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

gitea-github-sync.sh

@@ -0,0 +1,464 @@
+#!/usr/bin/env bash
+# =============================================================================
+# Gitea ↔ GitHub Mirror Sync
+#
+# Mirrors repos between your local Gitea and GitHub in both directions:
+#   GitHub → Gitea:  Pulls repos you own on GitHub into Gitea (backup/offline use)
+#   Gitea → GitHub:  Pushes Gitea repos to GitHub (remote backup)
+#
+# Usage:
+#   ./gitea-github-sync.sh                     — sync all configured repos
+#   ./gitea-github-sync.sh --pull-only         — GitHub → Gitea only
+#   ./gitea-github-sync.sh --push-only         — Gitea → GitHub only
+#   ./gitea-github-sync.sh --repo owner/name   — sync one specific repo
+#   ./gitea-github-sync.sh --list              — list what would sync (dry run)
+#   ./gitea-github-sync.sh --init              — interactive first-time setup
+#
+# Config:  ~/.config/gitea-github-sync/config
+# Tokens:  reads from .env in the same directory as this script (or $SYNC_ENV)
+#
+# Schedule: install the systemd timer with --install-timer
+#   ./gitea-github-sync.sh --install-timer     — every 6 hours (default)
+#   ./gitea-github-sync.sh --install-timer 1h  — custom interval
+#   ./gitea-github-sync.sh --remove-timer      — remove the timer
+# =============================================================================
+set -euo pipefail
+
+RED='\033[0;31m'; GREEN='\033[0;32m'; YELLOW='\033[1;33m'
+CYAN='\033[0;36m'; BOLD='\033[1m'; NC='\033[0m'
+info()  { echo -e "${CYAN}[sync]${NC}  $*"; }
+ok()    { echo -e "${GREEN}[ ok ]${NC}  $*"; }
+warn()  { echo -e "${YELLOW}[warn]${NC}  $*"; }
+err()   { echo -e "${RED}[err ]${NC}  $*" >&2; }
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CONFIG_DIR="${XDG_CONFIG_HOME:-$HOME/.config}/gitea-github-sync"
+CONFIG_FILE="$CONFIG_DIR/config"
+WORK_DIR="$CONFIG_DIR/repos"
+LOG_FILE="$CONFIG_DIR/sync.log"
+
+# ── load tokens from .env ───────────────────────────────────────────────────
+ENV_FILE="${SYNC_ENV:-$SCRIPT_DIR/.env}"
+if [[ -f "$ENV_FILE" ]]; then
+    # shellcheck disable=SC1090
+    set -a; source <(grep -E '^(GITEA_TOKEN|GITHUB_TOKEN|GITEA_URL)=' "$ENV_FILE" | sed 's/ *#.*//'); set +a
+fi
+
+GITEA_URL="${GITEA_URL:-http://localhost:3001}"
+GITEA_TOKEN="${GITEA_TOKEN:-}"
+GITHUB_TOKEN="${GITHUB_TOKEN:-}"
+
+# ── parse args ──────────────────────────────────────────────────────────────
+MODE="all"         # all | pull | push | list | init | install-timer | remove-timer
+SINGLE_REPO=""
+TIMER_INTERVAL="6h"
+
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --pull-only)      MODE="pull"; shift ;;
+        --push-only)      MODE="push"; shift ;;
+        --list)           MODE="list"; shift ;;
+        --init)           MODE="init"; shift ;;
+        --install-timer)  MODE="install-timer"; shift; [[ "${1:-}" =~ ^[0-9]+[smhd]$ ]] && { TIMER_INTERVAL="$1"; shift; } ;;
+        --remove-timer)   MODE="remove-timer"; shift ;;
+        --repo)           shift; SINGLE_REPO="${1:-}"; shift ;;
+        -h|--help)
+            sed -n '2,/^# =====/{ /^# =====/d; s/^# \?//p; }' "$0"; exit 0 ;;
+        *) err "Unknown arg: $1"; exit 1 ;;
+    esac
+done
+
+# ── helpers ─────────────────────────────────────────────────────────────────
+_gitea_api() {
+    local method="$1" path="$2"; shift 2
+    curl -sfL -X "$method" \
+        -H "Authorization: token $GITEA_TOKEN" \
+        -H "Content-Type: application/json" \
+        "$GITEA_URL/api/v1$path" "$@"
+}
+
+_github_api() {
+    local method="$1" path="$2"; shift 2
+    curl -sfL -X "$method" \
+        -H "Authorization: Bearer $GITHUB_TOKEN" \
+        -H "Accept: application/vnd.github+json" \
+        "https://api.github.com$path" "$@"
+}
+
+_log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" >> "$LOG_FILE"; }
+
+# ── config management ──────────────────────────────────────────────────────
+load_config() {
+    mkdir -p "$CONFIG_DIR" "$WORK_DIR"
+    GITHUB_USER=""
+    GITEA_USER=""
+    SYNC_REPOS=()           # explicit list (empty = auto-discover)
+    EXCLUDE_REPOS=()        # repos to skip
+    PUSH_PRIVATE=false      # push private Gitea repos to GitHub?
+    PULL_PRIVATE=true       # pull private GitHub repos to Gitea?
+    PULL_FORKS=false        # pull forked repos from GitHub?
+
+    if [[ -f "$CONFIG_FILE" ]]; then
+        # shellcheck disable=SC1090
+        source "$CONFIG_FILE"
+    fi
+}
+
