build.sh: serve Q4_1 on port 20000 and configure pi coding agent

build.sh

@@ -1,9 +1,185 @@
 #!/bin/bash
-# Build llama.cpp at patches/BASE_COMMIT with the flint CPU patch (Xeon with AVX512-VNNI + AMX).
-set -e
-git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
-git fetch --depth 1 origin $(cat ../patches/BASE_COMMIT) && git checkout $(cat ../patches/BASE_COMMIT)
-git apply ../patches/flint-cpu-opt.patch
-cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON -DGGML_CUDA=OFF -DLLAMA_CURL=OFF
-cmake --build build -j$(nproc) --target llama-cli llama-server llama-bench llama-perplexity
-# run: taskset -c 0-15 build/bin/llama-cli -m flint-UD-Q4_K_XL.gguf -t 16 -fa on -lm mlock
+# Build llama.cpp at patches/BASE_COMMIT with the flint CPU patch (Xeon with AVX512-VNNI + AMX),
+# then serve the Q4_1 model on http://127.0.0.1:20000 and point the pi coding agent at it.
+#
+# Overridable environment variables:
+#   MODEL_FILE    (default flint-int4-q4_1.gguf)
+#   LLAMA_PORT    (default 20000)
+#   MODEL_ALIAS   (default flint-q4_1)
+#   CONTEXT_SIZE  (default 32768)
+#   LLAMA_THREADS (default nproc)
+#   LLAMA_HOST    (default 127.0.0.1)
+set -euo pipefail
+
+ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+cd "$ROOT"
+
+MODEL_FILE="${MODEL_FILE:-flint-int4-q4_1.gguf}"
+MODEL_REPO="${MODEL_REPO:-HelloSun/flint_cpu_opt}"
+LLAMA_PORT="${LLAMA_PORT:-20000}"
+MODEL_ALIAS="${MODEL_ALIAS:-flint-q4_1}"
+CONTEXT_SIZE="${CONTEXT_SIZE:-32768}"
+LLAMA_THREADS="${LLAMA_THREADS:-$(nproc 2>/dev/null || echo 4)}"
+LLAMA_HOST="${LLAMA_HOST:-127.0.0.1}"
+MODEL_PATH="$ROOT/$MODEL_FILE"
+SERVER_BIN="$ROOT/llama.cpp/build/bin/llama-server"
+PIDFILE="$HOME/llama-server.pid"
+LOGFILE="$HOME/llama-server.log"
+PI_MODELS_JSON="$HOME/.pi/agent/models.json"
+PI_PROVIDER_ID="llama-cpp-local"
+PI_MODEL_ID="$MODEL_ALIAS"
+
+info() { echo -e "\033[36m[build.sh]\033[0m $*"; }
+ok()   { echo -e "\033[32m[build.sh]\033[0m $*"; }
+warn() { echo -e "\033[33m[build.sh]\033[0m $*" >&2; }
+
+mkdir -p "$HOME/.local/bin" "$(dirname "$PI_MODELS_JSON")"
+
+# ── 1. Build llama.cpp (skip when the binary is already there) ────────────────
+if [[ ! -x "$SERVER_BIN" ]]; then
+    if [[ ! -f "$ROOT/patches/BASE_COMMIT" || ! -f "$ROOT/patches/flint-cpu-opt.patch" ]]; then
+        warn "缺少 patches/BASE_COMMIT 或 patches/flint-cpu-opt.patch，無法套用 flint patch"
+        exit 1
+    fi
+    if [[ ! -d "$ROOT/llama.cpp/.git" ]]; then
+        info "clone llama.cpp"
+        git clone https://github.com/ggml-org/llama.cpp "$ROOT/llama.cpp"
+    fi
+    cd "$ROOT/llama.cpp"
+    BASE_COMMIT="$(cat "$ROOT/patches/BASE_COMMIT")"
+    info "checkout llama.cpp base commit $BASE_COMMIT"
+    git fetch --depth 1 origin "$BASE_COMMIT" || true
+    git checkout "$BASE_COMMIT"
+    if ! git apply --reverse --check "$ROOT/patches/flint-cpu-opt.patch" 2>/dev/null; then
+        git apply "$ROOT/patches/flint-cpu-opt.patch"
+    fi
+    info "cmake configure (GGML_NATIVE=ON, CUDA off)"
+    cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=ON -DGGML_CUDA=OFF -DLLAMA_CURL=OFF
+    info "cmake build (llama-server, -j$LLAMA_THREADS)"
+    cmake --build build -j"$LLAMA_THREADS" --target llama-server
+    cd "$ROOT"
+else
+    ok "llama-server 已存在，跳過編譯：$SERVER_BIN"
+fi
+
+# ── 2. Make sure the Q4_1 model is present ───────────────────────────────────
+if [[ ! -s "$MODEL_PATH" ]]; then
+    info "下載模型 $MODEL_REPO/$MODEL_FILE（約 24 GiB，支援續傳）"
+    URL="https://huggingface.co/$MODEL_REPO/resolve/main/$MODEL_FILE?download=true"
+    if command -v hf >/dev/null 2>&1; then
+        hf download "$MODEL_REPO" "$MODEL_FILE" --local-dir "$ROOT" || \
+            curl -L -C - --retry 3 --retry-delay 5 "$URL" -o "$MODEL_PATH"
+    else
+        curl -L -C - --retry 3 --retry-delay 5 "$URL" -o "$MODEL_PATH"
+    fi
+fi
+if [[ ! -s "$MODEL_PATH" ]]; then
