Frank
24c5eb4e28
K cache back to default f16 full precision with K_CACHE_HADAMARD=0; V cache at q6_0 with V_CACHE_HADAMARD=1. Env, serve script, bats, and KB updated.
110 lines
4.9 KiB
Bash
110 lines
4.9 KiB
Bash
# Tunables for frank-default-small (Ornith-1.5-35B-A3B CRACK Q8_0, MTP MoE,
|
||
# Dankserver, RTX 3090).
|
||
#
|
||
# Runtime: ik_llama.cpp llama-server (ikawrakow ik_llama.cpp, built
|
||
# 2026-08-26 from commit b166e269) on RTX 3090 24 GB; Ryzen 7 7700X; 62 GB
|
||
# RAM. Previous runtime was upstream llama.cpp b10121. MoE expert weights
|
||
# pinned to CPU with an explicit --override-tensor regex (not --n-cpu-moe).
|
||
# KV cache K q8_0 / V q5_0 with Hadamard transforms on both (see the KV
|
||
# section below). Split mode none (single GPU; graph would require 2+).
|
||
#
|
||
# Build flags (IK_FA note, see docs/LLM/ikllama-ornith-dankserver.md) must
|
||
# ALWAYS include: -DGGML_CUDA_FA_ALL_QUANTS=ON and -DGGML_IQK_FA_ALL_QUANTS=ON
|
||
# so the quantized + Hadamard KV cache combos compile into the CUDA/CPU FA
|
||
# kernels. Replacing the HauhauCS Qwen3.6-35B-A3B model on 2026-08-24.
|
||
# Switched to CRACK Q8_0 on 2026-08-26 (was Q6_K).
|
||
#
|
||
# WARNING: do not name any variable TEMP, HOST, or GPU_MAX_HEAP_SIZE. The
|
||
# ROCm HIP runtime reads those bare env vars (see the 2026-08-13 TEMP
|
||
# incident). This box is NVIDIA/CUDA, but the rule keeps the env portable.
|
||
|
||
# Model files (Q8_0 CRACK GGUF + f16 mmproj vision encoder + MTP head bundled
|
||
# in the GGUF). Dealign2026-08-24.
|
||
MODEL_DIR=/home/dank/models/Ornith-1.5-35B-A3B-CRACK
|
||
MODEL_FILE=Ornith-1.5-35B-A3B-CRACK-Q8_0.gguf
|
||
MMPROJ_FILE=mmproj-Ornith-1.5-35B-A3B-f16.gguf
|
||
# ik_llama.cpp llama-server (built 2026-08-26 from ikawrakow/ik_llama.cpp @
|
||
# b166e269; CUDA 13.2, -DGGML_CUDA_FA_ALL_QUANTS=ON -DGGML_IQK_FA_ALL_QUANTS=ON,
|
||
# arch 86 for the RTX 3090). Uses the staged pip cu13 toolkit at
|
||
# /home/dank/cuda13-stage. Rollback: switch to
|
||
# /home/dank/llama.cpp/build/bin/llama-server (upstream b10121).
|
||
LLAMA_SERVER=/home/dank/ik_llama.cpp/build/bin/llama-server
|
||
|
||
# Qwen-Sharp chat template from peculiar-ragdoll (tracked in the repo,
|
||
# staged on the host).
|
||
# Source: https://huggingface.co/peculiar-ragdoll/Qwen-Sharp-Chat-Templates
|
||
CHAT_TEMPLATE_FILE=/home/dank/serverAppConfigs/scripts/chat-template.jinja
|
||
# Render the Qwen-Sharp template with the true Jinja engine (--jinja). Without
|
||
# it ik_llama falls back to built-in non-Jinja handlers and drops the custom
|
||
# template's thinking/image/tool formatting.
|
||
USE_JINJA=1
|
||
|
||
ALIAS=frank-default-small
|
||
BIND_ADDR=0.0.0.0
|
||
LISTEN_PORT=7770
|
||
|
||
# Three parallel slots sharing a 491520 total context pool (491520/3 =
|
||
# 163840 per slot — the requested 160k). With PARALLEL=3 ik_llama limits
|
||