+save_config() {
+    mkdir -p "$CONFIG_DIR"
+    cat > "$CONFIG_FILE" << EOF
+# Gitea-GitHub Sync — configuration
+# Generated $(date '+%Y-%m-%d %H:%M:%S')
+
+# GitHub username (for discovering repos to pull)
+GITHUB_USER="$GITHUB_USER"
+
+# Gitea username (for discovering repos to push)
+GITEA_USER="$GITEA_USER"
+
+# Explicit repo list — if set, only these sync. Format: owner/repo
+# Leave empty () to auto-discover from both platforms.
+SYNC_REPOS=($(printf '"%s" ' "${SYNC_REPOS[@]}"))
+
+# Repos to skip (pattern matched against owner/repo)
+EXCLUDE_REPOS=($(printf '"%s" ' "${EXCLUDE_REPOS[@]}"))
+
+# Push private Gitea repos to GitHub as private repos?
+PUSH_PRIVATE=$PUSH_PRIVATE
+
+# Pull private GitHub repos to Gitea?
+PULL_PRIVATE=$PULL_PRIVATE
+
+# Pull forked repos from GitHub?
+PULL_FORKS=$PULL_FORKS
+EOF
+    ok "Config saved: $CONFIG_FILE"
+}
+
+# ── init (first-time setup) ────────────────────────────────────────────────
+do_init() {
+    echo -e "\n${BOLD}Gitea ↔ GitHub Sync — First-Time Setup${NC}\n"
+
+    # Check tokens
+    if [[ -z "$GITEA_TOKEN" || "$GITEA_TOKEN" == "your-gitea-token-here" ]]; then
+        err "GITEA_TOKEN not set. Add it to $ENV_FILE first."
+        echo "  Generate at: $GITEA_URL/user/settings/applications"
+        exit 1
+    fi
+    if [[ -z "$GITHUB_TOKEN" || "$GITHUB_TOKEN" == "your-github-token-here" ]]; then
+        err "GITHUB_TOKEN not set. Add it to $ENV_FILE first."
+        echo "  Generate at: https://github.com/settings/tokens"
+        echo "  Scopes needed: repo (full control)"
+        exit 1
+    fi
+
+    # Discover usernames
+    info "Detecting GitHub user..."
+    GITHUB_USER=$(_github_api GET /user | python3 -c "import sys,json; print(json.load(sys.stdin)['login'])" 2>/dev/null) \
+        || { err "Failed to reach GitHub API. Check GITHUB_TOKEN."; exit 1; }
+    ok "GitHub user: $GITHUB_USER"
+
+    info "Detecting Gitea user..."
+    GITEA_USER=$(_gitea_api GET /user | python3 -c "import sys,json; print(json.load(sys.stdin)['login'])" 2>/dev/null) \
+        || { err "Failed to reach Gitea API. Check GITEA_TOKEN and GITEA_URL ($GITEA_URL)."; exit 1; }
+    ok "Gitea user: $GITEA_USER"
+
+    # Ask about sync scope
+    echo ""
+    read -rp "Pull private GitHub repos to Gitea? [Y/n] " ans
+    PULL_PRIVATE=true; [[ "${ans,,}" == "n" ]] && PULL_PRIVATE=false
+
+    read -rp "Pull forked repos from GitHub? [y/N] " ans
+    PULL_FORKS=false; [[ "${ans,,}" == "y" ]] && PULL_FORKS=true
+
+    read -rp "Push private Gitea repos to GitHub? [y/N] " ans
+    PUSH_PRIVATE=false; [[ "${ans,,}" == "y" ]] && PUSH_PRIVATE=true
+
+    save_config
+
+    echo ""
+    info "Run '$(basename "$0") --list' to preview what would sync."
+    info "Run '$(basename "$0")' to sync now."
+    info "Run '$(basename "$0") --install-timer' to sync automatically."
+}
+
+# ── discover repos ─────────────────────────────────────────────────────────
+get_github_repos() {
+    local page=1 repos=()
+    while true; do
+        local batch
+        batch=$(_github_api GET "/user/repos?per_page=100&page=$page&affiliation=owner" \
+            | python3 -c "
+import sys, json
+for r in json.load(sys.stdin):
+    if r.get('fork') and not $($PULL_FORKS && echo True || echo False):
+        continue
+    if r.get('private') and not $($PULL_PRIVATE && echo True || echo False):
+        continue
+    print(r['full_name'] + '|' + r['clone_url'] + '|' + str(r.get('private',False)).lower())
+" 2>/dev/null) || break
+        [[ -z "$batch" ]] && break
+        while IFS= read -r line; do repos+=("$line"); done <<< "$batch"
+        ((page++))
+    done
+    printf '%s\n' "${repos[@]}"
+}
+
+get_gitea_repos() {
+    local page=1 repos=()
+    while true; do
+        local batch
+        batch=$(_gitea_api GET "/repos/search?limit=50&page=$page" \
+            | python3 -c "
+import sys, json
+for r in json.load(sys.stdin).get('data', []):
+    if r.get('private') and not $($PUSH_PRIVATE && echo True || echo False):
+        continue
+    print(r['full_name'] + '|' + r['clone_url'] + '|' + str(r.get('private',False)).lower())
+" 2>/dev/null) || break
+        [[ -z "$batch" ]] && break
+        while IFS= read -r line; do repos+=("$line"); done <<< "$batch"
+        ((page++))
+    done
+    printf '%s\n' "${repos[@]}"
+}
+
+is_excluded() {
+    local repo="$1"
+    for pat in "${EXCLUDE_REPOS[@]}"; do
+        [[ "$repo" == $pat ]] && return 0
+    done
