Experimental 128k context with rope scaling

config.json

@@ -11,14 +11,20 @@
   "hidden_size": 4096,
   "initializer_range": 0.02,
   "intermediate_size": 14336,
-  "max_position_embeddings": 65536,
+  "max_position_embeddings": 131072,
   "model_type": "llama",
   "num_attention_heads": 32,
   "num_hidden_layers": 32,
   "num_key_value_heads": 8,
   "pretraining_tp": 1,
   "rms_norm_eps": 1e-05,
-  "rope_scaling": null,
+  "rope_scaling": {
+    "factor": 4.0,
+    "low_freq_factor": 1.0,
+    "high_freq_factor": 4.0,
+    "original_max_position_embeddings": 32768,
+    "rope_type": "llama3"
+  },
   "rope_theta": 1500000.0,
   "tie_word_embeddings": false,
   "torch_dtype": "bfloat16",