72 lines
2.3 KiB
Bash
Executable File
72 lines
2.3 KiB
Bash
Executable File
#!/usr/bin/env bash
|
||
#
|
||
# Llama 4 Scout / Maverick 统一容器入口：
|
||
# 1. 幂等应用 CoreX vLLM 的 Llama 4 补丁（Scout 只用 chunked attention，
|
||
# Maverick 额外使用 FP8 补丁；补丁脚本会按 marker 自动跳过已应用项）
|
||
# 2. 检查模型目录
|
||
# 3. exec vLLM（信号可直接传给 vLLM）
|
||
#
|
||
set -euo pipefail
|
||
|
||
PATCH_SCRIPT="${LLAMA4_PATCH_SCRIPT:-/opt/llama4/apply_patch.py}"
|
||
VLLM_BIN="${VLLM_BIN:-vllm}"
|
||
|
||
python3 "${PATCH_SCRIPT}"
|
||
|
||
MODEL_PATH="${MODEL_PATH:-/models/llama4}"
|
||
SERVED_NAME="${SERVED_NAME:-llama4}"
|
||
HOST="${HOST:-0.0.0.0}"
|
||
PORT="${PORT:-8091}"
|
||
TP="${TP:-8}"
|
||
PP="${PP:-1}"
|
||
MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
|
||
MAX_NUM_SEQS="${MAX_NUM_SEQS:-1}"
|
||
GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
|
||
ENFORCE_EAGER="${ENFORCE_EAGER:-1}"
|
||
EXTRA_ARGS="${EXTRA_ARGS:-}"
|
||
|
||
if [[ ! -f "${MODEL_PATH}/config.json" ]]; then
|
||
echo "ERROR: 模型目录不完整：${MODEL_PATH}/config.json 不存在" >&2
|
||
echo "请确认已挂载模型，例如：" >&2
|
||
echo " -v /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct:/models/llama4:ro" >&2
|
||
echo " -v /mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8:/models/llama4:ro" >&2
|
||
exit 1
|
||
fi
|
||
|
||
PP_ARGS=()
|
||
if (( PP > 1 )); then
|
||
PP_ARGS=(--pipeline-parallel-size "${PP}")
|
||
fi
|
||
|
||
EAGER_ARGS=()
|
||
if [[ "${ENFORCE_EAGER}" == "1" ]]; then
|
||
EAGER_ARGS=(--enforce-eager)
|
||
fi
|
||
|
||
echo "============================================================"
|
||
echo "Llama 4 (BI-V150 unified image)"
|
||
echo " model : ${MODEL_PATH}"
|
||
echo " served name: ${SERVED_NAME}"
|
||
echo " listen : ${HOST}:${PORT}"
|
||
echo " TP / PP : ${TP} / ${PP} (total $((TP * PP)) GPUs)"
|
||
echo " max len : ${MAX_MODEL_LEN}"
|
||
echo " max seqs : ${MAX_NUM_SEQS}"
|
||
echo " gpu util : ${GPU_MEM_UTIL}"
|
||
echo " eager : ${ENFORCE_EAGER}"
|
||
echo " visible : ${CUDA_VISIBLE_DEVICES:-all}"
|
||
echo " extra : ${EXTRA_ARGS}"
|
||
echo "============================================================"
|
||
|
||
# shellcheck disable=SC2086
|
||
exec "${VLLM_BIN}" serve "${MODEL_PATH}" \
|
||
--host "${HOST}" \
|
||
--port "${PORT}" \
|
||
--served-model-name "${SERVED_NAME}" \
|
||
--tensor-parallel-size "${TP}" \
|
||
"${PP_ARGS[@]}" \
|
||
--max-model-len "${MAX_MODEL_LEN}" \
|
||
--max-num-seqs "${MAX_NUM_SEQS}" \
|
||
--gpu-memory-utilization "${GPU_MEM_UTIL}" \
|
||
"${EAGER_ARGS[@]}" \
|
||
${EXTRA_ARGS}
|