container

DeepSeek-V4 Reference Containers

Shared reference Dockerfiles for the DeepSeek-V4 family — used by both deepseek-v4-flash and deepseek-v4-pro. Nothing in either image is recipe-specific; the model is selected at runtime via --model-path (SGLang).

Backend	Dockerfile	Base image	Build flow
SGLang (B200)	sglang/Dockerfile.dsv4.sglang.b200	lmsysorg/sglang:deepseek-v4-blackwell (digest-pinned, amd64)	Two-stage; Dynamo runtime image as donor
SGLang (GB200)	sglang/Dockerfile.dsv4.sglang.gb200	lmsysorg/sglang:deepseek-v4-grace-blackwell (digest-pinned, arm64)	Two-stage; Dynamo runtime image as donor

NVIDIA also publishes the prebuilt images for vLLM and SGLang which manifests pull directly:

• nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.2.0-deepseek-v4-cuda13-dev.2 (multi-arch)
• nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.2.0-deepseek-v4-cuda13-dev.2 (arm64 only)
• nvcr.io/nvidia/ai-dynamo/sglang-runtime:1.2.0-deepseek-v4-cuda12-dev.2 (amd64 only)

The cudaXY suffix encodes the CUDA major version baked into the image, not the hardware target.

Optional: users may also build the standard Dynamo vLLM runtime image via container/render.py. See <repo_root>/container/README.md.

SGLang (sglang/Dockerfile.dsv4.sglang.b200)

Two-stage build: a Dynamo SGLang runtime image as the donor (for nats / etcd / UCX / NIXL and the Dynamo wheels + Python source), layered onto the upstream SGLang dsv4 base.

Step 1 — Build the Dynamo SGLang runtime

From the repo root:

container/render.py --framework sglang --target runtime --output-short-filename
docker build -t dynamo:latest-sglang-runtime -f container/rendered.Dockerfile .

This produces the local tag dynamo:latest-sglang-runtime, which Step 2 expects as DYNAMO_SRC_IMAGE. The donor must contain the V4 tool/reasoning parsers and the SGLang routed_experts fix; the build asserts on this with a post-install assert 'deepseek_v4' in get_tool_parser_names().

See <repo_root>/container/README.md for runtime-image build details and alternative tags.

Step 2 — Build the dsv4 overlay

Still from the repo root:

docker build \
  -f recipes/deepseek-v4/container/sglang/Dockerfile.dsv4.sglang.b200 \
  -t <your-registry>/sglang-dsv4:<tag> \
  .

The Dockerfile takes nothing from the build context (everything comes from FROM / COPY --from=), so any context directory works.

Build args

Arg	Default	Purpose
DYNAMO_SRC_IMAGE	dynamo:latest-sglang-runtime	Source for nats / etcd / UCX / NIXL and the V4-aware Dynamo wheels. Default matches Step 1; override with a published Dynamo SGLang runtime tag for reproducible builds without rebuilding locally.
DSV4_BASE_IMAGE	lmsysorg/sglang:deepseek-v4-blackwell@sha256:da2acdc8...	The DeepSeek-V4 SGLang base. Digest-pinned for byte-stable rebuilds.

Wire into a recipe

Push:

docker push <your-registry>/sglang-dsv4:<tag>

Set the image: field (Frontend + decode worker) in the recipe's SGLang manifest, then follow the recipe's Quick Start:

• Flash → ../deepseek-v4-flash/sglang/agg/deploy.yaml — see Quick Start.
• Pro → ../deepseek-v4-pro/sglang/agg/deploy.yaml — see Quick Start.