53a55e7317
Single-model deployment per Ryan's direction: - Primary model: casperhansen/deepseek-r1-distill-qwen-32b-awq, max_model_len=32768 - nomic-embed-text-v1.5 and Qwen3-8B-AWQ both disabled (single-model requirement) - kv_cache_dtype: int4_per_token_head required to fit 32768 ctx on 24GB RTX 3090 (fp16 KV: 26GB needed, doesn't fit at any utilization; fp8 KV: OOM'd during FlashInfer warmup with ~50-150MB margin; int4 KV: clean single-attempt start) - Added kv_cache_dtype / kv_cache_memory_bytes as new optional per-model template fields in vllm.service.j2 (guarded, no effect on other models) Verified live: /health 200, /v1/models confirms max_model_len=32768, live /v1/completions smoke test + manual chat completion both passed (genuine <think> reasoning trace, correct arithmetic). NRestarts=0, steady-state VRAM 23.2GB/24.576GB. Ansible idempotent re-run confirmed changed=0. Known follow-up (not done here): Hindsight's HINDSIGHT_API_LLM_MODEL cluster config still references the retired Qwen2.5-32B-Instruct-AWQ — needs separate GitOps update to point at the new model.
216 lines
12 KiB
YAML
216 lines
12 KiB
YAML
---
|
|
# ------------------------------------------------------------------------------
|
|
# FILE: ansible/host_vars/astro_orbiter/vars.yml
|
|
# HOST: astro-orbiter (10.1.71.130)
|
|
# ROLE: llama.cpp LLM inference host — Ryzen 7 5800XT / RTX 3090 (ATX rebuild,
|
|
# 2026-08-04). Superseded the prior AMD RX 5700 / Ollama config below;
|
|
# drive was transplanted into new hardware, not reinstalled.
|
|
# ------------------------------------------------------------------------------
|
|
|
|
ansible_host: 10.1.71.130
|
|
ansible_user: jarvis
|
|
ansible_ssh_private_key_file: ~/.ssh/id_jarvis
|
|
ansible_become: true
|
|
|
|
# LVM root expansion — xlarge template uses sda3 partition, standard VG/LV names
|
|
common_expand_root_lvm: true
|
|
common_root_pv: /dev/sda3
|
|
common_root_vg: ubuntu-vg
|
|
common_root_lv: ubuntu-lv
|
|
|
|
# --- Staged GGUF models for the llama.cpp router (:8002) ---------------------
|
|
# Data-driven list consumed by roles/llm-inference-multimodel tasks/models.yml
|
|
# (loop -> tasks/stage_model.yml). Each entry is idempotently staged into
|
|
# /opt/models: stat + EXACT-size check vs HF manifest; skip (no download, no
|
|
# restart) when present + size matches. Source repos are public bartowski GGUFs
|
|
# on HuggingFace (no auth). A router restart is notified ONLY when a new GGUF
|
|
# is actually downloaded.
|
|
# Added 2026-08-12 (War Machine): codify Phi-3.5-mini-instruct-Q8_0 and
|
|
# Meta-Llama-3.1-8B-Instruct-Q4_K_M as router models alongside the production
|
|
# Qwen3.6-35B-A3B-UD-Q4_K_S. The live files were already present/correct on
|
|
# astro-orbiter; this pass codifies them. Future adds = append to this list.
|
|
# Router --models-max override for astro-orbiter.
|
|
# Default in defaults/main.yml is 1 (conservative). Bumped to 4 on 2026-08-12
|
|
# (t_33acbb2e) so the router can keep more than one GGUF resident on-demand
|
|
# and LRU-evict when needed.
|
|
#
|
|
# VRAM NOTE (t_33acbb2e, updated t_55c164f5, updated t_34b96e83, updated t_f5f7e9ad, updated t_441470b9, updated t_c5cef2b2):
|
|
# With models-max=4 and all 6 GGUFs registered, worst case is all 6 loaded simultaneously:
|
|
# Qwen3.8-27B Q4_K_M: ~20.0GB (weights ~17.1GB + KV ~2.9GB @ 65536 ctx, q4_0) ← CORRECTED (ctx rolled back from 128K to 65536, t_c9fed26c 2026-08-18)
