Wire in GPU optimizations and Gitea↔GitHub sync

Ollama config (all 3 setup scripts):
- Models updated from Qwen 2.5 → Qwen 3.5 series (Feb 2026)
- Auto-detect multi-GPU: TOTAL_VRAM = per-card × count
- KV cache quantization (q8_0) + flash attention enabled by default
- Context windows scaled by total VRAM (4K→128K)
- RAG server CHAT_MODEL now uses detected model variable

Gitea↔GitHub sync (new):
- gitea-github-sync.sh: bidirectional mirror with --init wizard
- Modes: --pull-only, --push-only, --list (dry run), --repo single
- Auto-discovers repos from both platforms via API
- Systemd timer: --install-timer [interval] for scheduled sync
- MCP tool: gitea_github_sync() for on-demand from Claude/WebUI
- Sync script mounted read-only into mcp-server container
- .env gets GITEA_URL variable for sync script
- curl added to mcp-server container deps

https://claude.ai/code/session_01PtYTPherSJaxDEVPgF6Nxu

local-ai-setup.sh

@@ -23,19 +23,31 @@ IS_UPDATE=false; [[ -f "$BASE/docker-compose.yml" ]] && IS_UPDATE=true
 # ── detect VRAM and set models accordingly ────────────────────────────────────
 VRAM_GB=$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits 2>/dev/null \
           | head -1 | awk '{printf "%d", $1/1024}' 2>/dev/null || echo "0")
+GPU_COUNT=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | wc -l || echo "0")
+TOTAL_VRAM=$((VRAM_GB * GPU_COUNT))
 
-if   [[ "$VRAM_GB" -ge 14 ]]; then
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:14b"
-    CTX=32768; TIER="16GB — 14B models + 32k context"
-elif [[ "$VRAM_GB" -ge 8 ]]; then
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:7b"
-    CTX=16384; TIER="8-16GB — 14B chat, 7B code, 16k context"
-elif [[ "$VRAM_GB" -ge 4 ]]; then
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
-    CTX=8192;  TIER="6GB — 7B models, 8k context"
+# Ollama optimization flags (stacked — see docs/gpu-setup-research.md)
+OLLAMA_KV_CACHE="q8_0"          # halves KV cache VRAM (q4_0 for aggressive)
+OLLAMA_FLASH="1"                # flash attention: less VRAM, no quality loss
+
+if   [[ "$TOTAL_VRAM" -ge 40 ]]; then
+    CHAT_MODEL="qwen3.5:27b";  CODE_MODEL="qwen3.5:27b"
+    CTX=131072; TIER="${TOTAL_VRAM}GB — 27B dense, 128K context"
+elif [[ "$TOTAL_VRAM" -ge 28 ]]; then
+    CHAT_MODEL="qwen3.5-35b-a3b";  CODE_MODEL="qwen3.5-35b-a3b"
+    CTX=131072; TIER="${TOTAL_VRAM}GB — 35B MoE, 128K context"
+elif [[ "$TOTAL_VRAM" -ge 14 ]]; then
+    CHAT_MODEL="qwen3.5-35b-a3b";  CODE_MODEL="qwen3.5-35b-a3b"
+    CTX=65536;  TIER="${TOTAL_VRAM}GB — 35B MoE + KV quant, 64K context"
+elif [[ "$TOTAL_VRAM" -ge 8 ]]; then
+    CHAT_MODEL="qwen3.5:9b";  CODE_MODEL="qwen3.5:9b"
+    CTX=32768;  TIER="${TOTAL_VRAM}GB — 9B dense, 32K context"
+elif [[ "$TOTAL_VRAM" -ge 4 ]]; then
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
+    CTX=16384;  TIER="${TOTAL_VRAM}GB — 4B models, 16K context"
 else
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
-    CTX=4096;  TIER="CPU-only — 7B models, 4k context"
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
+    CTX=4096;   OLLAMA_KV_CACHE="q4_0"; TIER="CPU-only — 4B models, 4K context"
 fi
 EMBED_MODEL="nomic-embed-text"
 
@@ -488,6 +500,7 @@ if [[ ! -f "$BASE/.env" ]]; then
 # Local AI Stack — edit to add your API tokens
 GITEA_TOKEN=your-gitea-token-here
 GITHUB_TOKEN=your-github-token-here
+GITEA_URL=http://$LOCAL_IP:3001
 ENV
     ok "Created .env"
 else
@@ -527,9 +540,11 @@ services:
     volumes: [ollama-models:/root/.ollama]
     environment:
       - OLLAMA_NUM_GPU=999
-      - OLLAMA_NUM_CTX=
+      - OLLAMA_NUM_CTX=$CTX
       - OLLAMA_KEEP_ALIVE=24h
       - OLLAMA_MAX_LOADED_MODELS=1
+      - OLLAMA_KV_CACHE_TYPE=$OLLAMA_KV_CACHE
+      - OLLAMA_FLASH_ATTENTION=$OLLAMA_FLASH
     deploy:
       resources:
         reservations:
@@ -586,7 +601,7 @@ services:
       - OLLAMA_URL=http://ollama:11434
       - CHROMA_URL=http://chromadb:8000
       - EMBED_MODEL=nomic-embed-text
-      - CHAT_MODEL=
+      - CHAT_MODEL=$CHAT_MODEL
     command: >
       bash -c "apt-get update -qq && apt-get install -y --no-install-recommends git &&
                pip install --no-cache-dir -r requirements.txt &&
@@ -605,6 +620,7 @@ services:
       - $BASE/repos:/repos
       - $BASE/mcp_server.py:/app/mcp_server.py
       - $BASE/mcp_requirements.txt:/app/mcp_requirements.txt
+      - $SCRIPT_DIR/gitea-github-sync.sh:/app/gitea-github-sync.sh:ro
     working_dir: /app
     env_file: $BASE/.env
     environment:
@@ -613,7 +629,7 @@ services:
       - GITEA_URL=http://gitea:3000
       - RAG_URL=http://rag-server:8001
     command: >
-      bash -c "apt-get update -qq && apt-get install -y --no-install-recommends git ripgrep &&
+      bash -c "apt-get update -qq && apt-get install -y --no-install-recommends git ripgrep curl &&
                pip install --no-cache-dir -r mcp_requirements.txt &&
                python mcp_server.py"
     depends_on: [rag-server]