GPU NAM — Neural Amp Modeler player (separate repo)

GPU NAM is a Neural Amp Modeler (.nam) player that runs amp captures on the GPU, built on Pulp. It used to live here under examples/gpu-nam. It has grown into a full plugin — its own .nam / Keras format loaders, a reference-faithful CPU oracle for five capture architectures, a recreated editor, format packaging, and bundled models — so it now lives in its own repository and consumes Pulp as an SDK:

https://github.com/danielraffel/pulp-gpu-nam

Keeping it out of the framework repo lets Pulp stay focused on reusable capabilities rather than one large, domain-specific plugin. It also makes GPU NAM a worked example of building a real plugin against the Pulp SDK — Pulp is vendored there as a git submodule, and the plugin depends only on Pulp's public targets (pulp::render, pulp::gpu-audio, pulp::signal, pulp::view, pulp::canvas, pulp::runtime).

Install

Download the signed, notarized macOS installer from the repo's Releases page. It offers a Customize pane to pick formats (AU / VST3 / CLAP / Standalone). A default capture ships in the bundle, so it makes sound immediately; load your own .nam for real amp tones.

Build from source

git clone https://github.com/danielraffel/pulp-gpu-nam.git
cd pulp-gpu-nam
git submodule update --init --recursive
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --target GpuNam_Standalone GpuNam_CLAP GpuNam_VST3 GpuNam_AU \
      -j$(sysctl -n hw.ncpu)

Full build, test, packaging, and architecture-support details are in that repo's README and docs/nam-support.md.

What stayed in Pulp: the GPU WaveNet inference primitive

The one framework capability GPU NAM relies on remains part of the Pulp SDK: a fused, block-parallel, conditioned-WaveNet GPU forward — pulp::render::GpuCompute::prepare_wavenet / wavenet_forward (see core/render/include/pulp/render/gpu_compute.hpp). It is a general neural-inference primitive — a sequence of gated, dilated, causal 1-D conv layer-arrays computed GPU-resident with the CPU↔GPU round-trip paid once per block — not specific to any capture format. GPU NAM's repo owns the .nam translation onto it.

For the design rationale of running neural-amp inference on the GPU (why the naive per-sample approach loses and the fused block-parallel approach wins as models grow), see the honest write-up in the GPU NAM repo's README.