Frank
d9af7f1d19
- Move all previous scripts to old/scripts/ (retired PP/NPU-era) - frank-default-large: llama.cpp FastMTP on dankserver (HauhauCS Qwen3.8-27B IQ4_XS + FastMTP 32K draft + ngram-mod, K q8_0/V f16, 128K ctx, batch 2048/ubatch 1024, port 7770) - frank-default-small: vLLM ROCm on dankclaw (Avesed Qwen3.6-35B-A3B INT8 W8A8, full 262144 ctx, no KV quant, 4 seqs, batch 16384, ngram spec, port 52625) - retarget deploy.sh (dankserver/dankclaw, inline sudo, no helpers) - docs: deployment + index + AGENTS.md + README for new topology, superseded banners on retired docs, add frank-redo-2026-08-17 runbook
48 lines
2.0 KiB
Bash
48 lines
2.0 KiB
Bash
# Tunables for the Ling-3.0-flash TRAIL on dankclaw (frank-large-ling).
|
|
# This REPLACES frank-default-large (Qwen3.6-27B) temporarily. To switch
|
|
# back: systemctl stop frank-large-ling && systemctl start frank-default-large.
|
|
# Runtime: llama.cpp FORK bailingmoe3-support; Vulkan + AMD GTT.
|
|
#
|
|
# Memory model (amdgpu Strix Halo gfx1151, 8x16GB LPDDR5 = 128 GiB physical):
|
|
# - firmware UMA carve-out gives the iGPU ~96 GiB VRAM (rocm-smi "VRAM total").
|
|
# - CPU system RAM is the remainder (~30 GiB; kernel MemTotal).
|
|
# - Vulkan (RADV) reports an ~111.2 GiB heap (113910 MiB) - larger than ROCm's
|
|
# 96 GiB, because Vulkan can also reach system memory via GTT.
|
|
# - Q6_K is 98.26 GiB, leaving only ~13 GiB of heap for KV + graphs, so keep
|
|
# the context small (Ling uses MLA/KDA, so KV stays small). Tune CTX_SIZE.
|
|
MODEL_DIR=/home/dank/models
|
|
MODEL_FILE=Ling-3.0-flash-Q6_K.gguf
|
|
ALIAS=frank-default-large # keep the API alias so clients are unaffected
|
|
HOST=0.0.0.0
|
|
PORT=7770
|
|
# Context: Q6_K uses ~98.7 GiB of the ~111.2 GiB Vulkan heap; Ling's MLA/KDA
|
|
# KV is tiny (~2.5 GiB for 256k), so a large context fits. 262144 total = 2
|
|
# slots x 131072, ~10 GiB heap headroom after load. Verified 2026-08-14.
|
|
CTX_SIZE=262144
|
|
PARALLEL=2
|
|
# MTP is bundled in the Ling GGUF itself (block 42) - no separate draft file.
|
|
# Layering ngram-mod under MTP (draft-mtp,ngram-mod) re-drafts via ngram.
|
|
SPEC_TYPE=none
|
|
SPEC_DRAFT_N_MAX=3
|
|
SPEC_NGRAM_MOD_N_MIN=48
|
|
SPEC_NGRAM_MOD_N_MAX=64
|
|
SPEC_NGRAM_MOD_N_MATCH=24
|
|
# Vulkan exposes the larger heap (UMA + GTT) vs ROCm. Use Vulkan0.
|
|
DEVICE=Vulkan0
|
|
N_GPU_LAYERS=all
|
|
FLASH_ATTN=auto
|
|
# Batch kept small-leaning for the tight memory budget; tune with CTX_SIZE.
|
|
BATCH_SIZE=3072
|
|
UBATCH_SIZE=3072
|
|
|
|
# Diagnostics: high verbosity + a dedicated log file for investigating the
|
|
# premature 'stop' reports. -lv 4 = verbose (ggml); 5 = debug.
|
|
VERBOSITY=0
|
|
# Sampling defaults (model card: reasoning-capable; leave temp/top to default)
|
|
TEMPERATURE=1.0
|
|
TOP_K=20
|
|
TOP_P=0.95
|
|
MIN_P=0.0
|
|
THREADS=12
|
|
THREADS_HTTP=1
|