History

Fabio c36510c21e Some checks failed Linux Build / Linux cpu (push) Has been cancelled Details Linux Build / Linux cpu-clang (push) Has been cancelled Details Linux Build / Linux vulkan (push) Has been cancelled Details mac-build / macOS CPU (push) Has been cancelled Details Nix Build / Nix (cpu) (push) Has been cancelled Details Nix Build / Nix (vulkan) (push) Has been cancelled Details windows-build / Windows CPU (push) Has been cancelled Details Initial import: audio.cpp with Orange Pi documentation		2026-11-04 19:05:22 +01:00
..
benchmark.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
cases.json	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
codec_parity.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
greedy_parity.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
musiccoca_parity.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
musiccoca_reference.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
README.md	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
reference.py	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00
validation.cpp	Initial import: audio.cpp with Orange Pi documentation	2026-11-04 19:05:22 +01:00

README.md

Magenta RealTime 2 validation

Build the model test targets in a non-deployment Debug build:

cmake --build build/debug --target magenta_musiccoca_parity \
  magenta_codec_parity magenta_greedy_parity magenta_validation -j8

The Python reference scripts need the official magenta-realtime package with its JAX and TFLite dependencies. Use a separate reference environment.

Test	Coverage
`musiccoca_reference.py` / `magenta_musiccoca_parity`	Official text embeddings, mapper with aligned noise, RVQ codes, audio frontend, and audio embeddings.
`reference.py` / `magenta_codec_parity`	Teacher-forced official AR codes through the C++ codec; final-waveform cosine must exceed 0.99999. Tests one/four-frame blocks, padded final blocks, and reset/reuse.
`reference.py` / `magenta_greedy_parity`	Unforced greedy AR generation with aligned conditioning and F32 precision, then final-waveform comparison.
`magenta_validation`	Normal generation, repeated requests after different lengths/prompts, and exact concatenation of streaming events.
`benchmark.py`	Official Python system using the same prompts, durations, and seed as `cases.json`.

For greedy parity, convert the selected generator with --type orig and generate a matching F32 reference. Storage/compute differences can change greedy choices and then the subsequent sequence; use the teacher-forced waveform test to validate codec decoding independently.

python tests/magenta_realtime_2/reference.py --source checkpoint-dir \
  --variant small --frames 17 --temperature 0 --compute-dtype f32 \
  --output reference-small --log reference-small.log

build/debug/bin/magenta_greedy_parity gguf-dir/magenta-rt2-small-orig.gguf \
  reference-small/reference.safetensors cuda --log greedy-small.log

build/debug/bin/magenta_codec_parity gguf-dir/magenta-rt2-small-f16.gguf \
  reference-small/reference.safetensors cuda --log codec-small.log

Repeat with --variant base and the Base GGUF for the other transformer. For runtime validation, execute outside the checkout using absolute paths to the binary, case file, and output directory. Do not pass a model-spec override; confirm runtime.model.contract_spec selects @gguf/magenta_realtime_2.json.

magenta_validation MODEL.gguf cpu|cuda|vulkan offline|streaming CASES.json OUTPUT_DIR --log FILE

cases.json uses one session: warmup, a request and repeat, different short and long prompts, then the original request again. Keep timing runs sequential.