v0.9.0

Native install (Apple Silicon / homelab)

uv tool install mship
mship bootstrap --metal
mship start --config models.yaml

Docker images

Thin (control/coordinator — default, no torch/vllm)

docker pull ghcr.io/modelship-ai/modelship:0.9.0

CUDA (GPU node)

docker pull ghcr.io/modelship-ai/modelship:0.9.0-cuda

CPU (CPU node)

docker pull ghcr.io/modelship-ai/modelship:0.9.0-cpu

Floating tags (:latest, :latest-cuda, :latest-cpu) are single-node only — for any
multi-node cluster, pin every node to the same X.Y.Z tag to avoid a Ray version
mismatch between head and workers.

Helm chart (Kubernetes)

helm install modelship \
  oci://ghcr.io/modelship-ai/charts/modelship \
  --version 0.9.0 \
  -f values.yaml

Running an image

The images take the same subcommands as the mship CLI:

docker run --rm --shm-size=8g \
  -v ./models.yaml:/models.yaml \
  -v ./models-cache:/.cache \
  -p 8000:8000 \
  ghcr.io/modelship-ai/modelship:0.9.0-cpu start --config /models.yaml

Configuration

Create a models.yaml file, mount it and pass it with --config. Example configs for various GPU sizes are included in the image under /modelship/config/examples/ — use them as a reference for structure and available options, then tailor the models and GPU fractions to your hardware.

What's Changed

• docs: the bench results drop the host-RAM row by @alez007 in #251
• fix: vllm reasoning tokens are counted on the parser itself by @alez007 in #254
• fix: a request's reasoning effort switches thinking on the vllm loader by @alez007 in #255
• fix: /v1/responses/compact summarizes without thinking by @alez007 in #257
• fix: llama_server reports reasoning tokens in usage by @alez007 in #258
• Fix/responses reasoning replay by @alez007 in #259
• Perf/responses linear store by @alez007 in #260
• Feat/vllm thinking budget by @alez007 in #262
• Fix/vllm reasoning parser probe by @alez007 in #263
• Build/vllm 0.31.0 by @alez007 in #264
• build: bump llama.cpp to b11699 by @github-actions[bot] in #266
• build: llama.cpp b11699 llama_server changes by @alez007 in #265

Full Changelog: v0.8.1...v0.9.0