+    return 1
+}
+
+# ── sync: GitHub → Gitea (pull) ───────────────────────────────────────────
+sync_github_to_gitea() {
+    local full_name="$1" clone_url="$2" is_private="$3"
+    local repo_name="${full_name#*/}"
+    local local_path="$WORK_DIR/$full_name"
+
+    # Clone or fetch from GitHub
+    if [[ -d "$local_path" ]]; then
+        info "Fetching $full_name from GitHub..."
+        git -C "$local_path" fetch --all --prune --quiet 2>/dev/null || {
+            err "Failed to fetch $full_name"; return 1; }
+    else
+        info "Cloning $full_name from GitHub..."
+        mkdir -p "$(dirname "$local_path")"
+        local auth_url="${clone_url/https:\/\//https:\/\/$GITHUB_TOKEN@}"
+        git clone --bare --quiet "$auth_url" "$local_path" 2>/dev/null || {
+            err "Failed to clone $full_name"; return 1; }
+    fi
+
+    # Ensure repo exists on Gitea
+    local gitea_check
+    gitea_check=$(_gitea_api GET "/repos/$GITEA_USER/$repo_name" 2>/dev/null) || true
+    if ! echo "$gitea_check" | python3 -c "import sys,json; json.load(sys.stdin)['id']" &>/dev/null; then
+        info "Creating $repo_name on Gitea..."
+        _gitea_api POST "/user/repos" \
+            -d "{\"name\":\"$repo_name\",\"private\":$is_private,\"description\":\"Mirror of $full_name from GitHub\"}" \
+            >/dev/null || { err "Failed to create $repo_name on Gitea"; return 1; }
+    fi
+
+    # Push to Gitea
+    local gitea_push_url="${GITEA_URL/https:\/\//https:\/\/$GITEA_USER:$GITEA_TOKEN@}"
+    gitea_push_url="${gitea_push_url/http:\/\//http:\/\/$GITEA_USER:$GITEA_TOKEN@}"
+    gitea_push_url="$gitea_push_url/$GITEA_USER/$repo_name.git"
+
+    git -C "$local_path" push --mirror "$gitea_push_url" --quiet 2>/dev/null || {
+        err "Failed to push $full_name to Gitea"; return 1; }
+    ok "GitHub → Gitea: $full_name"
+    _log "PULL $full_name OK"
+}
+
+# ── sync: Gitea → GitHub (push) ───────────────────────────────────────────
+sync_gitea_to_github() {
+    local full_name="$1" clone_url="$2" is_private="$3"
+    local repo_name="${full_name#*/}"
+    local local_path="$WORK_DIR/gitea/$full_name"
+
+    # Clone or fetch from Gitea
+    local gitea_auth_url="${clone_url/https:\/\//https:\/\/$GITEA_USER:$GITEA_TOKEN@}"
+    gitea_auth_url="${gitea_auth_url/http:\/\//http:\/\/$GITEA_USER:$GITEA_TOKEN@}"
+
+    if [[ -d "$local_path" ]]; then
+        info "Fetching $full_name from Gitea..."
+        git -C "$local_path" fetch --all --prune --quiet 2>/dev/null || {
+            err "Failed to fetch $full_name from Gitea"; return 1; }
+    else
+        info "Cloning $full_name from Gitea..."
+        mkdir -p "$(dirname "$local_path")"
+        git clone --bare --quiet "$gitea_auth_url" "$local_path" 2>/dev/null || {
+            err "Failed to clone $full_name from Gitea"; return 1; }
+    fi
+
+    # Ensure repo exists on GitHub
+    local gh_check
+    gh_check=$(_github_api GET "/repos/$GITHUB_USER/$repo_name" 2>/dev/null) || true
+    if ! echo "$gh_check" | python3 -c "import sys,json; json.load(sys.stdin)['id']" &>/dev/null; then
+        info "Creating $repo_name on GitHub..."
+        _github_api POST "/user/repos" \
+            -d "{\"name\":\"$repo_name\",\"private\":$is_private,\"description\":\"Mirror from Gitea\"}" \
+            >/dev/null || { err "Failed to create $repo_name on GitHub"; return 1; }
+    fi
+
+    # Push to GitHub
+    local github_push_url="https://$GITHUB_TOKEN@github.com/$GITHUB_USER/$repo_name.git"
+    git -C "$local_path" push --mirror "$github_push_url" --quiet 2>/dev/null || {
+        err "Failed to push $full_name to GitHub"; return 1; }
+    ok "Gitea → GitHub: $full_name"
+    _log "PUSH $full_name OK"
+}
+
+# ── list (dry run) ─────────────────────────────────────────────────────────
+do_list() {
+    echo -e "\n${BOLD}Repos that would sync:${NC}\n"
+
+    if [[ ${#SYNC_REPOS[@]} -gt 0 ]]; then
+        echo -e "${CYAN}Explicit list:${NC}"
+        printf '  %s\n' "${SYNC_REPOS[@]}"
+    else
+        if [[ "$MODE" != "push" ]]; then
+            echo -e "${CYAN}GitHub → Gitea (pull):${NC}"
+            get_github_repos | while IFS='|' read -r name url priv; do
+                is_excluded "$name" && echo "  $name (excluded)" && continue
+                echo "  $name $([ "$priv" = "true" ] && echo "[private]")"
+            done
+        fi
+        echo ""
+        if [[ "$MODE" != "pull" ]]; then
+            echo -e "${CYAN}Gitea → GitHub (push):${NC}"
