From d7a13981c11d5eec12e543cca69ee310e722051f Mon Sep 17 00:00:00 2001 From: zhousha <736730048@qq.com> Date: Wed, 4 Nov 2026 17:21:18 +0800 Subject: [PATCH] Add Sage-A1.5 NVIDIA image build workflow and launch script --- .dockerignore | 3 + .gitea/workflows/docker-build-push.yml | 132 +++++++++++++++++++++++++ .gitea/workflows/task_info.env | 3 + Dockerfile | 6 ++ README.md | 37 +++++++ run_Sage.sh | 31 ++++++ 6 files changed, 212 insertions(+) create mode 100644 .dockerignore create mode 100644 .gitea/workflows/docker-build-push.yml create mode 100644 .gitea/workflows/task_info.env create mode 100644 Dockerfile create mode 100644 README.md create mode 100755 run_Sage.sh diff --git a/.dockerignore b/.dockerignore new file mode 100644 index 0000000..741ccbf --- /dev/null +++ b/.dockerignore @@ -0,0 +1,3 @@ +.git +.gitea +*.log diff --git a/.gitea/workflows/docker-build-push.yml b/.gitea/workflows/docker-build-push.yml new file mode 100644 index 0000000..0220ed3 --- /dev/null +++ b/.gitea/workflows/docker-build-push.yml @@ -0,0 +1,132 @@ +name: Docker Build and Push + +on: + push: + tags: + - "v*" + +jobs: + docker: + runs-on: amd64-ubuntu-24.04 + + steps: + - name: Clone repository + run: | + git clone "${{ gitea.server_url }}/${{ gitea.repository }}.git" . + git checkout "${{ gitea.ref_name }}" + + - name: Set image metadata + run: | + IMAGE_NAME="$(echo "${{ gitea.repository }}" | tr '[:upper:]' '[:lower:]' | tr '_' '-')" + IMAGE="${DOCKER_REGISTRY}/${DOCKER_USERNAME}/${IMAGE_NAME}:${{ gitea.ref_name }}" + + echo "IMAGE_NAME=${IMAGE_NAME}" >> "$GITEA_ENV" + echo "IMAGE=${IMAGE}" >> "$GITEA_ENV" + + - name: Load and Validate Task Info + run: | + set -a + . .gitea/workflows/task_info.env + set +a + + for name in FRAMEWORK GPU_TYPE TASK_TYPE; do + eval "value=\${${name}:-}" + if [ "$name" = "FRAMEWORK" ] && [ -z "$value" ]; then + echo "${name} is empty in .gitea/workflows/task_info.env" + exit 1 + fi + + echo "${name}=${value}" >> "$GITEA_ENV" + done + + - name: Validate Image Verify Metadata + run: | + if [ -z "${FIXED_TOKEN:-}" ]; then + echo "FIXED_TOKEN is not configured on runner" + exit 1 + fi + + if ! response="$(curl --silent --show-error --location --get 'https://modelhub.org.cn/adminApi/image-verify/validate' \ + --header "Xc-Token: ${FIXED_TOKEN}" \ + --data-urlencode "gpuType=${GPU_TYPE:-}" \ + --data-urlencode "taskType=${TASK_TYPE:-}")"; then + echo "failed to call image verify validate API" + exit 1 + fi + + VALIDATE_RESPONSE="$response" python3 - <<'PY' + import json + import os + import sys + + raw = os.environ.get("VALIDATE_RESPONSE", "") + try: + body = json.loads(raw) + except json.JSONDecodeError: + print("image verify validate API returned invalid JSON") + print(raw) + sys.exit(1) + + if body.get("code") == 0 and body.get("data") is True: + print("image verify metadata validation passed") + sys.exit(0) + + message = body.get("message") or "unknown error" + print(f"image verify metadata validation failed: {message}") + print(raw) + sys.exit(1) + PY + + - name: Login to Docker Registry + run: | + echo "$DOCKER_PASSWORD" | docker login "$DOCKER_REGISTRY" \ + -u "$DOCKER_USERNAME" \ + --password-stdin + + - name: Build Docker Image + run: | + docker build -t "$IMAGE" . + + - name: Push Docker Image + run: | + for attempt in 1 2 3; do + echo "Starting docker push attempt ${attempt}/3 for ${IMAGE}" + docker push "$IMAGE" & + PUSH_PID=$! + + while kill -0 "$PUSH_PID" 2>/dev/null; do + echo "docker push is still running at $(date -u '+%Y-%m-%dT%H:%M:%SZ')" + sleep 60 + done + + if wait "$PUSH_PID"; then + echo "docker push completed successfully" + exit 0 + fi + + echo "docker push failed on attempt ${attempt}/3" + sleep 30 + done + + echo "docker push failed after 3 attempts" + exit 1 + + - name: Notify Image Verify + run: | + if [ -z "${FIXED_TOKEN:-}" ]; then + echo "FIXED_TOKEN is not configured on runner" + exit 1 + fi + + curl --silent --show-error --fail-with-body --location --request POST 'https://modelhub.org.cn//adminApi/image-verify' \ + --header "Xc-Token: ${FIXED_TOKEN}" \ + --header 'Content-Type: application/json' \ + --data-raw "{ + \"framework\": \"${FRAMEWORK}\", + \"gpuType\": \"${GPU_TYPE}\", + \"imageUrl\": \"${IMAGE}\", + \"taskType\": \"${TASK_TYPE}\", + \"createBy\": \"${{ gitea.actor }}\", + \"repoUrl\": \"${{ gitea.server_url }}/${{ gitea.repository }}\", + \"tag\": \"${{ gitea.ref_name }}\" + }" diff --git a/.gitea/workflows/task_info.env b/.gitea/workflows/task_info.env new file mode 100644 index 0000000..dc2a6cb --- /dev/null +++ b/.gitea/workflows/task_info.env @@ -0,0 +1,3 @@ +FRAMEWORK=vllm +GPU_TYPE="Nvidia-H100" +TASK_TYPE=text-generation diff --git a/Dockerfile b/Dockerfile new file mode 100644 index 0000000..bb4114d --- /dev/null +++ b/Dockerfile @@ -0,0 +1,6 @@ +FROM harbor.4pd.io/hardcore-tech/vllm/vllm-openai:v0.25.0-cu129 + +USER root +# Remove the optional TorchCodec package requiring unavailable CUDA 13 libraries. +# This image is intended for text inference validation. +RUN python3 -m pip uninstall -y torchcodec diff --git a/README.md b/README.md new file mode 100644 index 0000000..6392656 --- /dev/null +++ b/README.md @@ -0,0 +1,37 @@ +# Sage-A1.5 NVIDIA 文本推理镜像 + +基于 `harbor.4pd.io/hardcore-tech/vllm/vllm-openai:v0.25.0-cu129`,卸载依赖缺失的可选 TorchCodec 包。模型权重不打包进镜像;本版本不用于验收 TorchCodec 视频解码。 + +## 自动构建 + +推送 `v*` 标签触发 `.gitea/workflows/docker-build-push.yml`。工作流沿用 step3p5-vllm-nv,Runner 必须提供 DOCKER_REGISTRY、DOCKER_USERNAME、DOCKER_PASSWORD 和 FIXED_TOKEN;不要将凭据提交到仓库。 + +任务信息:FRAMEWORK=vllm,GPU_TYPE=Nvidia-H100,TASK_TYPE=text-generation。GPU_TYPE 必须通过平台元数据校验。 + +实际镜像地址为 `${DOCKER_REGISTRY}/${DOCKER_USERNAME}/zhoushasha/sage-a1.5-nv:`。启动脚本默认按现有 Harbor 命名习惯设置;请以 Actions 构建日志中的 IMAGE 为准,若不同使用 IMAGE 环境变量覆盖。 + +## 本地构建和启动 + +```bash +docker build -t sage-a15-vllm:cu129-text-v1 . +IMAGE=sage-a15-vllm:cu129-text-v1 bash run_Sage.sh +``` + +发布镜像构建完成后: + +```bash +bash run_Sage.sh +docker logs -f --tail 100 sage-a15-h100-test +``` + +默认模型路径 `/mnt/disk0/models/Sage-A1.5`,可通过 MODEL 覆盖。使用 GPU 0、1 两张 H100,TP=2、BF16、4096 上下文。需确保卡可用、9099 端口空闲、无同名容器。 + +```bash +curl -sS --max-time 180 http://127.0.0.1:9099/v1/chat/completions \ + -H 'Content-Type: application/json' \ + -d '{"model":"Sage-A1.5","messages":[{"role":"user","content":"你好,请用一句话介绍自己。"}],"max_completion_tokens":512,"temperature":0}' +``` + +## 验证状态 + +源环境为 575.57.08 驱动、H100。卸载 TorchCodec 后服务模块可进入模型架构解析;显式映射 UVM 两个设备后,两卡 CUDA 初始化和基础计算通过。完整模型生成结果尚待验证,CI 构建成功不等于模型验收通过。 diff --git a/run_Sage.sh b/run_Sage.sh new file mode 100755 index 0000000..6033f4f --- /dev/null +++ b/run_Sage.sh @@ -0,0 +1,31 @@ +#!/usr/bin/env bash +set -euo pipefail +IMAGE=${IMAGE:-harbor.4pd.io/modelhubxc/zhoushasha/sage-a1.5-nv:v0.1.0} +MODEL=${MODEL:-/mnt/disk0/models/Sage-A1.5} +test -f "$MODEL/config.json" +docker run -d \ + --name sage-a15-h100-test \ + --gpus '"device=0,1"' \ + --device /dev/nvidia-uvm:/dev/nvidia-uvm \ + --device /dev/nvidia-uvm-tools:/dev/nvidia-uvm-tools \ + --shm-size=16g \ + -p 127.0.0.1:9099:8000 \ + -e NVIDIA_DRIVER_CAPABILITIES=compute,utility \ + -e HF_HUB_OFFLINE=1 \ + -e TRANSFORMERS_OFFLINE=1 \ + -e VLLM_WORKER_MULTIPROC_METHOD=spawn \ + --mount "type=bind,src=$MODEL,dst=/model,readonly" \ + --entrypoint python3 \ + "$IMAGE" \ + -u -m vllm.entrypoints.openai.api_server \ + --model /model \ + --served-model-name Sage-A1.5 \ + --host 0.0.0.0 --port 8000 \ + --tensor-parallel-size 2 \ + --distributed-executor-backend mp \ + --dtype bfloat16 \ + --max-model-len 4096 \ + --max-num-batched-tokens 4096 \ + --max-num-seqs 1 \ + --gpu-memory-utilization 0.90 \ + --enforce-eager