feat(astro-orbiter): add Qwen3-8B-Q4_K_M to inference stack (t_c5cef2b2)

- host_vars/astro-orbiter/vars.yml:
  - Stage Qwen3-8B-Q4_K_M GGUF (bartowski/Qwen_Qwen3-8B-GGUF, 5,027,784,224 bytes)
  - Fix stale VRAM comments: KV @ 128K ctx -> 65536 ctx (rolled back t_c9fed26c 2026-08-18)
  - Update worst-case VRAM table to include new 6th model

- defaults/main.yml:
  - Add Qwen3-8B-Q4_K_M to llm_swapmode_models (port 8106, GPU, 32K ctx, q4_0 KV)
  - Add row5 to llm_swapmode_matrix_rows (Qwen3-8B & nomic-embed)

- templates/llama-server-router-preset.ini.j2:
  - Add [Qwen3-8B-Q4_K_M] section (32K ctx, GPU, flash-attn, q4_0 KV, 60s idle evict)
  - Document thinking-mode handling: enabled by default; /no_think at call time for aux tasks

VRAM math: Qwen3-8B ~4.68GB weights + ~0.5GB KV @ 32K = ~5.2GB. Fits with nomic-embed
(~84MB) well within 24GB. Cannot co-reside with Qwen3.8-27B (17.8GB); LRU eviction applies.

Note: war-machine profile config.yaml also updated (Qwen3.6 stale entry -> Qwen3.8-27B;
added Qwen3-8B-Q4_K_M context_length: 32768). config.yaml not tracked in git.

ansible/host_vars/astro-orbiter/vars.yml

@@ -34,26 +34,27 @@ common_root_lv: ubuntu-lv
 # (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):
+# 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 5 GGUFs registered, worst case is all 5 loaded simultaneously:
+# With models-max=4 and all 6 GGUFs registered, worst case is all 6 loaded simultaneously:
-#   Qwen3.8-27B Q4_K_M:               ~23.3GB (weights ~17.1GB + KV ~6.2GB @ 128K ctx, q4_0) ← UPDATED
+#   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)
-#   Total worst-case:                 ~42.3GB  >>  24GB RTX 3090
+#   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 5 models but only keeps
+# 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 ~23.1GB (weights+KV); co-residency with Coder (~9GB) = ~32GB > 24GB.
+# Qwen3.8-27B alone uses ~17,804 MiB (weights+KV @ 65536 ctx); co-residency
-# LRU eviction handles this automatically — the router evicts the idle model before
+# with Coder (~9GB) = ~27GB > 24GB. LRU eviction handles this automatically.
-# loading the new one. Ryan should be aware this means model-switching always incurs
+# Ryan should be aware this means model-switching always incurs a ~30-60s
-# a ~30-60s cold-load latency when switching between Qwen3.8-27B and any other model.
+# 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
@@ -67,13 +68,16 @@ common_root_lv: ubuntu-lv
 # 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 (~20,302 MiB at 128K ctx) + nomic-embed (558 MiB, pinned) plus the
+# VRAM: Qwen3.8 (~17,804 MiB at 65536 ctx) + nomic-embed (558 MiB, pinned) plus
-# CUDA-context buffers llama.cpp 6ea215d allocates for the CPU models (~1.4-1.7GB
+# the CUDA-context buffers llama.cpp 6ea215d allocates for the CPU models (~1.4-1.7GB
-# each) = ~24,004 MiB steady-state, below the 24,576 MiB physical limit.
+# 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) can still be
+# Llama (slot 3, CPU) + Coder (slot 4, CPU). Phi (GPU, ~8.3GB) and new
-# requested but evicts Qwen3.8 due to VRAM constraint. models-max=4
+# Qwen3-8B (GPU, ~5.5GB) can also be requested but evict Qwen3.8 due to VRAM.
-# is required so CPU-offloaded models count as loaded without evicting Qwen3.8.
+# models-max=4 is required so CPU-offloaded models count as loaded without
+# evicting Qwen3.8.
 llm_router_models_max: 4
 
 llm_staged_models:
@@ -93,4 +97,15 @@ llm_staged_models:
     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"
 

ansible/roles/llm-inference-multimodel/defaults/main.yml

@@ -317,6 +317,26 @@ llm_swapmode_models:
     sleep_idle_seconds: -1 # never idle (always ready for embeddings)
     load_on_startup: true
 