+            get_gitea_repos | while IFS='|' read -r name url priv; do
+                is_excluded "$name" && echo "  $name (excluded)" && continue
+                echo "  $name $([ "$priv" = "true" ] && echo "[private]")"
+            done
+        fi
+    fi
+    echo ""
+}
+
+# ── main sync ──────────────────────────────────────────────────────────────
+do_sync() {
+    local pull_count=0 push_count=0 fail_count=0
+
+    _log "=== Sync started (mode=$MODE) ==="
+
+    # GitHub → Gitea
+    if [[ "$MODE" == "all" || "$MODE" == "pull" ]]; then
+        info "Discovering GitHub repos..."
+        while IFS='|' read -r name url priv; do
+            [[ -z "$name" ]] && continue
+            [[ -n "$SINGLE_REPO" && "$name" != "$SINGLE_REPO" ]] && continue
+            is_excluded "$name" && continue
+            if sync_github_to_gitea "$name" "$url" "$priv"; then
+                ((pull_count++))
+            else
+                ((fail_count++))
+            fi
+        done < <(get_github_repos)
+    fi
+
+    # Gitea → GitHub
+    if [[ "$MODE" == "all" || "$MODE" == "push" ]]; then
+        info "Discovering Gitea repos..."
+        while IFS='|' read -r name url priv; do
+            [[ -z "$name" ]] && continue
+            [[ -n "$SINGLE_REPO" && "${name#*/}" != "${SINGLE_REPO#*/}" ]] && continue
+            is_excluded "$name" && continue
+            # Skip repos that came from GitHub (already mirrored)
+            local repo_name="${name#*/}"
+            if [[ -d "$WORK_DIR/$GITHUB_USER/$repo_name" ]]; then
+                info "Skipping $name (already a GitHub mirror)"
+                continue
+            fi
+            if sync_gitea_to_github "$name" "$url" "$priv"; then
+                ((push_count++))
+            else
+                ((fail_count++))
+            fi
+        done < <(get_gitea_repos)
+    fi
+
+    echo ""
+    ok "Sync complete: ${pull_count} pulled, ${push_count} pushed, ${fail_count} failed"
+    _log "=== Sync complete: pull=$pull_count push=$push_count fail=$fail_count ==="
+}
+
+# ── systemd timer ──────────────────────────────────────────────────────────
+install_timer() {
+    local service_file="/etc/systemd/system/gitea-github-sync.service"
+    local timer_file="/etc/systemd/system/gitea-github-sync.timer"
+    local script_path
+    script_path="$(readlink -f "$0")"
+
+    info "Installing systemd timer (interval: $TIMER_INTERVAL)..."
+
+    sudo tee "$service_file" > /dev/null << EOF
+[Unit]
+Description=Gitea-GitHub Mirror Sync
+After=network-online.target docker.service
+Wants=network-online.target
+
+[Service]
+Type=oneshot
+User=$USER
+ExecStart=$script_path
+Environment=HOME=$HOME
+StandardOutput=append:$LOG_FILE
+StandardError=append:$LOG_FILE
+EOF
+
+    sudo tee "$timer_file" > /dev/null << EOF
+[Unit]
+Description=Gitea-GitHub Sync Timer
+
+[Timer]
+OnBootSec=5min
+OnUnitActiveSec=$TIMER_INTERVAL
+Persistent=true
+
+[Install]
+WantedBy=timers.target
+EOF
+
+    sudo systemctl daemon-reload
+    sudo systemctl enable --now gitea-github-sync.timer
+    ok "Timer installed: every $TIMER_INTERVAL"
+    ok "Check status: systemctl status gitea-github-sync.timer"
+    ok "Run now: sudo systemctl start gitea-github-sync.service"
+    ok "Logs: $LOG_FILE"
+}
+
+remove_timer() {
+    info "Removing systemd timer..."
+    sudo systemctl disable --now gitea-github-sync.timer 2>/dev/null || true
+    sudo rm -f /etc/systemd/system/gitea-github-sync.{service,timer}
+    sudo systemctl daemon-reload
+    ok "Timer removed"
+}
+
+# ── preflight checks ──────────────────────────────────────────────────────
+preflight() {
+    local ok=true
+    if [[ -z "$GITEA_TOKEN" || "$GITEA_TOKEN" == "your-gitea-token-here" ]]; then
+        err "GITEA_TOKEN not set. Edit $ENV_FILE"; ok=false
+    fi
+    if [[ -z "$GITHUB_TOKEN" || "$GITHUB_TOKEN" == "your-github-token-here" ]]; then
+        err "GITHUB_TOKEN not set. Edit $ENV_FILE"; ok=false
+    fi
+    if [[ -z "$GITEA_USER" || -z "$GITHUB_USER" ]]; then
+        err "Run --init first to configure usernames"; ok=false
+    fi
+    $ok || exit 1
+}
+
+# ── main ───────────────────────────────────────────────────────────────────
+load_config
+
+case "$MODE" in
+    init)           do_init ;;
+    install-timer)  install_timer ;;
+    remove-timer)   remove_timer ;;
+    list)           preflight; do_list ;;
+    *)              preflight; do_sync ;;
+esac

