Auto-detect Scout/Maverick from config.json and add verify script

run.sh

@@ -1,17 +1,24 @@
 #!/usr/bin/env bash
 #
-# Llama 4 Scout / Maverick 一键构建 + 启动（Iluvatar BI-V150）
+# Llama 4 统一构建 + 启动（Iluvatar BI-V150）
 #
-#   ./run.sh scout
-#   ./run.sh maverick
+# 自动读取模型目录下的 config.json，判断是 Scout 还是 Maverick，然后选择：
+#   Scout    : TP8 / PP1 / 8 卡 / 8091 / llama4-scout
+#   Maverick : TP8 / PP2 / 16 卡 / 8092 / llama4-maverick
+#
+# 用法：
+#   ./run.sh /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct
+#   ./run.sh /mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8
+#   MODEL_HOST_DIR=/path/to/model ./run.sh
+#   ./run.sh                    # 依次尝试两个默认模型目录
 #
 # 常用覆盖项：
-#   GPUS=8,9,10,11,12,13,14,15 ./run.sh scout
-#   MODEL_HOST_DIR=/path/to/model ./run.sh maverick
-#   EXTRA_ARGS="--max-model-len 8192" ./run.sh scout
-#   EXTRA_DOCKER_ARGS="-v /mnt/disk1:/data:ro" ./run.sh scout
-#   BUILD=0 ./run.sh scout          # 使用已有镜像，不重新构建
-#   DRY_RUN=1 ./run.sh scout        # 只打印 docker run 命令
+#   GPUS=8,9,10,11,12,13,14,15 ./run.sh /path/to/scout
+#   EXTRA_DOCKER_ARGS="-v /mnt/disk1:/data:ro" ./run.sh /path/to/model
+#   EXTRA_ARGS="--max-model-len 8192" ./run.sh /path/to/model
+#   TP=8 PP=1 PORT=8095 SERVED_NAME=my-llama4 ./run.sh /path/to/model
+#   BUILD=0 ./run.sh /path/to/model      # 使用已有镜像
+#   DRY_RUN=1 ./run.sh /path/to/model    # 只打印 docker run 命令
 #
 set -euo pipefail
 
@@ -19,67 +26,91 @@ cd "$(dirname "$0")"
 
 usage() {
   cat >&2 <<'EOF'
-用法: ./run.sh <scout|maverick>
+用法: ./run.sh [MODEL_DIR]
 
-预设默认值在 presets/<name>.env 中，所有值都可以用环境变量覆盖：
-  IMAGE NAME PORT MODEL_HOST_DIR MODEL_MOUNT SERVED_NAME
+不传 MODEL_DIR 时，按顺序尝试：
+  /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct
+  /mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8
+
+模型类型由 MODEL_DIR/config.json 自动判断，不需要 scout/maverick 参数。
+
+环境变量覆盖：
+  IMAGE NAME MODEL_HOST_DIR MODEL_MOUNT PORT SERVED_NAME
   TP PP GPUS MAX_MODEL_LEN MAX_NUM_SEQS GPU_MEM_UTIL
-  ENFORCE_EAGER EXTRA_ARGS EXTRA_DOCKER_ARGS
+  ENFORCE_EAGER AUTO_DETECT EXTRA_ARGS EXTRA_DOCKER_ARGS
   BUILD=0/1  DRY_RUN=0/1
 EOF
 }
 
-VARIANT="${1:-}"
-if [[ -z "${VARIANT}" || "${VARIANT}" == "-h" || "${VARIANT}" == "--help" ]]; then
+if [[ "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+  usage
+  exit 0
+fi
+
+MODEL_HOST_DIR="${1:-${MODEL_HOST_DIR:-}}"
+if [[ -z "${MODEL_HOST_DIR}" ]]; then
+  for candidate in \
+    "/mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct" \
+    "/mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8"; do
+    if [[ -f "${candidate}/config.json" ]]; then
+      MODEL_HOST_DIR="${candidate}"
+      break
+    fi
+  done
+fi
+
+if [[ -z "${MODEL_HOST_DIR}" ]]; then
+  echo "ERROR: 未指定模型目录，也没有找到默认模型目录。" >&2
   usage
   exit 2
 fi
 
-PRESET="presets/${VARIANT}.env"
-if [[ ! -f "${PRESET}" ]]; then
-  echo "ERROR: 未知预设: ${VARIANT}" >&2
-  usage
-  exit 2
+if [[ ! -f "${MODEL_HOST_DIR}/config.json" ]]; then
+  echo "ERROR: 模型目录不完整: ${MODEL_HOST_DIR}/config.json" >&2
+  echo "请先下载模型，或把 MODEL_DIR 指向正确目录。" >&2
+  exit 1
 fi
 
-# shellcheck disable=SC1090
-source "${PRESET}"
+DETECT_OUTPUT="$(python3 detect_model.py "${MODEL_HOST_DIR}")" || exit 1
+eval "${DETECT_OUTPUT}"
 
