rblundon/homelab
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
7867be688a1c73ae9e0af887f901d566251cd554
homelab/ansible/roles/llm-inference-multimodel/defaults/main.yml

338 lines
17 KiB
YAML
Raw Blame History

---
# ------------------------------------------------------------------------------
# FILE: roles/llm-inference-multimodel/defaults/main.yml
# DESCRIPTION: Overridable defaults for the llm-inference-multimodel role.
# Deploy target: astro-orbiter (10.1.71.130, RTX 3090 24GB).
# Built ALONGSIDE roles/llm-inference (not a replacement) — that
# role's CUDA/build/driver phases are the prerequisite; this role
# assumes /opt/llama.cpp/build/bin/llama-server already exists.
#
# See /home/hermes/astro-orbiter-multi-model-plan.md for the full
# approved design (VRAM math, rationale, rollback story).
# ------------------------------------------------------------------------------
# Shared
llm_service_user: jarvis
llm_binary_path: /opt/llama.cpp/build/bin/llama-server
llm_models_dir: /opt/models
# Bind address — deliberately NOT 0.0.0.0 (see plan §5). Default to the private
# LAN interface so both instances are reachable from Hermes but not the world.
# Override to 127.0.0.1 if even LAN-wide reachability is unwanted and a reverse
# proxy/localhost-only tunnel is used instead.
llm_bind_address: "10.1.71.130"
# Firewall scoping (Phase 3) — subnet/hosts allowed to reach the ports above.
# Override per-environment; default assumes Hermes runs somewhere on this /24.
llm_allowed_source_cidr: "10.1.70.0/24"