laptop_full_setup.sh

@@ -39,30 +39,42 @@ LOCAL_IP=$(ip route get 1.1.1.1 2>/dev/null | grep -oP 'src \K\S+' \
 
 # Models (defaults, may be adjusted below based on VRAM)
 EMBED_MODEL="nomic-embed-text"
-CHAT_MODEL="qwen2.5:14b"
+CHAT_MODEL="qwen3.5:9b"
-CODE_MODEL="qwen2.5-coder:7b"
+CODE_MODEL="qwen3.5:9b"
-FAST_MODEL="qwen2.5:7b"
+FAST_MODEL="qwen3.5:4b"
 
 # ── detect GPU ────────────────────────────────────────────────────────────────
 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")
 GPU_NAME=$(nvidia-smi --query-gpu=name --format=csv,noheader 2>/dev/null | head -1 || echo "None")
+TOTAL_VRAM=$((VRAM_GB * GPU_COUNT))
 
-if   [[ "$VRAM_GB" -ge 14 ]]; then
+# Ollama optimization flags (stacked — see docs/gpu-setup-research.md)
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:14b"
+OLLAMA_KV_CACHE="q8_0"          # halves KV cache VRAM (q4_0 for aggressive)
-    GPU_TIER="16GB VRAM — 14B models"
+OLLAMA_FLASH="1"                # flash attention: less VRAM, no quality loss
-elif [[ "$VRAM_GB" -ge 8 ]]; then
+
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:7b"
+if   [[ "$TOTAL_VRAM" -ge 40 ]]; then
-    GPU_TIER="8GB VRAM — 14B chat, 7B code"
+    CHAT_MODEL="qwen3.5:27b";  CODE_MODEL="qwen3.5:27b"
-elif [[ "$VRAM_GB" -ge 4 ]]; then
+    CTX=131072; GPU_TIER="${TOTAL_VRAM}GB VRAM — 27B dense, 128K context"
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
+elif [[ "$TOTAL_VRAM" -ge 28 ]]; then
-    GPU_TIER="6GB VRAM — 7B models"
+    CHAT_MODEL="qwen3.5-35b-a3b";  CODE_MODEL="qwen3.5-35b-a3b"
-elif [[ "$VRAM_GB" -gt 0 ]]; then
+    CTX=131072; GPU_TIER="${TOTAL_VRAM}GB VRAM — 35B MoE, 128K context"
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
+elif [[ "$TOTAL_VRAM" -ge 14 ]]; then
-    GPU_TIER="${VRAM_GB}GB VRAM — 7B models"
+    CHAT_MODEL="qwen3.5-35b-a3b";  CODE_MODEL="qwen3.5-35b-a3b"
+    CTX=65536;  GPU_TIER="${TOTAL_VRAM}GB VRAM — 35B MoE + KV quant, 64K context"
+elif [[ "$TOTAL_VRAM" -ge 8 ]]; then
+    CHAT_MODEL="qwen3.5:9b";  CODE_MODEL="qwen3.5:9b"
+    CTX=32768;  GPU_TIER="${TOTAL_VRAM}GB VRAM — 9B dense, 32K context"
+elif [[ "$TOTAL_VRAM" -ge 4 ]]; then
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
+    CTX=16384;  GPU_TIER="${TOTAL_VRAM}GB VRAM — 4B models, 16K context"
+elif [[ "$TOTAL_VRAM" -gt 0 ]]; then
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
+    CTX=8192;   GPU_TIER="${TOTAL_VRAM}GB VRAM — 4B models"
 else
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
-    GPU_TIER="CPU only — 7B models (slow)"
+    CTX=4096;   OLLAMA_KV_CACHE="q4_0"; GPU_TIER="CPU only — 4B models (slow)"
 fi
 
