A practical developer path from model file to device evidence.
The planned Proton SDK connects GGUF inspection, llama.cpp and GGML integration, runtime control, firmware, kernels, profiling, and numerical validation.
Planned developer artifacts are not yet downloadable
Understand the model before selecting execution paths.
The SDK direction includes GGUF metadata, tensor layout, operator mix, shape, and quantization inspection so software and hardware decisions share the same workload contract.
Meet developers where local models already run.
The initial integration path is a Proton GGML backend that maps supported operations to the runtime while preserving a reference execution path for comparison.
Make command submission explicit and inspectable.
Runtime APIs, queues, buffers, synchronization, and device commands are intended to expose the boundary between host orchestration and accelerator execution.
Target-specific code remains tied to deterministic checks.
Candidate kernels include quantized dot products, unpacking, activation quantization, normalization, positional operations, and matrix or vector tiles.
Generated or hand-written code is not accepted without compilation and independent numerical checks.
A device run is incomplete without comparison evidence.
Profiling and numerical comparison are intended to connect timing, memory movement, operator coverage, and output differences back to the model and architecture decision.
IREE remains a later compiler integration path.
The first path prioritizes GGUF and llama.cpp. IREE is retained as a future frontend direction rather than presented as current integration.