Frank
a27839519e
n_max 2 benched 12.4-14.3 t/s (slower than n_max 3 at 15.3-19.5); n_max 3 is the MTP sweet spot. ubatch 8192 (==batch) uses the VRAM headroom (65.5 GiB of 96 GiB) safely. MTP beats ngram-mod (~7 t/s) and matches DFlash (~20 t/s) without the parallel premature-EOS bug.
55 lines
2.5 KiB
Bash
55 lines
2.5 KiB
Bash
# Tunables for frank-default-large (HauhauCS Qwen3.6-27B Q8_K_P, DFlash+ngram-mod, dankclaw)
|
|
# Runtime: llama.cpp llama-server; AMD Strix Halo (gfx1151) 96GB UMA; Ryzen AI MAX+ 395.
|
|
MODEL_DIR=/home/dank/models
|
|
ALIAS=frank-default-large
|
|
HOST=0.0.0.0
|
|
PORT=7770
|
|
# 3 parallel slots x 200700 ctx each = 602100 total context pool.
|
|
CTX_SIZE=602100
|
|
PARALLEL=3
|
|
# Spec: speculative decoding mode.
|
|
# draft-dflash - DFlash block-diffusion draft (benchmark winner pre-2026-08-14)
|
|
# BUT: premature EOS under parallel load when a new task
|
|
# reuses a cached prompt (shared draft ctx corruption). Retired.
|
|
# draft-mtp - MTP sidecar draft (a4lg/Qwen3.6-27B-MTP-ONLY-Q8_0.gguf)
|
|
# ngram-mod - draftless ngram matching (no draft model, no shared ctx)
|
|
# Benched 2026-08-14: see docs/frank-default-large-benchmark.md for the winner.
|
|
SPEC_TYPE=draft-mtp
|
|
# Draft model file (only used when SPEC_TYPE contains 'draft-').
|
|
# DFlash: Qwen3.6-27B-DFlash-Q8_0.gguf ; MTP: Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
|
|
DRAFT_FILE=Qwen3.6-27B-MTP-ONLY-Q8_0.gguf
|
|
# Draft length per spec type: dflash 8, mtp 2-3.
|
|
SPEC_DRAFT_N_MAX=3
|
|
SPEC_NGRAM_MOD_N_MIN=48
|
|
SPEC_NGRAM_MOD_N_MAX=64
|
|
SPEC_NGRAM_MOD_N_MATCH=24
|
|
# Backend device from --list-devices. ROCm wins on generation; Vulkan renders
|
|
# DFlash spec unusable (NONE-op abort) and gen is slower on ngram (6.9 t/s).
|
|
DEVICE=ROCm0
|
|
# Dense model (Qwen3.6-27B, not MoE): no MoE CPU offload. Full offload.
|
|
N_GPU_LAYERS=all
|
|
# KV on GPU, quantized (K q8_0, V q8_0). V MUST match K: this llama.cpp
|
|
# build has GGML_CUDA_FA_ALL_QUANTS=OFF, and Flash Attention requires
|
|
# K type == V type. Mixed q8_0/f16 silently disables FA, collapsing
|
|
# long-context decode to ~1 t/s at 200k ctx. q8_0 V halves V memory too.
|
|
CACHE_TYPE_K=q8_0
|
|
CACHE_TYPE_V=q8_0
|
|
# Batch sizing: empirically verified 2026-08-13 with caches dropped.
|
|
# batch 8192 / ubatch 2048 loads and serves (61 GiB VRAM peak, ctx 602100).
|
|
# ubatch is the compute-graph driver; larger ubatch = more VRAM, faster prefill.
|
|
# Re-verified 2026-08-14 during the spec swap (see benchmark doc).
|
|
BATCH_SIZE=8192
|
|
UBATCH_SIZE=8192
|
|
# Sampling settings — per HauhauCS Qwen3.6-27B model card.
|
|
# Thinking-general: temp=1.0, top_p=0.95, top_k=20, min_p=0.
|
|
# Clients may override per-request; these are the server-side defaults.
|
|
# WARNING: do not name this variable TEMP. The ROCm HIP runtime reads the
|
|
# TEMP environment variable (JIT kernel-cache dir); TEMP=1.0 made the HIP
|
|
# runtime SEGV (NULL unref) during model load on Aug 13 2026.
|
|
TEMPERATURE=1.0
|
|
TOP_K=20
|
|
TOP_P=0.95
|
|
MIN_P=0.0
|
|
THREADS=12
|
|
THREADS_HTTP=1
|