Maple DeepSeek inference configuration

Tinfoil Containers configuration for DeepSeek-V4.1-Flash-1017 on four NVIDIA B300 GPUs. This repository publishes the measured deployment configuration; it reuses Tinfoil's digest-pinned inference image and verified model pack.

Upstream provenance

tinfoil-config.yml derives from tinfoilsh/confidential-deepseek-v4-1-flash v0.0.6+b300, commit 28775a51ed83ddd1cd295852db46499d7d50c480.

• Model: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Model revision: 1bae0e1088993b9b82de08f59de5be5a34e42170
• Image digest: sha256:01dae35f82bd4fb13376a9d1f0330e4109c1c03a07532f8df3bb12f9b0e58880
• Resources: four GPUs, 56 CPUs, 768 GiB RAM; tensor parallelism four.
• Confidential VM image: 0.14.13; verified model-pack schema: 2.
• Authentication: native vLLM bearer authentication with required VLLM_API_KEY; the CVM metrics handlers use required METRICS_API_KEY. Secret values are deployment settings. The upstream /tokenize and /metrics paths are omitted.

The release workflows derive from Tinfoil's container template. The measurement action is pinned to the same v0.14.0 commit used by the upstream DeepSeek release. These workflows do not rebuild or publish an inference image.

Publish a configuration release

Run Tinfoil Release with a new version such as v0.1.0. It creates the source tag, dispatches the publish workflow on that tag, measures the configuration, and publishes the release and attestation under this repository's identity. Wait for both workflow phases and the published release before selecting it in Tinfoil Containers.

Instance names, target hosts and custom domains are deployment settings. API keys and other credential values must remain outside this public repository.

The model's license is maintained by DeepSeek. See the linked upstream repositories for their source and notices.