vh/esh-pfi-infrastructure
1
0
Fork 0
Code Issues 7 Pull Requests Actions Packages Projects Releases Wiki Activity
Files
2e80e69ef5d783ccf3428462bede6367ac7d1dec
esh-pfi-infrastructure/stacks/llama-swap/config.yaml
T
vh 58f2a22966 llama-swap: pin qwen3.5-9b + qwen3.6-35-a3b as a coexistence group 
Adds a new `pinned` group with swap: false (models coexist in VRAM),
exclusive: false (group shares with other groups), persistent: true
(never unload). Each member also gets ttl: 0 so the per-model
idle-timeout can't drop them either — belt + suspenders.

Pair is currently qwen3.5-9b (~6 GB Q4) + qwen3.6-35-a3b (~29 GB Q6).
Plus the 128K KV caches, roughly 50-60 GB VRAM resident. Appropriate
for an A6000/H100-class card; verify fit after deploy.

Committed as a canonical change; push + restart still needed on
ana-ml2.
2026-04-22 09:05:19 -07:00

505 lines
17 KiB
YAML
Raw Blame History

# ============================================================================
# llama-swap configuration for PFI-ANA
# Optimized and synchronized with /models disk inventory
# Last updated: 2026-04-10
#
# KB Sources:
# - reference/nemotron-3-super-running-parameters.md
# - reference/nemotron-3-nano-running-parameters.md
# - reference/qwen3.5-running-parameters.md
# - reference/qwen3-coder-next-running-parameters.md
# - reference/gemma-4-running-parameters.md
# - reference/qwen3-embedding-running-parameters.md
# - reference/qwen3-reranker-running-parameters.md
#
# Changelog:
# 2025-07-22: Removed jina-reranker-v3 (unused, out of rotation).
# 2026-04-10: Fixed Qwen3-Embedding pooling (mean→last; causal LM uses last-token
# pooling). Fixed ctx-size 4096→8192 for embedding+reranker. Fixed
# reranker: removed --embeddings flag (not an embedding model).
# 2026-04-17: Added Qwen3.6-35B-A3B Abliterated Heretic Q8_0 via -hf syntax.
# Requires HF_HOME=/hfcache in compose (see docker-compose.yml).
# New convention: use -hf repo[:quant] instead of --model /path.
# 2026-04-20: Added stock unsloth Qwen3.6-35B-A3B at Q6_K_XL via -hf syntax.
# Swapped abliterated entry from IIEleven11 Heretic Q8_0 to
# mradermacher abliterated i1-Q6_K (already cached).
# Bumped both Qwen 3.6 entries ctx-size 32768 → 131072.
# 2026-04-22: New `pinned` group — qwen3.5-9b + qwen3.6-35-a3b coexist
# in VRAM with persistent=true and ttl=0. Means the two
# can be called concurrently and never idle-unload.
# ============================================================================
# Default 1200 seconds (20 min) to wait for model to be available to load.
healthCheckTimeout: 1200
# logLevel: sets the logging value
# - optional, default: info
# - Valid log levels: debug, info, warn, error
logLevel: info
# metricsMaxInMemory: maximum number of metrics to keep in memory
# - optional, default: 1000
metricsMaxInMemory: 1000
# startPort: sets the starting port number for the automatic ${PORT} macro.
# - optional, default: 5800
# - the ${PORT} macro can be used in model.cmd and model.proxy settings
# - it is automatically incremented for every model that uses it
# startPort: 10001
models:
# ==========================================================================
# QWEN 3.5 MODELS (KB-recommended settings)
# - Thinking mode: temp 1.0, top-p 0.95, top-k 20, min-p 0.0, presence_penalty 1.5
# - Coding (precise): temp 0.6, top-p 0.95, top-k 20, min-p 0.0, presence_penalty 0.0
# - Non-thinking general: temp 0.7, top-p 0.8, top-k 20, min-p 0.0, presence_penalty 1.5
# - Context: 256K native (start 16K-32K for responsiveness)
# - Gibberish fix: add --cache-type-k bf16 --cache-type-v bf16
# - No Ollama support for Qwen3.5 GGUFs — use llama.cpp only
# ==========================================================================
"qwen3.5-35-a3b":
