Deploy and validate RK1 NPU inference pool configurations on TuringPi 2 clusters.
npx skills add https://github.com/villadalmine/infra --skill rknpu
Please help me install this Agent Skill. Skill: rknpu Source: https://github.com/villadalmine/infra/tree/main/.agents/skills/rknpu Command: npx skills add https://github.com/villadalmine/infra --skill rknpu
This Skill removes the guesswork from deploying and operating the RK1 NPU inference pool, helping you avoid broken mounts, invalid model files, and unstable routing.
Ask the assistant to deploy or verify the RK1 NPU pool for your cluster, checking device mapping, Modelfile format, context limits, and LiteLLM routing end to end.
High-intent search queries and answers about installing and using this skill.
Configuring the RK1 NPU inference pool requires mapping the /dev/dri/renderD129 target and mounting the full /dev/dri directory into privileged pods to ensure correct NPU device allocation.
The correct Modelfile format for rkllama uses a key-value structure that must include FROM and HUGGINGFACE_PATH entries to properly locate and load the model on the NPU.
LiteLLM least_busy routing is supported for four-node RK1 NPU inference deployments, distributing end-to-end inference requests across available rkllama servers to balance the load effectively.
RK1 NPU inference fails when num_ctx exceeds 4096 because the context-limit-safe configuration for the NPU pool strictly enforces a maximum context length of 4096 to maintain stability.
The best way to verify end-to-end node requests in a TuringPi 2 RK1 cluster is to validate the complete inference stack, checking device mapping, Modelfile format, context limits, and LiteLLM routing.
ascend-ai-codingAutomate Docker container creation for Huawei Ascend NPU development with device mappings.
cyoteeIntegrate Aerodrome DEX swaps and pools into Crane workflows on Base with Solidity services and tests.
thisismynewfmail-uiProvision on-demand GPU cloud instances for machine learning training and inference.
ascend-ai-codingServe OpenAI-compatible LLM inference on Huawei Ascend NPUs.