|
|
# Phi-3.5-mini-instruct Q8_0: ~4.3GB (weights ~3.8GB + KV ~0.5GB @ 32K ctx)
|
|
# Meta-Llama-3.1-8B Q4_K_M: ~5.6GB (weights ~4.6GB + KV ~0.2GB @ 8K ctx)
|
|
# Qwen2.5-Coder-14B Q4_K_M: ~9.0GB (weights ~8.4GB + KV ~0.6GB @ 16K ctx)
|
|
# nomic-embed-text-v1.5 Q4_K_M: ~0.09GB (~84MB, embedding only — no KV cache)
|
|
# Qwen3-8B Q4_K_M: ~5.5GB (weights ~4.68GB + KV ~0.5GB @ 32K ctx, q4_0)
|
|
# Total worst-case: ~44.5GB >> 24GB RTX 3090
|
|
#
|
|
# OOM RISK: Full co-residency is impossible on 24GB. LRU eviction prevents this
|
|
# in practice: models-max=4 means the router can REGISTER 6 models but only keeps
|
|
# up to 4 LOADED simultaneously — the router will evict the LRU model when a new
|
|
# one is needed. nomic-embed-text-v1.5 is pinned via sleep-idle-seconds=-1 and
|
|
# load-on-startup=true but it uses only ~84MB, so it never meaningfully changes
|
|
# the budget. In single-user homelab operation, only one generative model is active
|
|
# at a time alongside the always-resident embedding model.
|
|
# Qwen3.8-27B alone uses ~17,804 MiB (weights+KV @ 65536 ctx); co-residency
|
|
# with Coder (~9GB) = ~27GB > 24GB. LRU eviction handles this automatically.
|
|
# Ryan should be aware this means model-switching always incurs a ~30-60s
|
|
# cold-load latency when switching between Qwen3.8-27B and any other model.
|
|
# Proceeding to models-max=4 as instructed; flagged for Ryan's attention.
|
|
# Router --models-max override for astro-orbiter.
|
|
# UPDATED (t_f5f7e9ad, 2026-08-16): Set to 2 because Qwen3.8-27B-Q4_K_M
|
|
# uses 17,804 MiB at 65536 ctx. Only nomic-embed (558MB, pinned) and ONE
|
|
# generative model can be resident simultaneously. Co-residency of Qwen3.8
|
|
# with any auxiliary model (Phi 8.3GB, Llama 5.9GB, Coder 9GB) exceeds 24GB.
|
|
# models-max=2: slot 1 = nomic-embed (pinned, always loaded), slot 2 = LRU
|
|
# generative model (Qwen3.8 primary, cold-loaded on first request ~30-60s;
|
|
# auxiliary models evict it on demand, and vice versa).
|
|
# NOTE: Qwen3.8 does NOT have load-on-startup — it loads on first request.
|
|
# This avoids an LRU eviction race with nomic-embed at startup.
|
|
# UPDATED (t_72646029, 2026-08-17): CPU offload for Coder + Llama changes the
|
|
# constraint. Coder and Llama now use CPU inference (n-gpu-layers=0). GPU-resident
|
|
# VRAM: Qwen3.8 (~17,804 MiB at 65536 ctx) + nomic-embed (558 MiB, pinned) plus
|
|
# the CUDA-context buffers llama.cpp 6ea215d allocates for the CPU models (~1.4-1.7GB
|
|
# each) = ~20,004 MiB steady-state, below the 24,576 MiB physical limit.
|
|
# CORRECTED (t_c5cef2b2, 2026-08-19): ctx-size was rolled back from 131072 to 65536
|
|
# (t_c9fed26c 2026-08-18). Qwen3.8 VRAM at 65536: 17,804 MiB (not 20,302 MiB).
|
|
# models-max raised to 4: nomic (slot 1, pinned) + Qwen3.8 (slot 2, GPU) +
|
|
# Llama (slot 3, CPU) + Coder (slot 4, CPU). Phi (GPU, ~8.3GB) and new
|
|
# Qwen3-8B (GPU, ~5.5GB) can also be requested but evict Qwen3.8 due to VRAM.
|
|
# models-max=4 is required so CPU-offloaded models count as loaded without
|
|
# evicting Qwen3.8.
|
|
llm_router_models_max: 4
|
|
|
|
llm_staged_models:
|
|
- filename: "Phi-3.5-mini-instruct-Q8_0.gguf"
|
|
url: "https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF/resolve/main/Phi-3.5-mini-instruct-Q8_0.gguf"
|
|
size_bytes: 4061222688
|
|
source_repo: "bartowski/Phi-3.5-mini-instruct-GGUF"
|
|
- filename: "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf"
|
|
url: "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/main/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf"
|
|
size_bytes: 4920739232
|
|
source_repo: "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
|
|
- filename: "Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf"
|
|
url: "https://huggingface.co/bartowski/Qwen2.5-Coder-14B-Instruct-GGUF/resolve/main/Qwen2.5-Coder-14B-Instruct-Q4_K_M.gguf"
|
|
size_bytes: 8988111072
|
|
source_repo: "bartowski/Qwen2.5-Coder-14B-Instruct-GGUF"
|
|
- filename: "nomic-embed-text-v1.5-Q4_K_M.gguf"
|
|
url: "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF/resolve/main/nomic-embed-text-v1.5.Q4_K_M.gguf"
|
|
size_bytes: 84106624
|
|
source_repo: "nomic-ai/nomic-embed-text-v1.5-GGUF"
|
|
# Added t_c5cef2b2 (2026-08-19, War Machine): Qwen3-8B dense 8B model for
|
|
# aux tasks (routing, rewriting, structured extraction, tool-call construction).
|
|
# Source: bartowski/Qwen_Qwen3-8B-GGUF (public, no auth). HF filename is
|
|
# Qwen_Qwen3-8B-Q4_K_M.gguf; stored locally as Qwen3-8B-Q4_K_M.gguf.
|
|
# Exact size verified from HF manifest (content-length): 5,027,784,224 bytes.
|
|
# VRAM: ~4.68GB weights + ~0.5GB KV @ 32K ctx (q4_0) ≈ 5.2GB total.
|
|
# Thinking mode ON by default; use /no_think for latency-sensitive aux tasks.
|
|
- filename: "Qwen3-8B-Q4_K_M.gguf"
|
|
url: "https://huggingface.co/bartowski/Qwen_Qwen3-8B-GGUF/resolve/main/Qwen_Qwen3-8B-Q4_K_M.gguf"
|
|
size_bytes: 5027784224
|
|
source_repo: "bartowski/Qwen_Qwen3-8B-GGUF"