name: "Qwen 3.5 35B-A3B Thinking"
description: "MoE reasoning model. 3B active params, general-purpose thinking/chat."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_Qwen3.5-35B-A3B-GGUF/Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.00
--presence-penalty 1.5
--chat-template-kwargs '{"enable_thinking":true}'
"qwen3.5-122b-a10b":
name: "Qwen 3.5 122B-A10B UD-Q4_K_XL"
description: "Large MoE reasoning model. 10B active params, heavy reasoning tasks."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_Qwen3.5-122B-A10B-GGUF/UD-Q4_K_XL/Qwen3.5-122B-A10B-UD-Q4_K_XL-00001-of-00003.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.00
--presence-penalty 1.5
--chat-template-kwargs '{"enable_thinking":true}'
"qwen3.5-9b":
name: "Qwen 3.5 9B UD-Q4_K_XL"
description: "Dense 9B model. Lightweight general-purpose chat and reasoning."
ttl: 0 # pinned — member of the `pinned` group, never unloads
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_Qwen3.5-9B-GGUF/Qwen3.5-9B-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.00
--presence-penalty 1.5
--chat-template-kwargs '{"enable_thinking":true}'
# --------------------------------------------------------------------------
# Qwen 3.6 — uses -hf syntax, reads from HF_HOME=/hfcache (host pre-download)
# --------------------------------------------------------------------------
"qwen3.6-35-a3b":
name: "Qwen 3.6 35B-A3B UD-Q6_K_XL"
description: "Qwen3.6 MoE, 3B active of 35B. Stock unsloth Q6_K_XL (~29GB). Thinking on by default."
ttl: 0 # pinned — member of the `pinned` group, never unloads
cmd: |
/app/llama-server
--context-shift
--jinja
-hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-Q6_K_XL
--port ${PORT}
--n-gpu-layers 999
--ctx-size 131072
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.00
--presence-penalty 1.5
--repeat-penalty 1.0
--reasoning on
--reasoning-format deepseek
"qwen3.6-35-a3b-abliterated":
name: "Qwen 3.6 35B-A3B Abliterated i1-Q6_K"
description: "Qwen3.6 MoE, 3B active of 35B. mradermacher abliterated imatrix Q6_K (~27GB)."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--jinja
-hf mradermacher/Qwen3.6-35B-A3B-abliterated-i1-GGUF:i1-Q6_K
--port ${PORT}
--n-gpu-layers 999
--ctx-size 131072
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 20
--min-p 0.00
--presence-penalty 1.5
--repeat-penalty 1.0
--reasoning on
--reasoning-format deepseek
# ==========================================================================
# NEMOTRON MODELS (KB-recommended settings)
# - General Chat: temp 1.0, top-p 1.0, min_p 0.01
# - Tool Calling: temp 0.6, top-p 0.95, min_p 0.01
# - NoPE architecture: no YaRN needed
# - DEPRECATED --special flag for reasoning tokens
# - --special flag causes issues.
# - Start ctx 16K-32K, increase cautiously
# ==========================================================================
"nemotron-3-super-120b":
name: "NVIDIA Nemotron 3 Super 120B-A12B UD-Q4_K_XL"
description: "Flagship NVIDIA reasoning model. 12B active of 120B, MoE. 64-72GB VRAM at Q4."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_NVIDIA-Nemotron-3-Super-120B-A12B-GGUF/UD-Q4_K_XL/NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_XL-00001-of-00003.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 16384
--flash-attn on
--temp 1.0
--top-p 1.0
--min-p 0.01
--seed 3407
"nemotron-3-nano-30b":
name: "NVIDIA Nemotron 3 Nano 30B-A3B UD-Q4_K_XL"
description: "Compact Nemotron. 3B active of 30B, MoE. ~24GB at Q4. Best performance/size on 24GB GPUs."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--special
--model /models/unsloth_Nemotron-3-Nano-30B-A3B-GGUF/Nemotron-3-Nano-30B-A3B-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 1.0