# --- RETIRED (2026-08-06): Aux / classification instance (port 8000, Phi-4-14B)
# and Tool-calling instance (port 8001, Mistral-Small-24B) --------------------
# Consolidated down to a single production model (Qwen2.5-14B-Instruct-1M,
# port 8002) serving BOTH the friday and war-machine Hermes profiles. Ryan
# explicitly accepted the tradeoffs (single model for chat + tool-calling +
# aux duties) over keeping the aux/toolcall split running.
# Both llama-server-aux and llama-server-toolcall services were stopped,
# disabled, and had their unit files removed from astro-orbiter; their GGUF
# weights (phi-4-14b-instruct-Q4_K_M.gguf, mistral-small-24b-instruct-2501-
# Q3_K_M.gguf) were deleted from /opt/models (~45GB reclaimed). The
# templates/tasks that deployed them have been removed from this role — see
# git log for the prior variable definitions and unit templates if a future
# rollback needs them restored.
# --- Production instance (port 8002, Qwen2.5-14B-Instruct-1M) ----------------
# History (2026-08-06): Qwen2.5-14B-Instruct (base) was deployed to this slot
# and DISQUALIFIED — live /v1/models meta reported n_ctx_train=32768, well
# under the 64K Hermes floor (the model card's "128K" figure conflated
# YaRN-extended inference-time scaling with actual trained context; disabled
# by default, not baked in). Llama-3.1-8B-Instruct was tried next — cleared
# the context gate (verified live n_ctx_train=131072) but failed the
# tool-calling validation harness badly (8/10 hallucination-stress prompts
# triggered spurious tool_calls even at temp=0.1 with the correct official
# chat template) — purged from disk and Ansible entirely, see git log.
# Current model: Qwen2.5-14B-Instruct-1M (bartowski GGUF) — distinct
# checkpoint with genuine additional long-context pretraining, NOT the same
# weights as the disqualified base model above. Live-verified 2026-08-06:
# /v1/models reports n_ctx=65536, n_ctx_train=1010000 (well over the 64K
# floor). Tool-calling verified live via a /v1/chat/completions probe with a
# tools= payload — returned a well-formed tool_calls response (finish_reason
# "tool_calls", valid JSON arguments), no hallucinated calls observed.
# PROMOTED TO PRODUCTION (2026-08-06): llm_qwen_service_enabled now defaults
# to true — this is the sole model serving both Hermes profiles. Ports
# 8000/8001 are permanently freed; no co-residency VRAM gate applies anymore.
llm_qwen_service_enabled: true
llm_qwen_port: 8002
llm_qwen_model_path: "{{ llm_models_dir }}/Qwen3.8-27B-Q4_K_M.gguf"
llm_qwen_model_min_bytes: 17000000000 # guard threshold; complete file ~17.1GB
llm_qwen_ctx_size: 65536
llm_qwen_parallel: 1
llm_qwen_gpu_layers: 99
llm_qwen_batch_size: 4096
llm_qwen_ubatch_size: 4096
llm_qwen_service_name: llama-server-qwen
llm_qwen_model_id: Qwen3.8-27B-Q4_K_M
llm_qwen_expected_vram_gb: 17 # Q4_K_M = 17.1GB weights + ~6GB KV @ 65536 ctx = ~23GB max
# NOTE (2026-08-16 t_f5f7e9ad): Qwen3.6-35B-A3B-UD-Q4_K_S superseded by
# Qwen3.8-27B-Q4_K_M per Ryan's direction. Qwen3.8-27B is a dense 27B VLM
# (Apache-2.0, Alibaba, Aug 2026) quantized by Unsloth Dynamic V3.0.
# Q4_K_M: 17,106,775,008 bytes. Downloaded out-of-band via wget.
# llm_qwen_model_url: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/main/Qwen3.8-27B-Q4_K_M.gguf
# --- Staged GGUF models (data-driven, idempotent staging) --------------------
# Additional GGUFs to ensure are present in llm_models_dir, alongside the
# production Qwen3.6-35B. Consumed by tasks/models.yml (loop over
# tasks/stage_model.yml). Each entry:
# filename: target filename in llm_models_dir
# url: HuggingFace resolve URL (public repos; no auth needed)
# size_bytes: EXACT expected byte size (HF manifest) — guard: download only
# if the file is missing OR its size != this value (idempotent;
# never re-pulls a correct file, never needlessly restarts).
# source_repo: upstream HF repo (audit/lineage)
# The REAL list is defined per-host in host_vars/astro-orbiter/vars.yml (NOT
# hardcoded here) so the role stays generic and reusable for future model adds.
# Empty default = nothing staged (safe no-op).
llm_staged_models: []