# speculative decoding to a single slot, so SPEC stays none (MTP is also
|
||
# disabled). The 3 slots + 480k pool make the KV cache large, so expert
|
||
# weights are CPU-pinned for the cache to fit on the 24 GB card.
|
||
CTX_SIZE=491520
|
||
PARALLEL=3
|
||
|
||
# Speculative decoding: DISABLED. ik_llama limits spec to a single slot; at
|
||
# PARALLEL=3 a configured spec stage hard-exits at load. The low-parameter
|
||
# MoE (3B active of 35B) does not need it anyway. MTP is also disabled.
|
||
SPEC_TYPE=none
|
||
|
||
# Expert offload via explicit tensor override (replaces --n-cpu-moe). The Q8_0
|
||
# audit found ~1088 MiB of expert tensors per block (4 tensor × 272 MiB). Keep
|
||
# every final/output-adjacent block 35-40 on CUDA; full-attention blocks 3,7,
|
||
# 11,15,19,23 have no exps and are untouched by this regex. All other expert
|
||
# tensors are pinned to CPU. Buffer type "CPU" is the CPU buft on this host.
|
||
#
|
||
# 2026-08-26: PARALLEL=3 (CTX 160000) triples the KV cache, so the blocks
|
||
# that were moved back to CUDA for fused-MoE (27-34) are pinned to CPU again
|
||
# to free GPU. Blocks 35-40 stay on CUDA. If it boot-loops, move the next
|
||
# lower block into the CPU regex one at a time.
|
||
OVERRIDE_EXPERT='blk\.(0|1|2|4|5|6|8|9|10|12|13|14|16|17|18|20|21|22|24|25|26|27|28|29|30|31|32|33|34)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
|
||
OVERRIDE_BUFT=CPU
|
||
# ik_llama --gpu-layers/-ngl does NOT accept the literal "all"; use 999 to
|
||
# mean "load as many layers as possible". Expert exps are pinned to CPU by
|
||
# OVERRIDE_EXPERT below.
|
||
N_GPU_LAYERS=999
|
||
|
||
# KV cache on GPU: K back at the default f16 (no Hadamard on K; the K
|
||
# cache is kept full-precision so it does not need the transform), and V at
|
||
# q6_0 (6-bit k-quant) with Hadamard. Flash-attention is on by default.
|
||
CACHE_TYPE_K=f16
|
||
CACHE_TYPE_V=q6_0
|
||
K_CACHE_HADAMARD=0
|
||
V_CACHE_HADAMARD=1
|
||
|
||
# Split mode: none. Dankserver has a single RTX 3090 and ik_llama requires
|
||
# >= 2 devices for split-mode graph ('It is not possible to use split mode
|
||
# graph with less than 2 devices' is a hard load error). none is correct
|
||
# here and still works with OVERRIDE_EXPERT CPU expert offload.
|
||
SPLIT_MODE=none
|
||
|
||
VERBOSITY=4
|
||
# Store context checkpoints in RAM to hold peak GPU memory down.
|
||
CTX_CHECKPOINTS=16
|
||
|
||
# Prefill batch sizing. ubatch 1024 is selected for reduced compute-memory
|
||
# pressure and predictable latency under the ngram-mod deployment.
|
||
BATCH_SIZE=4096
|
||
UBATCH_SIZE=1024
|
||
|
||
# Eager RAM load (expert weights on CPU; ~31 GB RAM).
|
||
NO_MMAP=1
|
||
|
||
# Sampling (Dealign Ornith model card, thinking-general).
|
||
TEMPERATURE=1.0
|
||
TOP_K=20
|
||
TOP_P=0.95
|
||
MIN_P=0.0
|
||
|
||
# Compute threads lowered to 8 so other processes keep a core.
|
||
THREADS=7
|
||
THREADS_HTTP=8 |