Qwen
/

Text Generation
Transformers
Safetensors
qwen3_moe
conversational
nbroad commited on
Commit
d854eb9
·
verified ·
1 Parent(s): c30ce1a

yarn scale to 122,880 context length

Browse files

Please don't merge or close this. I'm just going to use this pr revision to run the model at 122k sequence length

Files changed (1) hide show
  1. config.json +5 -1
config.json CHANGED
@@ -25,7 +25,11 @@
25
  "num_key_value_heads": 4,
26
  "output_router_logits": false,
27
  "rms_norm_eps": 1e-06,
28
- "rope_scaling": null,
 
 
 
 
29
  "rope_theta": 1000000.0,
30
  "router_aux_loss_coef": 0.001,
31
  "sliding_window": null,
 
25
  "num_key_value_heads": 4,
26
  "output_router_logits": false,
27
  "rms_norm_eps": 1e-06,
28
+ "rope_scaling": {
29
+ "rope_type": "yarn",
30
+ "factor": 3.0,
31
+ "original_max_position_embeddings": 40960
32
+ },
33
  "rope_theta": 1000000.0,
34
  "router_aux_loss_coef": 0.001,
35
  "sliding_window": null,