 # ── new vs update ─────────────────────────────────────────────────────────────
@@ -485,19 +497,19 @@ if $INSTALL_AI; then
             printf "  %-4s  %-8s  %-42s  %s\n" "3)" "22B"  "phi4:14b + codestral:22b"               "$(speed_label 13)"
             printf "  %-4s  %-8s  %-42s  %s\n" "4)" "70B"  "llama3.3:70b + codestral:22b"           "$(speed_label 41)"
             ;;
-        2)  # Performance
+        2)  # Performance (Qwen 3.5 — Feb 2026)
-            _TIER_NAMES=(7B 14B 32B 72B)
+            _TIER_NAMES=(4B 9B 35B 27B)
-            printf "  %-4s  %-8s  %-42s  %s\n" "1)" "7B"   "qwen2.5:7b + qwen2.5-coder:7b"         "$(speed_label 4)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "1)" "4B"   "qwen3.5:4b (chat+code)"                "$(speed_label 2)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "2)" "14B"  "qwen2.5:14b + qwen2.5-coder:14b"       "$(speed_label 9)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "2)" "9B"   "qwen3.5:9b (chat+code)"                "$(speed_label 5)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "3)" "32B"  "qwen2.5:14b + qwen2.5-coder:32b"       "$(speed_label 19)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "3)" "35B"  "qwen3.5-35b-a3b (MoE, 3B active)"      "$(speed_label 12)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "4)" "72B"  "qwen2.5:72b + qwen2.5-coder:32b"       "$(speed_label 41)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "4)" "27B"  "qwen3.5:27b (dense, A- quality)"        "$(speed_label 17)"
             ;;
-        3)  # Mixed
+        3)  # Mixed (Western chat + Qwen 3.5 code)
-            _TIER_NAMES=(7B 14B 32B 70B)
+            _TIER_NAMES=(7B 14B 35B 70B)
-            printf "  %-4s  %-8s  %-42s  %s\n" "1)" "7B"   "mistral:7b + qwen2.5-coder:7b"         "$(speed_label 4)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "1)" "7B"   "mistral:7b + qwen3.5:4b"               "$(speed_label 4)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "2)" "14B"  "phi4:14b + qwen2.5-coder:14b"          "$(speed_label 9)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "2)" "14B"  "phi4:14b + qwen3.5:9b"                 "$(speed_label 9)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "3)" "32B"  "phi4:14b + qwen2.5-coder:32b"          "$(speed_label 19)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "3)" "35B"  "phi4:14b + qwen3.5-35b-a3b"            "$(speed_label 19)"
-            printf "  %-4s  %-8s  %-42s  %s\n" "4)" "70B"  "llama3.3:70b + qwen2.5-coder:32b"      "$(speed_label 41)"
+            printf "  %-4s  %-8s  %-42s  %s\n" "4)" "70B"  "llama3.3:70b + qwen3.5-35b-a3b"        "$(speed_label 41)"
             ;;
     esac
 
