makepad/libs/ai/models/speech/tools/ref_infer.py
Admin 7f59912916 libs/ai: one AI stack, replacing libs/ggml, llama, mlx, cuda, tts, voice2 and pbr_paint
The model code was spread across eight crates that had grown into each other:
ggml and cuda and mlx each owned part of a tensor runtime, llama and tts and
voice2 each owned part of a model, and libs/diffusion owned everything else.
They are now one tree with an explicit shape:

  libs/ai/cuda     — kernels and launch surface
  libs/ai/metal    — Metal shaders and the shim
  libs/ai/llm      — the language-model runtime (sessions, lanes, contexts,
                     the CUDA and Metal executors, the compiled Metal path)
  libs/ai/models/  — common, flux, h3, music, paint, speech, stems, vision

libs/diffusion is not deleted but demoted: what remains is the VALIDATOR
crate — several dozen `*_validate.rs` oracles that check a native
implementation against a reference, which is where they belong now that the
implementations live next door.

The functional work inside the move is mostly in the LLM runtime: N lanes that
draft while one verify batch serves all of them, per-slot prefill over a shared
folded attention arena, speculation that survives batching, and a scheduler
that reports rather than publishes. And in the CUDA build: a machine without
usable CUDA must still LINK (and say so), the default kernel arch is the
building machine's GPU, `NO_CUDA` forces the stub even where the toolkit
exists, and kernels compile in parallel with progress.

libs/video_flow is new here: classical optical flow estimation and the `mkfl`
motion-field payload — a flow field measured from a clip without a model,
which is what drives free-rate bounce-looping playback and the uprez/tween
enhance pipe.
2026-08-23 01:34:35 +02:00

101 lines
3.5 KiB
Python

#!/usr/bin/env python3
"""Run the reference Kokoro ONNX on tokens produced by *our* Rust phonemizer.
Ground truth for the port: same text, same tokens, same voice. Anything the Rust
graph produces later gets diffed against this.
/tmp/kref/bin/python libs/tts/tools/ref_infer.py "Escape the Gummer." out.wav
"""
import struct
import subprocess
import sys
import numpy as np
import onnxruntime as ort
MODEL = "kokoro_ref.onnx"
VOICE = "af_heart.mkvoice"
SAMPLE_RATE = 24_000
def rust_tokens(text):
"""Ask the Rust binary, so the reference sees exactly what Rust will feed it."""
out = subprocess.run(
["cargo", "run", "--release", "--quiet",
"--manifest-path", "libs/tts/Cargo.toml", "--bin", "g2p_test", "--", "--ids", text],
capture_output=True, text=True, check=True,
)
return [int(x) for x in out.stdout.strip().split(",")]
def load_voice(path):
"""Read the `.mktts` container: one [510, 1, 256] tensor."""
blob = open(path, "rb").read()
assert blob[:8] == b"MKTTS\0\0\0", "bad voice magic"
count = struct.unpack("<I", blob[12:16])[0]
assert count == 1, count
at = 16
name_len = struct.unpack("<I", blob[at:at + 4])[0]
at += 4 + name_len
_dtype, ndim = blob[at], blob[at + 1]
at += 2
shape = struct.unpack(f"<{ndim}I", blob[at:at + 4 * ndim])
at += 4 * ndim
offset, nbytes = struct.unpack("<QQ", blob[at:at + 16])
data = np.frombuffer(blob[offset:offset + nbytes], dtype=np.float32)
return data.reshape(shape)
def write_wav(path, samples, rate):
pcm = np.clip(samples, -1.0, 1.0)
pcm = (pcm * 32767.0).astype("<i2")
with open(path, "wb") as out:
out.write(b"RIFF")
out.write(struct.pack("<I", 36 + pcm.nbytes))
out.write(b"WAVEfmt ")
out.write(struct.pack("<IHHIIHH", 16, 1, 1, rate, rate * 2, 2, 16))
out.write(b"data")
out.write(struct.pack("<I", pcm.nbytes))
out.write(pcm.tobytes())
def main():
text = sys.argv[1] if len(sys.argv) > 1 else "Escape the Gummer, a squishy purple blob."
dst = sys.argv[2] if len(sys.argv) > 2 else "kokoro_ref.wav"
ids = rust_tokens(text)
phonemes = len(ids) - 2 # the ids are zero-padded at both ends
voice = load_voice(VOICE)
# One style vector per phoneme count. Row `phonemes`, both 128-halves.
# kokoro/pipeline.py uses `pack[len(ps) - 1]`. Row `phonemes` also sounds
# fine, which is exactly how this off-by-one survives a listening test.
style = voice[phonemes - 1]
print(f"text : {text}")
print(f"tokens : {len(ids)} ids ({phonemes} phonemes)")
print(f"style row: {phonemes - 1} of {voice.shape[0]} -> {style.shape}")
session = ort.InferenceSession(MODEL, providers=["CPUExecutionProvider"])
waveform = session.run(
None,
{
"input_ids": np.array([ids], dtype=np.int64),
"style": style.astype(np.float32).reshape(1, 256),
"speed": np.array([1.0], dtype=np.float32),
},
)[0][0]
print(f"waveform : {waveform.shape[0]} samples "
f"({waveform.shape[0]/SAMPLE_RATE:.2f}s), peak={np.abs(waveform).max():.4f}")
write_wav(dst, waveform, SAMPLE_RATE)
print(f"wrote : {dst}")
# A 16k copy so Whisper can score it with the same harness.
index = np.arange(0, len(waveform), SAMPLE_RATE / 16_000)
resampled = np.interp(index, np.arange(len(waveform)), waveform)
write_wav(dst.replace(".wav", "_16k.wav"), resampled, 16_000)
np.save(dst.replace(".wav", ".npy"), waveform)
if __name__ == "__main__":
main()