v1.78 Context size is bigger then configured

[PeterPeet]

If I set the context size to any value (doesn't matter if I use GUI or manual) then n_ctx gets bigger then the configured value. This is a new behaviour in my opinion since I updated. The difference is always 256. I used 1.76 before.

Is just the warning new or does this get calculated/set in the wrong way? My Model is optimized for 8192, but I use it in a trained version that gives me 16384. I tried other models they behave the same.

Here is an example for context_size: 16384, but it happens with every value I set.

Automatic RoPE Scaling: Using (scale:1.000, base:10000.0).
llama_new_context_with_model: n_seq_max = 1
llama_new_context_with_model: n_ctx = 16640
llama_new_context_with_model: n_ctx_per_seq = 16640
llama_new_context_with_model: n_batch = 512
llama_new_context_with_model: n_ubatch = 512
llama_new_context_with_model: flash_attn = 1
llama_new_context_with_model: freq_base = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_new_context_with_model: n_ctx_pre_seq (16640) > n_ctx_train (16384) -- possible training context overflow
llama_kv_cache_init: CUDA0 KV buffer size = 2080.00 MiB
llama_new_context_with_model: KV self size = 2080.00 MiB, K (f16): 1040.00 MiB, V (f16): 1040.00 MiB
llama_new_context_with_model: CUDA_Host output buffer size = 0.12 MiB
llama_new_context_with_model: CUDA0 compute buffer size = 108.75 MiB

My configuration for this example is:
{"model": "", "model_param": "D:/SillyTavern/Modelle/Text-Creation/daybreak-kunoichi-2dpo-7b-q4_k_m.gguf", "port": 5001, "port_param": 5001, "host": "", "launch": false, "config": null, "threads": 7, "usecublas": ["normal", "0", "mmq"], "usevulkan": null, "useclblast": null, "usecpu": false, "contextsize": 16384, "gpulayers": 35, "tensor_split": null, "ropeconfig": [0.0, 10000.0], "blasbatchsize": 512, "blasthreads": null, "lora": null, "noshift": false, "nofastforward": false, "nommap": false, "usemlock": false, "noavx2": false, "debugmode": 0, "onready": "", "benchmark": null, "prompt": "", "promptlimit": 100, "multiuser": 1, "remotetunnel": false, "highpriority": false, "foreground": false, "preloadstory": null, "quiet": false, "ssl": null, "nocertify": false, "mmproj": null, "password": null, "ignoremissing": false, "chatcompletionsadapter": null, "flashattention": true, "quantkv": 0, "forceversion": 0, "smartcontext": false, "unpack": "", "nomodel": false, "showgui": false, "skiplauncher": false, "hordemodelname": "", "hordeworkername": "", "hordekey": "", "hordemaxctx": 0, "hordegenlen": 0, "sdmodel": "", "sdthreads": 7, "sdclamped": 0, "sdt5xxl": "", "sdclipl": "", "sdclipg": "", "sdvae": "", "sdvaeauto": false, "sdquant": false, "sdlora": "", "sdloramult": 1.0, "whispermodel": "", "hordeconfig": null, "sdconfig": null, "noblas": false}
llama_new_context_with_model: CUDA_Host compute buffer size = 40.51 MiB
llama_new_context_with_model: graph nodes = 903
llama_new_context_with_model: graph splits = 2

[askmyteapot]

Has always done that. Its just the warning that's new.
The additional 256 context is for buffer i believe. Won't actually impact usage, as you'd still only be requesting 16384 max context when actually interacting with the model.

Safe to ignore in this circumstance.