@@ -535,16 +547,16 @@ if $INSTALL_AI; then
             1:14B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="starcoder2:15b";  REASON_MODEL="phi4:14b" ;;
             1:22B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="codestral:22b";   REASON_MODEL="phi4:14b" ;;
             1:70B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="llama3.3:70b"; CODE_MODEL="codestral:22b";   REASON_MODEL="llama3.3:70b" ;;
-            # Performance-first
+            # Performance-first (Qwen 3.5)
-            2:7B)   FAST_MODEL="qwen2.5:7b";   CHAT_MODEL="qwen2.5:7b";   CODE_MODEL="qwen2.5-coder:7b";   REASON_MODEL="" ;;
+            2:4B)   FAST_MODEL="qwen3.5:4b";          CHAT_MODEL="qwen3.5:4b";          CODE_MODEL="qwen3.5:4b";          REASON_MODEL="" ;;
-            2:14B)  FAST_MODEL="qwen2.5:7b";   CHAT_MODEL="qwen2.5:14b";  CODE_MODEL="qwen2.5-coder:14b";  REASON_MODEL="deepseek-r1:14b" ;;
+            2:9B)   FAST_MODEL="qwen3.5:4b";          CHAT_MODEL="qwen3.5:9b";          CODE_MODEL="qwen3.5:9b";          REASON_MODEL="" ;;
-            2:32B)  FAST_MODEL="qwen2.5:7b";   CHAT_MODEL="qwen2.5:14b";  CODE_MODEL="qwen2.5-coder:32b";  REASON_MODEL="deepseek-r1:14b" ;;
+            2:35B)  FAST_MODEL="qwen3.5:4b";          CHAT_MODEL="qwen3.5-35b-a3b";     CODE_MODEL="qwen3.5-35b-a3b";     REASON_MODEL="" ;;
-            2:72B)  FAST_MODEL="qwen2.5:7b";   CHAT_MODEL="qwen2.5:72b";  CODE_MODEL="qwen2.5-coder:32b";  REASON_MODEL="deepseek-r1:14b" ;;
+            2:27B)  FAST_MODEL="qwen3.5:9b";          CHAT_MODEL="qwen3.5:27b";         CODE_MODEL="qwen3.5:27b";         REASON_MODEL="" ;;
-            # Mixed
+            # Mixed (Western chat + Qwen 3.5 code)
-            3:7B)   FAST_MODEL="mistral:7b";   CHAT_MODEL="mistral:7b";   CODE_MODEL="qwen2.5-coder:7b";   REASON_MODEL="" ;;
+            3:7B)   FAST_MODEL="mistral:7b";   CHAT_MODEL="mistral:7b";   CODE_MODEL="qwen3.5:4b";          REASON_MODEL="" ;;
-            3:14B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="qwen2.5-coder:14b";  REASON_MODEL="phi4:14b" ;;
+            3:14B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="qwen3.5:9b";          REASON_MODEL="phi4:14b" ;;
-            3:32B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="qwen2.5-coder:32b";  REASON_MODEL="phi4:14b" ;;
+            3:35B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="phi4:14b";     CODE_MODEL="qwen3.5-35b-a3b";     REASON_MODEL="phi4:14b" ;;
-            3:70B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="llama3.3:70b"; CODE_MODEL="qwen2.5-coder:32b";  REASON_MODEL="llama3.3:70b" ;;
+            3:70B)  FAST_MODEL="mistral:7b";   CHAT_MODEL="llama3.3:70b"; CODE_MODEL="qwen3.5-35b-a3b";     REASON_MODEL="llama3.3:70b" ;;
             *)
                 warn "Unrecognised tier '$TIER_PICK' — keeping detected defaults"
                 ;;
@@ -763,8 +775,11 @@ WEBUI_URL=
 # Gitea — generate at http://$LOCAL_IP:3001/user/settings/applications
 GITEA_TOKEN=your-gitea-token-here
 
-# GitHub — optional, for GitHub API access via MCP
+# GitHub — optional, for GitHub API access via MCP and Gitea↔GitHub sync
 GITHUB_TOKEN=your-github-token-here
+
+# Gitea URL — used by sync script (default: http://localhost:3001)
+GITEA_URL=http://$LOCAL_IP:3001
 ENV
     ok "Created .env — add your tokens before using MCP Gitea/GitHub tools"
 else
@@ -816,10 +831,12 @@ services:
     volumes:
 ${OLLAMA_VOLUME_LINE}
     environment:
-      - OLLAMA_NUM_GPU=999       # use all available VRAM (auto-detects GPU size)
+      - OLLAMA_NUM_GPU=999              # use all available VRAM (auto-detects GPU size)
-      - OLLAMA_NUM_CTX=8192      # lower to 4096 if you hit OOM
+      - OLLAMA_NUM_CTX=$CTX             # auto-set by detected VRAM
       - OLLAMA_KEEP_ALIVE=24h
       - OLLAMA_MAX_LOADED_MODELS=1
+      - OLLAMA_KV_CACHE_TYPE=$OLLAMA_KV_CACHE  # q8_0 halves KV cache; q4_0 = 1/3 size
+      - OLLAMA_FLASH_ATTENTION=$OLLAMA_FLASH    # tiled attention: less VRAM, no quality loss
     deploy:
       resources:
         reservations:
@@ -885,7 +902,7 @@ ${OLLAMA_VOLUME_LINE}
       - OLLAMA_URL=http://ollama:11434
       - CHROMA_URL=http://chromadb:8000
       - EMBED_MODEL=nomic-embed-text
-      - CHAT_MODEL=qwen2.5:14b
+      - CHAT_MODEL=$CHAT_MODEL
       - PAPERS_DIR=/papers
       - REPOS_DIR=/repos
     command: >
@@ -913,6 +930,7 @@ ${OLLAMA_VOLUME_LINE}
       - $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:
@@ -922,7 +940,7 @@ ${OLLAMA_VOLUME_LINE}
       - RAG_URL=http://rag-server:8001
     command: >
       bash -c "apt-get update -qq &&
-               apt-get install -y --no-install-recommends git ripgrep &&
+               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:
@@ -1360,6 +1378,14 @@ if $INSTALL_AI; then
     echo -e "  ${YELLOW}Add API tokens to:${NC}  $BASE/.env"
     $SVC_RAG && echo -e "  ${YELLOW}Drop PDFs into:${NC}     $BASE/papers/"
     echo -e "  ${YELLOW}Your workspace:${NC}     $BASE/workspace/"
+    $SVC_GITEA && {
+        echo ""
+        echo -e "  ${YELLOW}Gitea ↔ GitHub sync:${NC}"
+        echo "    First time:   $SCRIPT_DIR/gitea-github-sync.sh --init"
+        echo "    Sync now:     $SCRIPT_DIR/gitea-github-sync.sh"
+        echo "    Auto (6h):    $SCRIPT_DIR/gitea-github-sync.sh --install-timer"
+        echo "    Via MCP:      gitea_github_sync(mode='all')"
+    }
 fi
 if $SVC_KIWIX && [[ "$ZIM_CHOICE" == "3" ]]; then
     echo -e "  ${YELLOW}ZIM downloads:${NC}      ./kiwix_download.sh  (not started)"

