Document browser graph and parity
Browse files
README.md
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
library_name: onnxruntime
|
| 4 |
+
tags:
|
| 5 |
+
- onnx
|
| 6 |
+
- webgpu
|
| 7 |
+
- continuous-thought
|
| 8 |
+
- coconut
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Neuralese COCONUT browser graph
|
| 12 |
+
|
| 13 |
+
This repository hosts the compact ONNX graph used by the **Neuralese realtime
|
| 14 |
+
latent instrument**. It is derived from the MIT-labelled
|
| 15 |
+
[`ModalityDance/latent-tts-coconut`](https://huggingface.co/ModalityDance/latent-tts-coconut)
|
| 16 |
+
checkpoint at revision `89501ce8cc7cefd020b6e30c475a6772494de0a4`.
|
| 17 |
+
|
| 18 |
+
The graph implements strict COCONUT recurrence: a decoder final hidden state can
|
| 19 |
+
be supplied directly as the next input embedding before returning to token mode.
|
| 20 |
+
It is not an ordinary token model with post-hoc activation sonification.
|
| 21 |
+
|
| 22 |
+
## Artifact
|
| 23 |
+
|
| 24 |
+
`coconut-decoder-fp16-storage.onnx`
|
| 25 |
+
|
| 26 |
+
- 124M-parameter GPT-2 architecture
|
| 27 |
+
- 12 layers, 12 heads, hidden size 768
|
| 28 |
+
- FP16 weight storage with FP32 inputs, arithmetic, outputs, and KV cache
|
| 29 |
+
- SHA-256: `3bc35680260edb1c66276bbd65ae347535ec89ccb80914b469c1d8cf55f400bd`
|
| 30 |
+
- Source graph: 622 MiB FP32
|
| 31 |
+
- Compact graph: 311 MiB
|
| 32 |
+
|
| 33 |
+
Native FP16 arithmetic was tested but rejected because its recurrent trajectory
|
| 34 |
+
fell below the project's browser fidelity gate. The published graph halves the
|
| 35 |
+
download while preserving FP32 computation.
|
| 36 |
+
|
| 37 |
+
## Measured browser parity
|
| 38 |
+
|
| 39 |
+
Against the publisher-faithful PyTorch trajectory for the fixed six-state probe:
|
| 40 |
+
|
| 41 |
+
- identical generated tokens;
|
| 42 |
+
- minimum latent cosine: `0.99999919`;
|
| 43 |
+
- maximum absolute latent error: `1.742e-2`;
|
| 44 |
+
- approximately 29–30 ms per model call after warm-up on the development M4 Pro.
|
| 45 |
+
|
| 46 |
+
The checkpoint's answer to the probe is incorrect (`### 6` for `2 + 2`). This is
|
| 47 |
+
a checkpoint capability result, not a conversion discrepancy. The artifact is
|
| 48 |
+
published for artistic research, browser inference, and reproducibility—not as
|
| 49 |
+
a reliable general reasoning model.
|
| 50 |
+
|
| 51 |
+
## Attribution
|
| 52 |
+
|
| 53 |
+
Model weights originate from `ModalityDance/latent-tts-coconut`. The recurrent
|
| 54 |
+
continuous-thought method is based on COCONUT, *Training Large Language Models
|
| 55 |
+
to Reason in a Continuous Latent Space* (Meta FAIR, 2024).
|