# --- Existing Gemma baseline (rollback target — never modified by this role) -
# Populated by Phase 0 discovery (tasks/discover.yml) if not already known.
# Set here only as a fallback name to search for; discovery is authoritative.
llm_existing_gemma_service_name_guess: llama-server
# --- Router mode shadow deployment (port 8003) --------------------------------
# Deploy llama-server in router/supervisor mode (no -m flag) on a shadow port.
# Production unit (llama-server-qwen, port 8002) is UNCHANGED until validation
# gates pass and Ryan explicitly approves cutover.
#
# Default: llm_router_enabled: false — all router tasks are no-ops until you
# flip this to true (either in host_vars, extra-vars, or the shadow playbook).
#
# CRITICAL: llm_router_models_max default is 1 here for safety. It is
# overridden to 4 in host_vars/astro-orbiter/vars.yml (t_33acbb2e) with
# a full VRAM budget note. DO NOT raise it without a VRAM budget review.
# Default llama-server cap is 4 simultaneous — that would OOM a 24GB card
# immediately when Qwen3.6-35B (20GB) is the resident model.
#
# Added 2026-08-12 (t_0cca74a2): router mode migration — War Machine.
llm_router_enabled: false
llm_router_port: 8003
llm_router_service_name: llama-server-router
llm_router_models_dir: "{{ llm_models_dir }}" # /opt/models — same dir as production
llm_router_models_max: 1 # CRITICAL: RTX 3090 24GB, single model only
llm_router_ctx_size: 65536 # 64K — must match production (Hermes floor)
llm_router_parallel: 1
llm_router_gpu_layers: 99
llm_router_batch_size: 4096
llm_router_ubatch_size: 4096
llm_router_cache_type_k: q4_0 # required to fit 64K KV in 24GB
llm_router_cache_type_v: q4_0
llm_router_flash_attn: "auto"
llm_router_bind_address: "{{ llm_bind_address }}" # 10.1.71.130
llm_router_allowed_source_cidr: "{{ llm_allowed_source_cidr }}" # 10.1.70.0/24
llm_router_expected_model_id: "Qwen3.8-27B-Q4_K_M" # verified at Gate 1
llm_router_vram_max_mib: 23000 # Gate 3: fail if exceeded under load
# --- Router preset mode (--models-preset INI) ---------------------------------
# Set llm_router_preset_enabled: true to switch from --models-dir to
# --models-preset. Preset mode is REQUIRED to support model aliases.
# The template at llama-server-router-preset.ini.j2 defines all 3 router models:
# - Qwen3.6-35B-A3B-UD-Q4_K_S (no alias — primary ID unchanged)
# - Phi-3.5-mini-instruct-Q8_0 (alias: Phi-3.5-mini-instruct-8bit) <-- t_9adf0889
# - Meta-Llama-3.1-8B-Instruct-Q4_K_M (no alias — primary ID unchanged)
#
# llm_router_preset_path: on-disk path where the rendered INI is deployed.
# Default: /opt/llama-server-router-preset.ini (owned by root, readable by all).
#
# GH #22364 note: --models-preset causes an extra "default" entry in /v1/models.
# This is cosmetic and does not affect model selection by name. Accept it.
#
# Added 2026-08-12 (t_9adf0889) — War Machine.
llm_router_preset_enabled: false # flip true to activate preset mode
# Per-model ctx-size / flash-attn overrides for preset mode (t_ryan_per_model_ctx).
# Defaults mirror the prior uniform 65536/auto behavior; host_vars or the
# deploy playbook override these to the values Ryan requested per workload.
llm_router_llama_ctx_size: "{{ llm_router_ctx_size }}"
llm_router_llama_flash_attn: "{{ llm_router_flash_attn }}"
llm_router_phi_ctx_size: "{{ llm_router_ctx_size }}"
llm_router_phi_flash_attn: "{{ llm_router_flash_attn }}"
# Qwen2.5-Coder-14B: ctx_size=16384, flash_attn=true per task t_55c164f5
llm_router_coder_ctx_size: 16384
llm_router_coder_flash_attn: "true"
# CPU offload vars (t_72646029, 2026-08-17): n-gpu-layers=0 moves Coder and Llama to
# full CPU inference. Allows concurrent residency with Qwen3.8-27B. NOTE: llama.cpp
# 6ea215d still allocates ~1.4-1.7GB CUDA-context VRAM per CPU model, so steady-state
# is ~24,004 MiB (at the 24,576 MiB physical limit), not the 0-VRAM the spec assumed.
llm_router_coder_gpu_layers: 0
llm_router_llama_gpu_layers: 0
llm_router_preset_path: /opt/llama-server-router-preset.ini
# Qwen3.8-27B: ctx=65536 (64K). Bumped 32768 -> 131072 (t_441470b9, 2026-08-16);
# rolled back to 65536 (t_c9fed26c follow-up, 2026-08-18) after t_72646029 CPU-offload
# deployment moved Phi-3.5mini back to GPU, exceeding RTX 3090 24,576 MiB ceiling.
# At 131072 ctx + all 5 models resident, Qwen3.8 fails to load (HTTP 500 OOM).
# 64K satisfies the 2026-08-12 cutover validation Gate 1 (n_ctx >= 64000).
# Full VRAM analysis and Phase 2 options documented in
# playbooks/day2_qwen38_ctx128k_rollback.yml.
llm_router_qwen38_ctx_size: 65536
# nomic-embed-text-v1.5: embedding model, ctx-size=8192 per task t_34b96e83
# No flash_attn or KV cache params - embedding models use bidirectional forward pass,
# not autoregressive KV cache. load-on-startup=true / sleep-idle-seconds=-1 keep it
# always warm at negligible VRAM cost (~84MB).
llm_router_nomic_ctx_size: 8192
# FIX (2026-08-14, t_openviking_embed_batch): batch-size/ubatch-size were
# previously omitted from this section entirely, so llama-server silently
# defaulted the physical batch (ubatch-size) to 512 tokens. Embedding requests
# cannot be split across ubatches in llama.cpp, so any OpenViking chunk over
# ~512 tokens (observed 2000-3400 tokens/chunk from openviking-config's
# embedding.dense chunking) hard-failed with "input (N tokens) is too large to
# process. increase the physical batch size" - this fed OpenViking's circuit
# breaker into a permanent fail/re-enqueue loop. 4096 covers the observed max
# comfortably while staying under ctx-size=8192.
llm_router_nomic_batch_size: 4096
llm_router_nomic_ubatch_size: 4096
# --- Monitoring: VRAM exporter + Prometheus scrape + Grafana dashboard -------
# Phase 3: GPU/LLM monitoring deployment (Wong, 2026-08-18)
# Provides: VRAM textfile exporter, Prometheus scrape config for llama-swap
# /metrics endpoint, Grafana 6-panel dashboard, PrometheusRule alert rules.
#
# Ref: roles/llm-inference-multimodel/references/monitoring-llm-homelab-ciro-luciotta-2026.md
llm_monitoring_enabled: true # gate for monitoring tasks
llm_vram_exporter_script: /opt/llama-server-monitoring/nvidia-smi-vram-exporter.sh
llm_vram_exporter_cron_minute: "*" # run every minute
llm_vram_exporter_gpu_index: 0 # GPU 0 (RTX 3090 on astro-orbiter)
llm_vram_textfile_dir: /var/lib/node_exporter/textfile_collector
# Alert thresholds (per Ciro Luciotta pattern)
llm_vram_critical_mib: 24000 # ~90% of 24GB RTX 3090
llm_kv_cache_spill_ratio: 0.92 # KV-cache spill threshold
llm_throughput_baseline_tokens_per_min: 50 # baseline for degradation alert
# Grafana dashboard
llm_grafana_dashboard_uid: llama-swap-monitor
llm_grafana_dashboard_title: "llama-swap GPU/LLM Monitoring"
llm_grafana_dashboard_tags:
- llm
- llama-swap
- gpu-monitoring
- ciro-luciotta
llm_grafana_dashboard_refresh: "30s"
llm_grafana_dashboard_time_from: "now-24h"
# Prometheus scrape job
llm_prometheus_scrape_interval: "30s"
llm_prometheus_scrape_timeout: "10s"