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   [[ "$VRAM_GB" -ge 14 ]]; then
+# Ollama optimization flags (stacked — see docs/gpu-setup-research.md)
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:14b"
+OLLAMA_KV_CACHE="q8_0"          # halves KV cache VRAM (q4_0 for aggressive)
-    CTX=32768; TIER="16GB — 14B models + 32k context"
+OLLAMA_FLASH="1"                # flash attention: less VRAM, no quality loss
-elif [[ "$VRAM_GB" -ge 8 ]]; then
+
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:7b"
+if   [[ "$TOTAL_VRAM" -ge 40 ]]; then
-    CTX=16384; TIER="8-16GB — 14B chat, 7B code, 16k context"
+    CHAT_MODEL="qwen3.5:27b";  CODE_MODEL="qwen3.5:27b"
-elif [[ "$VRAM_GB" -ge 4 ]]; then
+    CTX=131072; TIER="${TOTAL_VRAM}GB — 27B dense, 128K context"
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
+elif [[ "$TOTAL_VRAM" -ge 28 ]]; then
-    CTX=8192;  TIER="6GB — 7B models, 8k context"
+    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"
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
-    CTX=4096;  TIER="CPU-only — 7B models, 4k context"
+    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]

mcp_server.py

@@ -180,6 +180,25 @@ def github_api(method: str, endpoint: str, body: str = "") -> str:
     except Exception:
         return r.text
 
+# ── Gitea ↔ GitHub sync ──────────────────────────────────────────────────────
+@mcp.tool()
+def gitea_github_sync(mode: str = "all", repo: str = "") -> str:
+    """Run Gitea↔GitHub mirror sync. mode: all|pull|push|list. repo: optional owner/name."""
+    cmd = ["/app/gitea-github-sync.sh"]
+    if mode == "pull":   cmd.append("--pull-only")
+    elif mode == "push": cmd.append("--push-only")
+    elif mode == "list": cmd.append("--list")
+    if repo:
+        cmd.extend(["--repo", repo])
+    try:
+        r = subprocess.run(cmd, capture_output=True, text=True, timeout=600,
+                           env={**os.environ, "SYNC_ENV": "/app/.env"})
+        return (r.stdout + r.stderr).strip() or "Sync completed (no output)"
+    except subprocess.TimeoutExpired:
+        return "Sync timed out after 10 minutes"
+    except Exception as e:
+        return f"Sync failed: {e}"
+
 # ── RAG ingest ────────────────────────────────────────────────────────────────
 @mcp.tool()
 def ingest_repo(url: str, name: str = "", branch: str = "main") -> str:

ubuntu-post-install.sh

@@ -1106,19 +1106,31 @@ 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   [[ "$VRAM_GB" -ge 14 ]]; then
+# Ollama optimization flags (stacked — see docs/gpu-setup-research.md)
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:14b"
+OLLAMA_KV_CACHE="q8_0"          # halves KV cache VRAM (q4_0 for aggressive)
-    CTX=32768; TIER="16GB — 14B models + 32k context"
+OLLAMA_FLASH="1"                # flash attention: less VRAM, no quality loss
-elif [[ "$VRAM_GB" -ge 8 ]]; then
+
-    CHAT_MODEL="qwen2.5:14b"; CODE_MODEL="qwen2.5-coder:7b"
+if   [[ "$TOTAL_VRAM" -ge 40 ]]; then
-    CTX=16384; TIER="8-16GB — 14B chat, 7B code, 16k context"
+    CHAT_MODEL="qwen3.5:27b";  CODE_MODEL="qwen3.5:27b"
-elif [[ "$VRAM_GB" -ge 4 ]]; then
+    CTX=131072; TIER="${TOTAL_VRAM}GB — 27B dense, 128K context"
-    CHAT_MODEL="qwen2.5:7b"; CODE_MODEL="qwen2.5-coder:7b"
+elif [[ "$TOTAL_VRAM" -ge 28 ]]; then
-    CTX=8192;  TIER="6GB — 7B models, 8k context"
+    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"
+    CHAT_MODEL="qwen3.5:4b";  CODE_MODEL="qwen3.5:4b"
-    CTX=4096;  TIER="CPU-only — 7B models, 4k context"
+    CTX=4096;   OLLAMA_KV_CACHE="q4_0"; TIER="CPU-only — 4B models, 4K context"
 fi
 EMBED_MODEL="nomic-embed-text"
 
@@ -1611,9 +1623,11 @@ 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: