build-zeroday

name	build-zeroday
description	Use when the user asks to build, package, or create Docker images based on vllm/vllm-openai or ascend/vllm-ascend base images. Handles building zeroday images with engine-requirements and optional patches, then exporting as tgz files (gzipped via pigz) to /os_nfs/06_images/.

Build Zeroday Image

Build Docker images based on vLLM / vLLM Ascend base images, install engine-level dependencies, optionally apply patches, and export as .tgz (gzipped via pigz).

Quick reference

# Generic image (no patches, engine-requirements only)
./build-zeroday-image.sh \
  -b vllm/vllm-openai:v0.23.0 \
  -e vllm

# With hardware/model naming
./build-zeroday-image.sh \
  -b vllm/vllm-openai:v0.23.0 \
  -e vllm \
  --hardware RTXPRO5000 \
  --model glm5.2

# With patches
./build-zeroday-image.sh \
  -b vllm/vllm-openai:v0.22.0 \
  -e vllm \
  -v v0.22.0 \
  --hardware H20 \
  --model deepseekv4flash

How it works

The script runs build-zeroday-image.sh from /home/scd/build-zeroday/.

Required arguments

Arg	Description
-b, --base-image	Base image, e.g. vllm/vllm-openai:v0.23.0
-e, --engine	Engine: vllm or vllm-ascend

Patch mode

Arg	Description
-v, --patch-version	Patch version directory under engine dir (required when applying patches)
--skip-patch	Skip patches, only install engine-requirements (makes -v optional)

Naming (auto-generates output filename)

Arg	Description
--hardware	Hardware name, e.g. RTXPRO5000, H20, 800I_A3
--model	Model name, e.g. glm5.2, deepseekv4flash
--arch	Architecture: x86_64 or aarch64 (default: auto-detect from uname -m)
--output-dir	Output directory (default: /os_nfs/06_images)
--prefix	Filename prefix (default: Wings)

Output naming convention

All hyphens in the filename are converted to underscores, and the image is saved as .tgz (gzip via pigz: docker save <img> | pigz > name.tgz).

With hardware + model: {prefix}_{engine}_{version}_{hardware}_{model}_{arch}.tgz

Generic (no hardware/model): {prefix}_{engine}_{version}_{arch}.tgz

Examples:

• Wings_vllm_v0.23.0_RTXPRO5000_glm5.2_x86_64.tgz
• Wings_vllm_v0.23.0_x86_64.tgz
• Wings_vllm_ascend_v0.19.1rc1_800I_A3_glm5.1_aarch64.tgz
• Wings_vllm_ascend_v0.21.0rc1_a3_20260709145242_aarch64.tgz (tag v0.21.0rc1-a3 → underscore)

Use -o to override the auto-generated path.

Other options

Arg	Description
-o, --output-tar	Explicit output path (overrides auto-naming)
-t, --target-image	Docker image tag (default: auto-generated with timestamp)
--push	Push image after build
--no-cache	Build without Docker cache

Output files

Each build produces:

File	Content
{name}.tgz	Docker image archive, gzipped via pigz
{name}.tgz.manifest.json	Metadata: base image, engine, patches, deps, image ID, md5, build time
{output_dir}/md5_checksums.txt	Cumulative md5 records for all builds in the directory

Engine-requirements

Engine-level pip dependencies are stored in engine-requirements/:

• vllm-zeroday.txt for -e vllm
• vllm-ascend-zeroday.txt for -e vllm-ascend

These are always installed regardless of patch mode.

Patch submodules

vllm-zeroday/ and vllm-ascend-zeroday/ are git submodules containing version-specific patches. Sync them with:

git submodule update --init --recursive
git submodule update --remote --merge vllm-ascend-zeroday vllm-zeroday