1DOLLY 72PACKAGE llama3REQUIRES HOST runtime@04REQUIRES HOST gpu@056# dolly-llama, linked on the llama-cpp package: llama.cpp with its WebGPU7# backend, serving the model it is given over stdin and stdout. A model package8# brings the GGUF.9FROM /v0.1.4/Dollyfile-system-tools 7f6c86fcc73e544e7007646eba2718bbc4d43b7f0516d9558e2dc500b1a05a8010USE /v0.1.4/demos/local-llm/Dollyfile-llama-cpp 36a6c8fe63818088d8baa5bc62fe8984935bd16c9654185b5b63a2fe57016dea11REQUIRES HEADER llama12REQUIRES LIB llama13REQUIRES LIB llama-common14REQUIRES LIB ggml15REQUIRES LIB ggml-base16REQUIRES LIB ggml-cpu17REQUIRES LIB ggml-webgpu18REQUIRES LIB dolly-webgpu19REQUIRES LIB dolly-gpu20REQUIRES TOOL c++21REQUIRES TOOL cp22REQUIRES TOOL mkdir23REQUIRES TOOL rm24REQUIRES TOOL tar2526SOURCE /v0.1.4/dist/static/llama/engine.tar 320eac803d1b96305f6422c9d2257e60ea539e27aec0c7ce670721d77c705ed1 /tmp/llm-engine.tar27SLOP tar -xf /tmp/llm-engine.tar -C /28SLOP c++ -std=c++20 -O1 -I/usr/include/dolly-llm /tmp/dolly-llama/main.cpp \29  -L/usr/lib/dolly-llm -lllama-common -lllama -lggml -lggml-webgpu -ldolly-webgpu -lggml-cpu -lggml-base -ldolly-gpu -lm -o /usr/bin/dolly-llama30# The program carries llama.cpp's and Dawn's headers' licences.31SLOP mkdir -p /usr/share/licenses/dolly-llama32SLOP cp -R /usr/share/licenses/dolly-llm/. /usr/share/licenses/dolly-llama33SLOP rm -rf /tmp/dolly-llama /tmp/llm-engine.tar3435EXPORTS TOOL dolly-llama36FOLDER /usr/share/licenses/dolly-llama37FILE /usr/share/man/cat1/dolly-llama.138    dolly-llama: llama.cpp on WebGPU, serving a GGUF model to Pi (provider webgpu).39    40      dolly-llama MODEL.gguf    serves the model41      dolly-llama --check       prints the shader kind this adapter gets42    43    Each line of standard input is a request as llama-server takes it at44    /v1/chat/completions; standard output is the event stream it would answer45    with, and a field it does not implement is an error naming it. It uses the46    adapter gpu@0 grants: with shader-f16 llama.cpp's WGSL as upstream wrote it,47    without it the f16 shader code widened to f32. The chat template, tool48    calls, stop tokens and sampling defaults are the GGUF's own. Pi's /model,49    /thinking and /local choose and tune the model.50