|
|
|
|
# --- deploy-vllm role: vllm_models override (t_r1d32b_swap, 2026-09-01) -----
|
|
# Ansible's hash_behaviour is "replace" (see ansible.cfg) — a host_vars list
|
|
# variable REPLACES the role default list wholesale, it does not deep-merge.
|
|
#
|
|
# SWAP (Ryan direction, 2026-09-01): Qwen2.5-32B-Instruct-AWQ retired,
|
|
# replaced with DeepSeek-R1-Distill-Qwen-32B-AWQ, max_model_len=32768.
|
|
# "Single model only" — nomic-embed-text-v1.5 (embedding, :8020) and
|
|
# Qwen3-8B-AWQ (aux, :8010, already disabled) are BOTH disabled here.
|
|
# DeepSeek gets the full 24GB card to itself. Nothing in production
|
|
# consumed nomic-embed at the time of this swap (Hindsight uses its own
|
|
# bundled 384-dim embedder; OpenViking pointed at the old llama-swap
|
|
# endpoint, already stopped) — confirmed with Ryan before disabling.
|
|
#
|
|
# Model choice: casperhansen/deepseek-r1-distill-qwen-32b-awq — same
|
|
# quantizer/toolchain (AutoAWQ) as the outgoing Qwen2.5-32B-Instruct-AWQ,
|
|
# widely used, 4-bit GEMM AWQ, ~19.3GB on disk (4 safetensors shards).
|
|
# Architecture: Qwen2ForCausalLM (DeepSeek-R1 distilled onto Qwen2.5-32B
|
|
# base) — same vLLM code path as the outgoing model, no new serving
|
|
# support needed. Native max_position_embeddings=131072; we cap at 32768
|
|
# per the task's explicit max-model-len requirement.
|
|
#
|
|
# VRAM math: ~19.3GB weights (4-bit AWQ) + KV cache at 32768 ctx (GQA,
|
|
# 8 KV heads, 128 head_dim, 64 layers, fp16 KV by default) ≈ 19.3GB +
|
|
# ~4GB KV+overhead ≈ 23.3GB — tight but the FULL 24GB card is now
|
|
# available (no co-resident nomic-embed/Qwen3-8B taking a share, unlike
|
|
# the outgoing Qwen2.5-32B config). gpu_memory_utilization=0.95 (role
|
|
# default) + enforce_eager retained as the proven-stable mitigation from
|
|
# t_e6facb19/t_ca1af9fb (avoids CUDA graph capture VRAM spike; this host's
|
|
# only validated way to avoid crash-loop-to-stabilize behavior on this
|
|
# card). If 0.95 OOMs at 32768 ctx once tested live, drop to 0.90 next
|
|
# (documented fallback, same pattern as the outgoing model).
|
|
#
|
|
# DeepSeek-R1 output note: reasoning traces stream in <think> tags before
|
|
# the final answer — this is expected R1-distill behavior, not a bug.
|
|
# Model card recommends temperature 0.5-0.7 (not 0, not vLLM's greedy
|
|
# default) to avoid repetition/incoherence; not set here (server-side
|
|
# default), left to be set client-side per the model card's guidance —
|
|
# flagging for whoever wires this into Hermes profile configs next.
|
|
vllm_models:
|
|
- id: "DeepSeek-R1-Distill-Qwen-32B-AWQ"
|
|
hf_repo: "casperhansen/deepseek-r1-distill-qwen-32b-awq"
|
|
role: primary
|
|
quantization: awq
|
|
port: 8000
|
|
max_model_len: 32768
|
|
# VRAM correction (2026-09-01, live test): weights alone load at 18.17
|
|
# GiB (confirmed via journalctl "Model loading took 18.17 GiB memory").
|
|
# fp16 KV cache at 32768 ctx needs 8.0 GiB per vLLM's own error message
|
