings: true
|
| 20 |
+
qk_norm: true
|
| 21 |
+
|
| 22 |
+
# Ultra-Deep MoE Specifications
|
| 23 |
+
moe_every: 1
|
| 24 |
+
moe_num_experts: 28
|
| 25 |
+
moe_top_k: 2
|
| 26 |
+
moe_expert_div: 4
|
| 27 |
+
moe_shared: false
|
| 28 |
+
moe_balance_weight: 0.05
|
| 29 |
+
moe_capacity_factor: 1.25
|
| 30 |
+
|
| 31 |
+
# Multi-Token Prediction (Forward Planning)
|
| 32 |
+
switch_gate: false
|
| 33 |
+
multi_token: true
|