Frank
accf90b6fa
GGUF block scan (2026-08-17): FA at 3,7,...,63; linear blocks 383 MiB, FA blocks 446 MiB. 35 blocks (0-34) keep 8 FA layers on RPC0 -> KV stays 8192 MiB, predicted ~0.8 GiB free. Block 35 is FA: +1 GiB KV -> OOM. Measured at 0.50: 22650/24576 used, 1477 free (the earlier 714 MiB projection was conservative).
82 lines
3.0 KiB
Bash
82 lines
3.0 KiB
Bash
# Head tunables — Qwen3.8-27B pipeline-parallel head (Dankclaw).
|
|
#
|
|
# The head is the LOCAL side of the split. It runs llama-server (b10433,
|
|
# HIP+Vulkan+RPC build in /home/dank/llama.cpp/build) with --rpc pointing at
|
|
# the worker on Dankserver. The model file lives on this machine only; the
|
|
# worker receives its tensor share over the network at load time.
|
|
#
|
|
# CRITICAL: this script must NOT pass --device. When --device is set,
|
|
# llama.cpp uses exactly that device list and silently drops the RPC device
|
|
# (src/llama.cpp llama_prepare_model_devices). Default device selection
|
|
# inserts RPC servers at the FRONT of the device list, so the split becomes
|
|
# [RPC0, iGPU] and --tensor-split index 0 is the worker.
|
|
#
|
|
# WARNING: do not name any variable here TEMP, HOST, or GPU_MAX_HEAP_SIZE
|
|
# (HIP runtime collision, see the 2026-08-13 TEMP incident).
|
|
|
|
MODEL_DIR=/home/dank/models
|
|
MODEL_FILE=Qwen3.8-27B-UD-Q8_K_XL.gguf
|
|
HEAD_BIN=/home/dank/llama.cpp/build/bin/llama-server
|
|
ALIAS=frank-default-large
|
|
BIND_ADDR=0.0.0.0
|
|
LISTEN_PORT=7770
|
|
|
|
# Comma-separated RPC endpoints (host:port). One worker here.
|
|
RPC_SERVERS=192.168.0.235:50052
|
|
|
|
# Model split fractions, one per device in order [RPC0, iGPU].
|
|
# 0.55,0.45 = 35 of 65 blocks on the 3090 (blocks 0-34) - the maximum
|
|
# that fits its 24 GB card at the 262144 f16 KV pool. VRAM math from the
|
|
# live load (2026-08-17): RPC0 buffer ~14.4 GiB + KV 8192 MiB (8 full-
|
|
# attention layers) + RS/compute/overhead -> ~0.8 GiB free. Block 35 is a
|
|
# full-attention layer: adding it costs +1 GiB KV -> OOM. GGUF block scan:
|
|
# FA at 3,7,...,63 (every 4th, offset 3).
|
|
TENSOR_SPLIT=0.55,0.45
|
|
|
|
# Context pool. 262144 total = 1 parallel slot x 262144 per slot
|
|
# (n_ctx_seq = n_ctx / n_seq_max, llama-context.cpp:293).
|
|
CTX_SIZE=262144
|
|
PARALLEL=1
|
|
|
|
# Speculative decoding. Matches production frank-default-large: MTP (bundled
|
|
# head, n_max 4 per the published tuning curve) with the ngram-mod fallback.
|
|
# Set SPEC_NONE=1 for a raw baseline with no speculation.
|
|
SPEC_NONE=0
|
|
SPEC_DEFAULT=1
|
|
MTP=1
|
|
SPEC_DRAFT_N_MAX=4
|
|
# Optional standalone MTP draft head (e.g. the Q4_0 ggml-org file). When set,
|
|
# it replaces the bundled in-model head via --spec-draft-model.
|
|
SPEC_DRAFT_MODEL=
|
|
SPEC_NGRAM_MOD_N_MIN=48
|
|
SPEC_NGRAM_MOD_N_MAX=64
|
|
SPEC_NGRAM_MOD_N_MATCH=24
|
|
|
|
# KV on GPU, unquantized (f16). 16 full-attention layers x 2 (K,V) x 4 heads
|
|
# x 256 dim x 2 bytes = 64 KiB/token -> 262144 tokens = 16.6 GiB total,
|
|
# held in proportion to the full-attention layers each device holds.
|
|
# Keep K == V so Flash Attention stays enabled.
|
|
CACHE_TYPE_K=f16
|
|
CACHE_TYPE_V=f16
|
|
|
|
# Batch sizing: production-verified values (2026-08-14).
|
|
BATCH_SIZE=8192
|
|
UBATCH_SIZE=256
|
|
|
|
# Sampling — Qwen3.8 thinking-mode defaults.
|
|
TEMPERATURE=1.0
|
|
TOP_K=20
|
|
TOP_P=0.95
|
|
MIN_P=0.0
|
|
|
|
THREADS=12
|
|
THREADS_HTTP=1
|
|
|
|
# 4 (info): clean journal for interactive use. Raise to 10 only when
|
|
# debugging the layer split (the per-layer assignment lines are debug-level).
|
|
VERBOSITY=4
|
|
|
|
# 1: --direct-io model load. Required in practice for large models loaded
|
|
# over RPC (load hangs otherwise; visorcraft Strix RPC finding).
|
|
DIRECT_IO=1
|