--min-p 0.01
--seed 3407
# ==========================================================================
# GEMMA 4 MODELS (KB-recommended settings)
# - All variants: temp 1.0, top-p 0.95, top-k 64, repeat_penalty 1.0
# - Thinking: enable via --chat-template-kwargs '{"enable_thinking":true}'
# - Multi-turn: only keep final visible answer in history (not thought blocks)
# - Context: E2B/E4B=128K, 26B-A4B/31B=256K. Start at 32K.
# - ⚠️ Do NOT use CUDA 13.2 runtime — causes poor outputs
# - Use llama-server (not llama-cli) for thinking control
# ==========================================================================
"gemma4-26b-a4b":
name: "Gemma 4 26B-A4B"
description: "Google DeepMind Gemma 4 MoE, 4B active params, 256K context. Best speed/quality tradeoff."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 64
--repeat-penalty 1.0
--chat-template-kwargs '{"enable_thinking":true}'
"gemma4-31b-dense":
name: "Gemma 4 31B Dense"
description: "Google DeepMind Gemma 4 dense 31B. Maximum quality for complex reasoning, 256K context."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_gemma-4-31B-it-GGUF/gemma-4-31B-it-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 64
--repeat-penalty 1.0
--chat-template-kwargs '{"enable_thinking":true}'
# ==========================================================================
# GLM MODELS
# ==========================================================================
"glm4.7-flash":
name: "GLM 4.7 Flash UD-Q4_K_XL"
description: "THUDM GLM 4.7 Flash. Fast inference, general-purpose chat."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_GLM-4.7-Flash-GGUF/GLM-4.7-Flash-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 40000
--flash-attn on
--temp 0.6
--top-p 0.95
"glm-steam-106b":
name: "GLM Steam 106B-A12B Q4_K_M"
description: "TheDrummer GLM Steam MoE. 12B active of 106B. Creative and RP-focused."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/RP/bartowski_TheDrummer_GLM-Steam-106B-A12B-v1-GGUF/TheDrummer_GLM-Steam-106B-A12B-v1-Q4_K_M-00001-of-00002.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 40000
--flash-attn on
--temp 0.6
--top-p 0.95
# ==========================================================================
# SKYFALL MODELS
# ==========================================================================
"skyfall-r1-31b-q6k":
name: "Skyfall R1 31B v4 Q6_K_L"
description: "TheDrummer Skyfall R1 31B v4. General-purpose reasoning."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/bartowski_TheDrummer_Skyfall-31B-v4-GGUF/TheDrummer_Skyfall-31B-v4-Q6_K_L.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 40000
--flash-attn on
"skyfall-r1-31b-v4a":
name: "Skyfall R1 31B v4a Q6_K (RP)"
description: "BeaverAI Skyfall R1 v4a variant. RP/creative-focused."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/RP/BeaverAI_Skyfall-R1-31B-v4a-GGUF/Skyfall-R1-31B-v4a-Q6_K.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 40000
--flash-attn on
# ==========================================================================
# CODER MODELS
# ==========================================================================
"qwen3-coder-next":
name: "Qwen3 Coder Next UD-Q4_K_XL"
description: "Latest Qwen3 Coder. Non-reasoning model, optimized for code gen. KB: temp 1.0, top-k 40, min-p 0.01."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-UD-Q4_K_XL.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 1.0
--top-p 0.95
--top-k 40
--min-p 0.01
--repeat-penalty 1.0
# ==========================================================================
# LARGE / SPECIAL-PURPOSE MODELS
# ==========================================================================
"kimik2-q2kxl":
name: "Kimi K2 Instruct UD-Q2_K_XL"
description: "Moonshot Kimi K2. Huge MoE model (8-shard Q2). Limited GPU layers due to size."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_Kimi-K2-Instruct-0905-GGUF/UD-Q2_K_XL/Kimi-K2-Instruct-0905-UD-Q2_K_XL-00001-of-00008.gguf
--port ${PORT}
--n-gpu-layers 2
--temp 0.6
--top-p 0.95
# ==========================================================================
# GRANITE MODELS (IBM)
# ==========================================================================
"granite-4-small":
name: "Granite 4.0 Small Q4_K_M"
description: "IBM Granite 4.0 Small. Deterministic utility model for structured tasks."
