Increase `sliding_window` to 32k

Some backends, such as vLLM, do not dynamically adjust the sliding window param based on the max context length provided. Modifying it here seems to fix that.

config.json

@@ -20,7 +20,7 @@
   "rms_norm_eps": 1e-05,
   "rope_theta": 1000000.0,
   "router_aux_loss_coef": 0.02,
-  "sliding_window": 4096,
+  "sliding_window": 32768,
   "tie_word_embeddings": false,
   "torch_dtype": "bfloat16",
   "transformers_version": "4.36.0.dev0",