Frank
fd528ab364
- spec-type → ngram-mod (no separate draft model) - drop --model-draft, --spec-draft-n-max; add --spec-ngram-mod-n-min/max/match - DFlash batch default fallback SPEC_TYPE → ngram-mod - env: SPEC_TYPE=ngram-mod, NGRAM_MOD_N_MIN=48, N_MAX=64, N_MATCH=24 - DRAFT_FILE env var emptied (not used under ngram-mod)
34 lines
1.3 KiB
Bash
34 lines
1.3 KiB
Bash
# 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
|
|
# 3 parallel slots x 200000 ctx each = 600000 total context pool.
|
|
# (Reduced from 262144/stream per operator: allows the larger batch to fit VRAM.)
|
|
CTX_SIZE=600000
|
|
PARALLEL=3
|
|
SPEC_TYPE=ngram-mod
|
|
# Ngram-mod speculative decoding: look back this many tokens to guess the
|
|
# next tokens. Defaults (48/64/24) are a reasonable starting point; tune if
|
|
# you want more or fewer draft lookahead tokens.
|
|
NGRAM_MOD_N_MIN=48
|
|
NGRAM_MOD_N_MAX=64
|
|
NGRAM_MOD_N_MATCH=24
|
|
CACHE_TYPE_K=q8_0
|
|
CACHE_TYPE_V=f16
|
|
# All 40 layers keep their MoE expert weights on CPU (targeted expert offload).
|
|
N_CPU_MOE=40
|
|
# Every non-expert tensor (attention, norms, embedding, LM head) goes to GPU.
|
|
N_GPU_LAYERS=all
|
|
# Larger batches for prefill throughput. The bigger physical batch grows GPU
|
|
# activation buffers; ctx-checkpoints parks attention checkpoints in RAM to
|
|
# hold peak GPU memory down so the larger ubatch fits without OOM.
|
|
BATCH_SIZE=3072
|
|
UBATCH_SIZE=1536
|
|
CTX_CHECKPOINTS=4
|
|
# 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
|
|
THREADS=8
|
|
THREADS_HTTP=8 |