ttl: 0
cmd: |
/app/llama-server
--context-shift
--model /models/unsloth_granite-4.0-h-small-GGUF/granite-4.0-h-small-Q4_K_M.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 120000
--flash-attn on
--top-p 1.0
--temp 0.0
--top-k 0
"granite-4-micro":
name: "Granite 4.0 Micro Q4_K_M"
description: "IBM Granite 4.0 Micro. Ultra-lightweight for fast structured responses."
ttl: 600
cmd: |
/app/llama-server
--context-shift
--model /models/ibm-granite_granite-4.0-micro-GGUF/granite-4.0-micro-Q4_K_M.gguf
--port ${PORT}
--n-gpu-layers 999
--ctx-size 32768
--flash-attn on
--temp 0.0
--top-p 1.0
# ==========================================================================
# EMBEDDING MODELS (persistent, always loaded)
# ==========================================================================
"embeddinggemma-300M":
name: "Embedding Gemma 300M"
description: "Google Embedding Gemma for vectorization."
ttl: 0
cmd: |
/app/llama-server
--embedding
--pooling cls
--model /models/ggml-org_embeddinggemma-300M-GGUF/embeddinggemma-300M-Q8_0.gguf
--port ${PORT}
--n-gpu-layers 0
--ctx-size 2048
--batch-size 1024
--no-mmap
--ubatch-size 1024
--cont-batching
--threads 24
"qwen3-embedding-0.6B":
name: "Qwen3 Embedding 0.6B"
description: "Qwen3 Embedding model for vectorization. 32K context, last-token pooling (decoder/causal LM)."
ttl: 0
cmd: |
/app/llama-server
--embeddings
--pooling last
--model /models/Qwen_Qwen3-Embedding-0.6B-GGUF/Qwen3-Embedding-0.6B-Q8_0.gguf
--port ${PORT}
--n-gpu-layers 0
--ctx-size 8192
--batch-size 8192
--ubatch-size 2048
--no-mmap
--cont-batching
--threads 24
# ==========================================================================
# RERANKING MODELS (persistent, always loaded)
# ==========================================================================
"qwen3-reranker-0.6B":
name: "Qwen3 Reranker 0.6B"
description: "Qwen3 Reranker for retrieval reranking. Causal LM scoring yes/no logits at last token. Replaces BGE v2."
ttl: 0
cmd: |
/app/llama-server
--reranking
--pooling rank
--model /models/ggml-org_Qwen3-Reranker-0.6B-Q8_0-GGUF/qwen3-reranker-0.6b-q8_0.gguf
--port ${PORT}
--n-gpu-layers 0
--ctx-size 8192
--batch-size 8192
--ubatch-size 2048
--cont-batching
--threads 24
# NOTE: jina-reranker-v3 removed 2025-07-22 — unused, out of rotation.
# Qwen3 Reranker handles all reranking duties.
# ============================================================================
# GROUPS
# - swap: false = models in group can coexist in memory
# - exclusive: false = group can share memory with other groups
# - persistent: true = models never unload (for utility/embedding)
# ============================================================================
groups:
"high-reasoning":
swap: false
exclusive: false
members:
- "qwen3.5-35-a3b"
- "gemma4-31b-dense"
- "nemotron-3-nano-30b"
"heavy-moe":
swap: false
exclusive: false
members:
- "qwen3.5-122b-a10b"
- "nemotron-3-super-120b"
- "kimik2-q2kxl"
- "glm-steam-106b"
"utility":
swap: false
exclusive: false
persistent: true
members:
- "embeddinggemma-300M"
- "qwen3-embedding-0.6B"
- "qwen3-reranker-0.6B"
# Pinned general-purpose chat models. Coexist in VRAM, never unload.
# Members also have ttl: 0 individually so idle-timeout can't drop them.
# VRAM budget check: qwen3.5-9b (~6 GB at Q4) + qwen3.6-35-a3b (~29 GB
# at Q6_K_XL) + 128K KV cache for each ≈ 50-60 GB. Size for your GPU.
"pinned":
swap: false
exclusive: false
persistent: true
members:
- "qwen3.5-9b"
- "qwen3.6-35-a3b"
Reference in New Issue View Git Blame Copy Permalink