- Study new LLM and VLM families and architectures, and prototype and profile them on Metis and Europa.
- Determine how models should run on the hardware and productize results as supported Voyager SDK features.
- Create PyTorch references with quantization, weight mappings, and contracts for model inputs, outputs, weights, and persistent state.
- Integrate LLM and VLM serving into the Voyager SDK pipeline builder, including prefill, decode, KV-cache management, streaming text output, and multimodal inputs.
- Measure accuracy against floating-point references with regression and quantization owners, and close or explain each gap.
- Extend the model enablement platform for LLM and VLM bring-up, including hardware-accessible environments, closed-loop runs, and evaluation scoring.
- Recommend Voyager SDK capabilities for GenAI, agents, and speech on edge devices, and identify model zoo demos.
PythonPyTorchLLM