81 lines
2.5 KiB
Text
81 lines
2.5 KiB
Text
|
# llama.cpp.config.example
|
||
|
|
#
|
||
|
|
# Copy this file to llama.cpp.config and adjust it for your host before
|
||
|
|
# running llamacppctl. See docs/SECURITY_AND_OPERATIONS.md for details on
|
||
|
|
# every field and on the security model for --system-url/--prompt-url.
|
||
|
|
#
|
||
|
|
# Section types:
|
||
|
|
# [default] global defaults, merged over the built-in fallbacks
|
||
|
|
# [model.<name>] a model profile, selected via --profile <name>
|
||
|
|
# [prompt.<name>] a system-prompt profile, selected via --system-prompt-profile <name>
|
||
|
|
#
|
||
|
|
# container_name is MANDATORY and must be unique per profile you intend to
|
||
|
|
# run concurrently: it is the single identity anchor for Docker naming,
|
||
|
|
# locking (--change), and --stop/--check targeting.
|
||
|
|
|
||
|
|
[default]
|
||
|
|
image = ghcr.io/ggml-org/llama.cpp:server-cuda
|
||
|
|
hf_home = ${HF_HOME}
|
||
|
|
model_path = models/qwen3/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf
|
||
|
|
container_name = llama_cpp_server
|
||
|
|
host_port = 8001
|
||
|
|
container_port = 8000
|
||
|
|
model_alias = default_llm
|
||
|
|
gpu_device = 1
|
||
|
|
restart_policy = unless-stopped
|
||
|
|
# Groß-Profil für literarische Texte / durchdachte Reden: volles
|
||
|
|
# 256k-Kontextfenster (= n_ctx_train des Modells) und sehr lange Antworten.
|
||
|
|
# Empirisch gemessen passt der q4_0-KV-Cache dafür knapp auf EINE 24-GB-3090
|
||
|
|
# (~22,6 GB inkl. 21-GB-Modell). Bei Bedarf auf 131072/65536 senken.
|
||
|
|
ctx_size = 262144
|
||
|
|
n_predict = 32768
|
||
|
|
temp = 0.65
|
||
|
|
top_p = 0.80
|
||
|
|
top_k = 20
|
||
|
|
min_p = 0.01
|
||
|
|
repeat_penalty = 1.05
|
||
|
|
main_gpu = 0
|
||
|
|
ngl = 999
|
||
|
|
fa = true
|
||
|
|
kv_unified = true
|
||
|
|
jinja = true
|
||
|
|
reasoning = on
|
||
|
|
no_context_shift = true
|
||
|
|
cache_type_k = q4_0
|
||
|
|
cache_type_v = q4_0
|
||
|
|
batch_size = 1024
|
||
|
|
ubatch_size = 512
|
||
|
|
parallel = 1
|
||
|
|
cont_batching = true
|
||
|
|
health_endpoint = /health
|
||
|
|
models_endpoint = /v1/models
|
||
|
|
chat_endpoint = /v1/chat/completions
|
||
|
|
timeout = 300
|
||
|
|
poll_interval = 2
|
||
|
|
# Chat-Antwortbudget (Reasoning-Modell braucht viel; für lange Texte hoch):
|
||
|
|
max_tokens = 32768
|
||
|
|
# chat_temperature leer lassen -> Server-Temp (temp oben) gilt.
|
||
|
|
chat_temperature =
|
||
|
|
|
||
|
|
# Example second model profile, pinned to the second GPU (e.g. RTX 3090 #2)
|
||
|
|
# with its own port and container name so it can run alongside [default].
|
||
|
|
[model.qwen35]
|
||
|
|
model_path = qwen3/Qwen3-35B-A3B-Q4_K_M.gguf
|
||
|
|
model_alias = qwen35
|
||
|
|
container_name = llama_cpp_qwen35
|
||
|
|
host_port = 8002
|
||
|
|
gpu_device = 1
|
||
|
|
|
||
|
|
[model.deepseek]
|
||
|
|
model_path = deepseek/deepseek-r1-q4.gguf
|
||
|
|
model_alias = deepseek
|
||
|
|
container_name = llama_cpp_deepseek
|
||
|
|
host_port = 8003
|
||
|
|
gpu_device = 1
|
||
|
|
ctx_size = 131072
|
||
|
|
|
||
|
|
[prompt.concise]
|
||
|
|
system_prompt = Du antwortest kurz, präzise und technisch.
|
||
|
|
|
||
|
|
[prompt.coding]
|
||
|
|
system_prompt = Du bist ein erfahrener Linux-, Python- und LLM-Infrastruktur-Engineer.
|