-IMAGE="${IMAGE:-${DEFAULT_IMAGE}}"
-NAME="${NAME:-${DEFAULT_NAME}}"
-PORT="${PORT:-${DEFAULT_PORT}}"
-MODEL_HOST_DIR="${MODEL_HOST_DIR:-${DEFAULT_MODEL_HOST_DIR}}"
+MODEL_KIND="${DETECTED_MODEL_KIND}"
+IMAGE="${IMAGE:-llama4-bi150:2026-10-26}"
+NAME="${NAME:-llama4-${MODEL_KIND}}"
+PORT="${PORT:-${DETECTED_PORT}}"
 MODEL_MOUNT="${MODEL_MOUNT:-/models/llama4}"
-SERVED_NAME="${SERVED_NAME:-${DEFAULT_SERVED_NAME}}"
-TP="${TP:-${DEFAULT_TP}}"
-PP="${PP:-${DEFAULT_PP}}"
-GPUS="${GPUS:-${CUDA_VISIBLE_DEVICES:-${DEFAULT_GPUS}}}"
-MAX_MODEL_LEN="${MAX_MODEL_LEN:-${DEFAULT_MAX_MODEL_LEN}}"
-MAX_NUM_SEQS="${MAX_NUM_SEQS:-${DEFAULT_MAX_NUM_SEQS}}"
-GPU_MEM_UTIL="${GPU_MEM_UTIL:-${DEFAULT_GPU_MEM_UTIL}}"
+SERVED_NAME="${SERVED_NAME:-${DETECTED_SERVED_NAME}}"
+TP="${TP:-${DETECTED_TP}}"
+PP="${PP:-${DETECTED_PP}}"
+MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
+MAX_NUM_SEQS="${MAX_NUM_SEQS:-1}"
+GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
 ENFORCE_EAGER="${ENFORCE_EAGER:-1}"
+AUTO_DETECT="${AUTO_DETECT:-1}"
 EXTRA_ARGS="${EXTRA_ARGS:-}"
 EXTRA_DOCKER_ARGS="${EXTRA_DOCKER_ARGS:-}"
 BUILD="${BUILD:-1}"
 DRY_RUN="${DRY_RUN:-0}"
 
+GPU_COUNT=$((TP * PP))
+
+if [[ -n "${GPUS:-}" ]]; then
+  GPUS="${GPUS}"
+elif [[ -n "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+  GPUS="${CUDA_VISIBLE_DEVICES}"
+else
+  GPUS="$(seq -s, 0 $((GPU_COUNT - 1)))"
+fi
 GPUS="${GPUS// /}"
 
-if [[ ! -f "${MODEL_HOST_DIR}/config.json" ]]; then
-  echo "ERROR: 模型目录不完整: ${MODEL_HOST_DIR}/config.json" >&2
-  echo "请先下载模型，或设置 MODEL_HOST_DIR 指向正确目录。" >&2
-  exit 1
-fi
-
-EXPECTED_GPUS=$((TP * PP))
 if [[ -n "${GPUS}" ]]; then
   IFS=',' read -r -a GPU_LIST <<< "${GPUS}"
-  if (( ${#GPU_LIST[@]} < EXPECTED_GPUS )); then
-    echo "ERROR: TP*PP=${EXPECTED_GPUS}，但 GPUS 只给了 ${#GPU_LIST[@]} 张: ${GPUS}" >&2
+  if (( ${#GPU_LIST[@]} < GPU_COUNT )); then
+    echo "ERROR: ${MODEL_KIND} 需要 TP*PP=${GPU_COUNT} 张卡，但 GPUS 只给了 ${#GPU_LIST[@]} 张: ${GPUS}" >&2
     exit 1
   fi
-  if (( ${#GPU_LIST[@]} > EXPECTED_GPUS )); then
-    echo "WARN: TP*PP=${EXPECTED_GPUS}，但 GPUS 给了 ${#GPU_LIST[@]} 张；vLLM 只会使用其中 ${EXPECTED_GPUS} 张。" >&2
+  if (( ${#GPU_LIST[@]} > GPU_COUNT )); then
+    echo "WARN: TP*PP=${GPU_COUNT}，但 GPUS 给了 ${#GPU_LIST[@]} 张；vLLM 只会使用其中 ${GPU_COUNT} 张。" >&2
   fi
 fi
 
@@ -101,6 +132,8 @@ DOCKER_RUN=(
   --ulimit stack=67108864
   -e CUDA_VISIBLE_DEVICES="${GPUS}"
   -e MODEL_PATH="${MODEL_MOUNT}"
+  -e AUTO_DETECT="${AUTO_DETECT}"
+  -e MODEL_KIND="${MODEL_KIND}"
   -e SERVED_NAME="${SERVED_NAME}"
   -e PORT="${PORT}"
   -e TP="${TP}"
@@ -115,6 +148,11 @@ DOCKER_RUN=(
   "${IMAGE}"
 )
 
+echo "==> detected ${MODEL_KIND}: ${DETECTED_DESCRIPTION}"
+echo "    model : ${MODEL_HOST_DIR}"
+echo "    TP/PP : ${TP}/${PP} (${GPU_COUNT} GPUs), port=${PORT}, name=${SERVED_NAME}"
+echo "    GPUs  : ${GPUS}"
+
 if [[ "${DRY_RUN}" == "1" ]]; then
   printf '%q ' "${DOCKER_RUN[@]}"
   printf '\n'
@@ -131,7 +169,7 @@ fi
 echo "==> recreate ${NAME}"
 docker rm -f "${NAME}" 2>/dev/null || true
 
-echo "==> run ${VARIANT}: GPUs=${GPUS}, TP=${TP}, PP=${PP}, model=${MODEL_HOST_DIR}"
+echo "==> run ${MODEL_KIND}"
 "${DOCKER_RUN[@]}"
 
 echo "==> wait for http://127.0.0.1:${PORT}/health"