+    warn "模型不存在或下載失敗：$MODEL_PATH"
+    exit 1
+fi
+ok "模型：$MODEL_PATH"
+
+# ── 3. Install pi if needed, and configure it to use this llama-server ───────
+export PATH="$HOME/.local/bin:$HOME/.pi/bin:${XDG_DATA_HOME:-$HOME/.local/share}/pi-node/current/bin:$PATH"
+if ! command -v pi >/dev/null 2>&1; then
+    if command -v npm >/dev/null 2>&1; then
+        info "安裝 pi coding agent 到 $HOME/.local"
+        npm install -g --ignore-scripts --no-fund --no-audit --progress=false \
+            --prefix "$HOME/.local" @earendil-works/pi-coding-agent || warn "pi 安裝失敗"
+    else
+        warn "找不到 npm，略過 pi 安裝（僅寫入設定）"
+    fi
+fi
+
+info "設定 pi 使用本地 llama-server：http://127.0.0.1:$LLAMA_PORT/v1"
+python3 - "$PI_MODELS_JSON" "$LLAMA_PORT" "$PI_MODEL_ID" "$PI_PROVIDER_ID" "$CONTEXT_SIZE" "$MODEL_ALIAS" <<'PY'
+import json, os, sys
+path, port, model_id, provider, ctx_size, alias = sys.argv[1:7]
+ctx_size = int(ctx_size)
+os.makedirs(os.path.dirname(path), exist_ok=True)
+cfg = {"providers": {}}
+if os.path.exists(path):
+    try:
+        with open(path) as f:
+            loaded = json.load(f)
+        if isinstance(loaded, dict):
+            cfg = loaded
+            cfg.setdefault("providers", {})
+    except Exception as exc:
+        print(f"[WARN] 無法解析舊 models.json，將覆蓋：{exc}")
+cfg["providers"][provider] = {
+    "baseUrl": f"http://127.0.0.1:{port}/v1",
+    "api": "openai-completions",
+    "apiKey": "not-needed",
+    "compat": {"supportsDeveloperRole": True, "supportsReasoningEffort": False},
+    "models": [{
+        "id": model_id,
+        "name": f"Flint Q4_1 (Local {alias})",
+        "reasoning": True,
+        "input": ["text"],
+        "contextWindow": ctx_size,
+        "maxTokens": ctx_size,
+        "cost": {"input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0},
+    }],
+}
+with open(path, "w") as f:
+    json.dump(cfg, f, indent=2)
+print(f"[build.sh] 已寫入 {path}")
+PY
+
+WRAPPER="$HOME/.local/bin/pi-local"
+cat > "$WRAPPER" <<EOS
+#!/usr/bin/env bash
+exec pi --provider $PI_PROVIDER_ID --model $PI_MODEL_ID "\$@"
+EOS
+chmod +x "$WRAPPER"
+ok "pi-local wrapper：$WRAPPER"
+
+# ── 4. Start llama-server on port 20000 ──────────────────────────────────────
+if curl -sf "http://127.0.0.1:$LLAMA_PORT/health" >/dev/null 2>&1; then
+    ok "llama-server 已在 127.0.0.1:$LLAMA_PORT 執行，沿用"
+else
+    if [[ -f "$PIDFILE" ]]; then
+        OLD_PID="$(cat "$PIDFILE" 2>/dev/null || true)"
+        if [[ -n "$OLD_PID" ]] && kill -0 "$OLD_PID" 2>/dev/null; then
+            info "停止舊 llama-server (PID $OLD_PID)"
+            kill "$OLD_PID" 2>/dev/null || true
+            sleep 2
+            kill -9 "$OLD_PID" 2>/dev/null || true
+        fi
+    fi
+    info "啟動 llama-server：127.0.0.1:$LLAMA_PORT（model=$MODEL_FILE, alias=$MODEL_ALIAS, ctx=$CONTEXT_SIZE, threads=$LLAMA_THREADS）"
+    nohup "$SERVER_BIN" \
+        -m "$MODEL_PATH" \
+        --alias "$MODEL_ALIAS" \
+        -t "$LLAMA_THREADS" \
+        -fa on \
+        --jinja \
+        -c "$CONTEXT_SIZE" \
+        --host "$LLAMA_HOST" \
+        --port "$LLAMA_PORT" \
+        >"$LOGFILE" 2>&1 &
+    echo $! > "$PIDFILE"
+    info "llama-server PID: $(cat "$PIDFILE")，log: $LOGFILE"
+
+    READY=0
+    for _ in $(seq 1 120); do
+        if curl -sf "http://127.0.0.1:$LLAMA_PORT/health" >/dev/null 2>&1; then
+            READY=1
+            break
+        fi
+        if ! kill -0 "$(cat "$PIDFILE")" 2>/dev/null; then
+            warn "llama-server 程序已結束，最後 40 行 log："
+            tail -n 40 "$LOGFILE" >&2 || true
+            break
+        fi
+        sleep 5
+    done
+    if [[ "$READY" -ne 1 ]]; then
+        warn "llama-server 未就緒（見 $LOGFILE）"
+        exit 1
+    fi
+fi
+
+ok "完成：llama-server http://127.0.0.1:$LLAMA_PORT/v1（model=$PI_MODEL_ID, provider=$PI_PROVIDER_ID）"
+ok "pi 用法：pi --provider $PI_PROVIDER_ID --model $PI_MODEL_ID  或  pi-local"