version = 1

; ============================================================================
; llama.cpp router presets — Strix Halo (Radeon 8060S, 126 GiB unified VRAM)
; ============================================================================

; ---------------------------------------------------------------------------
; Global defaults shared by every model instance (overridable per-model)
; ---------------------------------------------------------------------------
[*]
n-gpu-layers  = all
flash-attn    = on
jinja         = true
metrics       = true ; Prometheus /metrics endpoint (per-model: /metrics?model=...)

; ---------------------------------------------------------------------------
; Qwen3.6-35B-A3B — Alibaba, MoE 3B-active + multimodal. Fast all-rounder.
; ---------------------------------------------------------------------------
[qwen3.6-35b-a3b]
load-on-startup    = true
model              = /home/spencer/.cache/models/Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf
mmproj             = /home/spencer/.cache/models/Qwen3.6-35B-A3B-mmproj-BF16.gguf
spec-type          = draft-mtp
spec-draft-n-max   = 3
ubatch-size        = 2048
chat-template-file = /home/spencer/.config/llama.cpp/qwen-fixed-chat_template.jinja
reasoning-format   = deepseek
reasoning          = auto
reasoning-preserve = true
temperature        = 1.0
top-p              = 0.95
top-k              = 20
min-p              = 0.0
presence-penalty   = 1.5
repeat-penalty     = 1.0

; ---------------------------------------------------------------------------
; Qwen3.8-Flash-Next — Qwen4-preview arch (qwen4exp), 125B total / 6B active.
;
; Speculative decoding: DRAFT-MTP (unsloth qwen4exp/mtp fork build provides it; upstream PR
; ggml-org/llama.cpp#27836 still can't load head-only MTP GGUFs). Head = unsloth shared-Q4_K_M.
; n-gram/PLE table is lazy-from-disk in the fork loader (TENSOR_READ_LAZY, no flag needed).
; ---------------------------------------------------------------------------
[qwen3.8-flash-next]
; unsloth UD-Q4_K_XL + MTP head (their qwen4exp/mtp fork build — REQUIRED for MTP).
; n-gram/PLE table is served from disk lazily by the fork loader by default (TENSOR_READ_LAZY).
; The shared head excludes embed_tokens (shares the main model's).
model              = /home/spencer/.cache/models/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf
mmproj             = /home/spencer/.cache/models/Qwen3.8-Flash-Next-mmproj-BF16.gguf
spec-draft-model   = /home/spencer/.cache/models/mtp-Qwen3.8-Flash-Next-shared-Q4_K_M.gguf
chat-template-file = /home/spencer/.config/llama.cpp/qwen-fixed-chat_template.jinja
reasoning-format   = deepseek
reasoning          = auto
reasoning-preserve = true
spec-type          = draft-mtp
spec-draft-n-max   = 5
temperature        = 1.0
top-p              = 0.95
top-k              = 20
min-p              = 0.0
presence-penalty   = 0.0
repeat-penalty     = 1.0