|
# ("To serve at least one request with the model's max seq len (32768),
|
|
# 8.0 GiB KV cache is needed") — 18.17 + 8.0 = 26.17GB, DOES NOT FIT a
|
|
# 24GB card even at gpu_memory_utilization=1.0. Confirmed via 3 live
|
|
# crash-loop attempts at 0.95 (usable ~23.35GiB budget): vLLM's own
|
|
# ValueError reported "estimated maximum model length is 14528" at
|
|
# whatever KV budget was actually available — nowhere close to 32768.
|
|
# FIX: --kv-cache-dtype fp8 halves KV cache memory (~4.0GiB instead of
|
|
# 8.0GiB) — 18.17 + 4.0 = ~22.2GB, fits with ~1.4GB headroom at 0.95.
|
|
# fp8 KV cache is a standard vLLM feature (not experimental for this
|
|
# vLLM version), minor precision loss in attention scores, no known
|
|
# material quality impact for a distilled reasoning model at this scale.
|
|
# VRAM correction round 3 (2026-09-01, live test): manual
|
|
# kv_cache_memory_bytes did NOT fix it either — same OOM pattern
|
|
# (small ~150MB alloc failing with only ~52MB actually free), meaning
|
|
# real GPU usage at warmup exceeds what profiling/reservation account
|
|
# for by roughly ~1GB (unaccounted FlashInfer warmup workspace buffers,
|
|
# not sized during profiling or reservation). Rather than keep guessing
|
|
# at a KV byte budget with ~0 margin, switched KV dtype from fp8 (8-bit)
|
|
# to int4_per_token_head (4-bit) — roughly HALVES KV cache footprint
|
|
# for the same 32768 ctx (~2.0GiB instead of ~4.0-4.3GiB), buying back
|
|
# ~2GiB of real headroom to absorb whatever the unaccounted warmup
|
|
# overhead actually is. Reverted to standard percentage-based
|
|
# gpu_memory_utilization (no kv_cache_memory_bytes override) so vLLM's
|
|
# own profiling determines the exact KV budget again, now with much
|
|
# more slack in play. Some precision loss vs fp8 is expected for KV
|
|
# cache in int4 — acceptable tradeoff for a homelab box; revisit if
|
|
# DeepSeek's output quality visibly degrades once running.
|
|
kv_cache_dtype: int4_per_token_head
|
|
gpu_memory_utilization: 0.95
|
|
enforce_eager: true
|
|
enabled: true
|
|
- id: "Qwen3-8B-AWQ"
|
|
hf_repo: "Qwen/Qwen3-8B-AWQ"
|
|
role: aux
|
|
quantization: awq
|
|
port: 8010
|
|
max_model_len: 32768
|
|
gpu_memory_utilization: 0.15
|
|
enforce_eager: true
|
|
enabled: false # single-model deployment — see swap note above
|
|
- id: "nomic-embed-text-v1.5"
|
|
hf_repo: "nomic-ai/nomic-embed-text-v1.5"
|
|
role: embedding
|
|
quantization: none
|
|
port: 8020
|
|
max_model_len: 2048
|
|
gpu_memory_utilization: 0.05
|
|
trust_remote_code: true
|
|
enabled: false # single-model deployment — see swap note above
|
|
|
|
# --- deploy-vllm role: boot persistence (unchanged) -------------------------
|
|
# Still permanent/boot-persistent — same policy as the outgoing Qwen2.5-32B
|
|
# deployment (t_5508360a), just now serving one model instead of two.
|
|
vllm_service_enabled: true
|
|
vllm_service_state: started
|
|
|