# Tunables for frank-default-small (HauhauCS Q6_K_P, ngram-mod spec, dankserver)
# Runtime: llama-server b10108; RTX 3090 24GB; Ryzen 7 7700X; 62GB RAM.
MODEL_DIR=/home/dank/models
ALIAS=frank-default-small
HOST=0.0.0.0
PORT=7770
# 4 parallel slots x 200000 ctx each = 800000 total context pool.
# (Increased from 3×200k to add a 4th parallel stream.)
CTX_SIZE=800000
PARALLEL=4
# Spec: ngram-mod. n_min=48, n_max=64, n_match=24 (defaults). 8 expert
# layers back on GPU; rest on CPU. Batch 8192 / ubatch 2048.
SPEC_TYPE=ngram-mod
NGRAM_MOD_N_MIN=48
NGRAM_MOD_N_MAX=64
NGRAM_MOD_N_MATCH=24
# Expert offload: 33 of 40 layers keep MoE weights on CPU; remaining
# 7 layers have their expert weights resident on GPU for faster routing.
N_CPU_MOE=33
N_GPU_LAYERS=all
# KV cache on GPU, K quantized to q8_0, V stays f16 (user-locked decision).
CACHE_TYPE_K=q8_0
CACHE_TYPE_V=f16
# Store SWA checkpoints in RAM to hold peak GPU memory down for the
# larger batch/ubatch (leaves headroom for ctx-checkpoints operation).
CTX_CHECKPOINTS=4
# Larger batches / ubatch for big prefill workloads.
BATCH_SIZE=8192
UBATCH_SIZE=2048
# Load the model into RAM rather than lazy-mapping it. The MoE experts live on
# CPU, so eager RAM load avoids page faults during decode (uses ~31GB RAM).
NO_MMAP=1
# Compute threads lowered to 7 so other processes on the box keep a core.
THREADS=7
THREADS_HTTP=8