A JavaScript library for local model execution through WebGPU: generation, embeddings, reranking, and choice scoring, qualified per model and host.
Try the demo · Get started · Documentation
Run the pinned published quickstart with Node 20+ and a working WebGPU provider:
npx --package doppler-gpu@0.6.1 doppler-gpu --model gemma3-270m --prompt "Describe WebGPU briefly"First execution downloads model assets. Browser and Node support are qualified separately. The installation guide covers provider requirements and embedding the published compatibility API.
The checkout's signed Run API and resident partition exports have separate release requirements; do not assume they exist in npm 0.6.1.
Rig prepares signed Capsules with declared TargetPlans. Run verifies artifacts and selects an accepted plan; session operations reuse resident resources through the model program, binder, and command executor. Applications supply trust and host ports. JSON declares execution, JavaScript orchestrates, and WGSL computes. Doe and Reploid remain optional integrations.
Architecture and source owners · Resident partition lifetimes
Model catalog: text generators, multimodal generators, diffusion language models, translation specialists, language embedders, rerankers, protein encoders, nucleotide encoders. The support matrix separates catalog entries from verified execution and links the lifecycle evidence for each model.
The comparison scoreboard reports throughput and loading separately against declared controls. Runtime correctness, task quality, memory, and lifecycle acceptance remain separate checks. W8A8 remains experimental; it does not qualify distributed chat.
Choice scores are uncalibrated. Unsupported paths fail closed. The retained search starter needs renewed signed eligibility before fresh-install use; see installation limits.
- API index — operations and public entrypoints.
- Architecture — preparation and execution boundaries.
- Goals · Contributing — priorities and development.