+  # Added t_c5cef2b2 (2026-08-19, War Machine): Qwen3-8B aux model.
+  # Dense 8B (not MoE). ~4.68GB weights at Q4_K_M. GPU-resident (~5.2GB total
+  # including KV at 32K ctx). Thinking mode on by default — callers MUST send
+  # /no_think prefix for latency-sensitive aux tasks. See INI template comments.
+  # VRAM budget: Qwen3.8-27B (~17,804 MiB @ 65536 ctx) + Qwen3-8B (~5,300 MiB
+  # @ 32K ctx) = ~23,104 MiB — fits in 24GB (24,576 MiB) with ~1.4GB headroom.
+  # llama-swap evicts Qwen3.8 when Qwen3-8B is requested (cannot co-reside).
+  # LRU eviction is the operative strategy; cold load ~5s for 5GB GGUF.
+  - id: Qwen3-8B-Q4_K_M
+    gguf_path: "{{ llm_models_dir }}/Qwen3-8B-Q4_K_M.gguf"
+    port: 8106
+    n_gpu_layers: -1       # -1 = auto-detect / all layers to GPU (~5GB, fits easily)
+    ctx_size: 32768
+    batch_size: 4096
+    ubatch_size: 4096
+    parallel: 1
+    cache_type: q4_0
+    flash_attn: true
+    sleep_idle_seconds: 60  # idle after 60s no requests
+
 # 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)
@@ -335,3 +355,6 @@ llm_swapmode_matrix_rows:
 
   - row: row4
     expr: "Phi-3.5-mini-instruct-Q8_0 & nomic-embed-text-v1.5"      # Mini + embed
+
+  - row: row5
+    expr: "Qwen3-8B-Q4_K_M & nomic-embed-text-v1.5"                  # Aux 8B + embed

ansible/roles/llm-inference-multimodel/templates/llama-server-router-preset.ini.j2

@@ -207,3 +207,46 @@ rope-scaling       = yarn
 rope-freq-scale    = 0.75
 load-on-startup    = true
 sleep-idle-seconds = -1
+
+; --- Aux model: Qwen3-8B-Q4_K_M -----------------------------------------------
+; Added t_c5cef2b2 (2026-08-19, War Machine).
+; Qwen3-8B is a dense 8B model (Qwen3 family, Alibaba) for aux tasks:
+; intent classification, query rewriting, structured extraction, tool-call
+; construction, and draft generation. Chosen over Llama-3.1-8B for superior
+; json_schema grammar support and stronger instruction-following.
+;
+; Architecture: 36 layers, 32 Q heads, 8 KV heads (GQA), 32K native context,
+; 131K via YaRN. Q4_K_M GGUF from bartowski/Qwen_Qwen3-8B-GGUF. ~4.68GB weights.
+;
+; ctx-size=32768: native training context (safe, no YaRN extension needed).
+; VRAM at 32K ctx with q4_0 KV: ~4.68GB weights + ~0.5GB KV ≈ 5.2GB total.
+; Fits comfortably on RTX 3090 24GB. Co-resident with nomic-embed (~84MB):
+; ~5.3GB total — well within 24GB budget even as a warm auxiliary model.
+;
+; THINKING MODE NOTE (JARVIS clarification, t_c5cef2b2):
+;   Qwen3-8B has thinking (chain-of-thought) mode ENABLED BY DEFAULT.
+;   For latency-sensitive aux tasks (routing, rewriting, structured extraction),
+;   disable at call time — NOT in this deployment config:
+;     - Prompt prefix: begin the user message with "/no_think"
+;     - Chat template override: pass enable_thinking=False in the template vars
+;       (llama.cpp Jinja2 template: {% if enable_thinking is false %} ... )
+;   Do NOT hardcode thinking=false here — this is a call-site concern.
+;   Thinking mode IS appropriate for complex multi-step tool plans and long-context
+;   summarization; leave that decision to the caller.
+;
+; Source: bartowski/Qwen_Qwen3-8B-GGUF (public, no auth). Stored locally as
+; Qwen3-8B-Q4_K_M.gguf (canonical name, trimmed from HF's Qwen_Qwen3-8B prefix).
+; n-gpu-layers=99: GPU (all layers). At ~5GB, fully GPU-resident with headroom.
+; flash-attn=true: Qwen3-8B uses standard attention; flash-attn is safe.
+; sleep-idle-seconds=60: evict after 60s idle to free GPU VRAM for Qwen3.8-27B.
+[Qwen3-8B-Q4_K_M]
+model        = {{ llm_models_dir }}/Qwen3-8B-Q4_K_M.gguf
+n-gpu-layers = 99
+ctx-size     = 32768
+flash-attn   = true
+cache-type-k = {{ llm_router_cache_type_k }}
+cache-type-v = {{ llm_router_cache_type_v }}
+batch-size   = {{ llm_router_batch_size }}
+ubatch-size  = {{ llm_router_ubatch_size }}
+parallel     = {{ llm_router_parallel }}
+