Add architecture config (max_loops=16 table capacity, run at 8)
Browse files- config.json +32 -0
config.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"vocab_size": 4096,
|
| 3 |
+
"dim": 96,
|
| 4 |
+
"n_heads": 8,
|
| 5 |
+
"prelude_layers": 0,
|
| 6 |
+
"coda_layers": 1,
|
| 7 |
+
"shared_loops": 16,
|
| 8 |
+
"lora_rank": 4,
|
| 9 |
+
"ffn_hidden": 2112,
|
| 10 |
+
"max_seq_len": 4096,
|
| 11 |
+
"max_context_len": 8192,
|
| 12 |
+
"rope_base": 10000.0,
|
| 13 |
+
"index_dim": 16,
|
| 14 |
+
"index_top_k": 128,
|
| 15 |
+
"local_window": 64,
|
| 16 |
+
"sparse_chunk_threshold": 4096,
|
| 17 |
+
"sparse_chunk_queries": 1024,
|
| 18 |
+
"ssa_block_size": 64,
|
| 19 |
+
"ssa_top_k_blocks": 8,
|
| 20 |
+
"attention_window": 256,
|
| 21 |
+
"fp4_weights": false,
|
| 22 |
+
"thinking_enabled": false,
|
| 23 |
+
"thinking_steps_min": 0,
|
| 24 |
+
"thinking_steps_max": 0,
|
| 25 |
+
"thinking_loss_weight": 0.0,
|
| 26 |
+
"moe_experts_per_layer": 0,
|
| 27 |
+
"moe_expert_hidden": 0,
|
| 28 |
+
"moe_hot_experts": 0,
|
| 29 |
+
"moe_capacity_factor": 1.5,
|
| 30 |
+
"moe_seed": 0,
|
| 31 |
+
"max_loops": 16
|
| 32 |
+
}
|