# --- llama-swap mode (port 8001) -----------------------------------------------
# Deploy llama-swap — Go-based hot-swap proxy (v250+) for model orchestration.
# Replaces router mode entirely: single binary + YAML config.json, no --models-preset INI.
# Additive deployment (non-invasive); production router (port 8002) stays running during Phase 1 shadow.
#
# Default: llm_swapmode_enabled: false — all llama-swap tasks are no-ops until flipped to true.
# Gated by Phase 3 go/no-go once War Machine Phase 1-2 validation completes.
#
# NOTE: llama-swap v250 config format differs from evaluation docs (§4b).
# Uses routing.router DSL with expression-based matrix, not old list-of-arrays syntax.
# See /etc/llama-swap/config.yaml on astro-orbiter (Phase 1 artifact) for reference.
#
# Added 2026-08-18 (t_c1e44190): llama-swap Phase 3 Ansible integration — Wong.
llm_swapmode_enabled: false # Gate for llama-swap tasks (Phase 3)
llm_swapmode_port: 8001 # Shadow port (Phase 1), becomes production in Phase 3
llm_swapmode_bind_address: "{{ llm_bind_address }}" # 10.1.71.130
llm_swapmode_allowed_source_cidr: "{{ llm_allowed_source_cidr }}" # 10.1.70.0/24
# Binary installation
llm_swapmode_binary_url: "https://github.com/mostlygeek/llama-swap/releases/download/v250/llama-swap-linux-amd64.tar.gz"
llm_swapmode_binary_version: "v250"
llm_swapmode_checksum: "sha256:60226b64fcc78e8de6e9d4fac78de95372c2c2a0a31fd6b7d26d1e77ea7c9d9d" # From Phase 1 deployment
# Directories
llm_swapmode_config_dir: /etc/llama-swap
llm_swapmode_config_file: "{{ llm_swapmode_config_dir }}/config.yaml"
llm_swapmode_models_dir: "{{ llm_models_dir }}" # /opt/models — same as production
# Service
llm_swapmode_service_name: llama-swap
llm_swapmode_service_user: "{{ llm_service_user }}" # jarvis
llm_swapmode_vram_max_mib: 23000 # Gate 3: fail if exceeded under load
# Consolidated model list for llama-swap config.yaml
# Each model specifies full per-model config (ctx_size, n_gpu_layers, cmd args)
# Instead of scattered llm_router_* variables, this is the structure llama-swap expects
# (matches the v250 config.yaml YAML structure, not the router's INI/per-model variables)
llm_swapmode_models:
- id: Qwen3.8-27B-Q4_K_M
gguf_path: "{{ llm_models_dir }}/Qwen3.8-27B-Q4_K_M.gguf"
port: 8105
n_gpu_layers: -1 # -1 = auto-detect / all layers to GPU
ctx_size: 65536
batch_size: 4096
ubatch_size: 4096
parallel: 1
cache_type: q8_0
flash_attn: true
sleep_idle_seconds: -1 # never idle (primary model — always ready)
load_on_startup: true
- id: Qwen2.5-Coder-14B-Instruct-Q4_K_M
gguf_path: "{{ llm_models_dir }}/Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf"
port: 8101
n_gpu_layers: 0 # CPU-offload (aux model)
ctx_size: 16384
batch_size: 4096
ubatch_size: 4096
parallel: 1
flash_attn: "true"
sleep_idle_seconds: 60 # idle after 60s no requests
- id: Meta-Llama-3.1-8B-Instruct-Q4_K_M
gguf_path: "{{ llm_models_dir }}/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf"
port: 8102
n_gpu_layers: 0 # CPU-offload (aux model)
ctx_size: 8192
batch_size: 4096
ubatch_size: 4096
parallel: 1
flash_attn: "true"
sleep_idle_seconds: 60
- id: Phi-3.5-mini-instruct-Q8_0
gguf_path: "{{ llm_models_dir }}/Phi-3.5-mini-instruct-Q8_0.gguf"
port: 8104
n_gpu_layers: 0 # CPU-offload (aux model)
ctx_size: 32768
batch_size: 4096
ubatch_size: 4096
parallel: 1
flash_attn: "true"
sleep_idle_seconds: 60
- id: nomic-embed-text-v1.5
gguf_path: "{{ llm_models_dir }}/nomic-embed-text-v1.5-Q4_K_M.gguf"
port: 8103
n_gpu_layers: 0 # CPU-offload (embedding model — always on)
ctx_size: 8192
batch_size: 4096
ubatch_size: 4096
parallel: 1
sleep_idle_seconds: -1 # never idle (always ready for embeddings)
load_on_startup: true
# llama-swap matrix routing configuration
# Each row defines a set of models that can be co-resident and hot-swappable
# Syntax: "model1 & model2" = both models in same row (via v250 expression DSL)
llm_swapmode_matrix_rows:
- row: row0
expr: "nomic-embed-text-v1.5" # Embedding-only row
- row: row1
expr: "Qwen3.8-27B-Q4_K_M & nomic-embed-text-v1.5" # Primary + embed
- row: row2
expr: "Meta-Llama-3.1-8B-Instruct-Q4_K_M & nomic-embed-text-v1.5" # Aux LLM + embed
- row: row3
expr: "Qwen2.5-Coder-14B-Instruct-Q4_K_M & nomic-embed-text-v1.5" # Coder + embed
- row: row4
expr: "Phi-3.5-mini-instruct-Q8_0 & nomic-embed-text-v1.5" # Mini + embed
Reference in New Issue View Git Blame Copy Permalink