Source files / install · raw

tersler.local.conf.example

# Copy to install/tersler.local.conf (git-ignored) and adjust.
# Everything here overrides install/tersler.conf.

# Direct download link to the GGUF (must include the MTP head):
# MODEL_URL="https://huggingface.co/<owner>/<repo>/resolve/main/<file>.gguf"
# MODEL_SHA256=""
# MODEL_HF_REPO="<owner>/<repo>"

# Serve on the LAN (llama-server has no auth by default!):
# LLAMA_HOST=0.0.0.0

# Try a newer fork commit without editing the main config:
# LLAMA_COMMIT=<sha>

# Fan control, after `tersler-fan-control --discover`:
# FAN_CONTROL=1
# FAN_HWMON_NAME=nct6779
# FAN_PWM=pwm2

# ---- Alternative model: Nanbeige4.3-33B (dense Llama architecture, no MTP head) ----
# Untested on the rig. Uses K-quants (the fork has P100 matvec paths for Q4_K/Q5_K;
# IQ quants run on the generic path with emulated DP4A). See docs/research.md.
# MODEL_URL="https://huggingface.co/bartowski/Nanbeige_Nanbeige4.3-33B-GGUF/resolve/main/Nanbeige_Nanbeige4.3-33B-Q4_K_M.gguf"
# MODEL_HF_REPO="bartowski/Nanbeige_Nanbeige4.3-33B-GGUF"
# LLAMA_ENV=(GGML_CUDA_GRAPHS_PRE_VOLTA=3)     # the LLAMA_SPEC_* variables only matter with MTP
# LLAMA_ARGS=(
#   -ngl 99 -sm tensor -fa 1 -ctk q8_0 -ctv q8_0
#   -c 32768 -ub 512 -np 1                     # KV q8_0 ≈ 110 KB/token; try 65536 once it runs
#   --jinja --chat-template-kwargs '{"preserve_thinking":true}'
#   --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0
# )