True Ternary Refactor 10 — Sidecar Int8 Verification And Full-System Smoke

Scope

The sidecar model files are now available locally:

arbitor/encoders/models/dinov2-small
arbitor/encoders/models/moonshine-base
arbitor/encoders/models/pig-vae/model.safetensors

This pass verifies int8 quantization for the available sidecars and checks the full multimodal path through the Triton-backed ternary system.

Sidecar Quantization Metadata

Added explicit metadata to quantized imported sidecars:

_arb_quantize_requested
_arb_quantized
_arb_quantized_int8

The helper also freezes every sidecar parameter after quantization.

Applied to:

ImageSequencer.vit
AudioSequencer.audio_encoder
pig_vae loader paths

Verified Int8 Sidecars

ImageSequencer(quantize_weights='int8'):

image quant_requested= int8
image quantized_int8= True
image quantized= True
image trainable= 0
image quant_classes= {'QConv2d': 1, 'QLinear': 72}

AudioSequencer(quantize_weights='int8'):

audio quant_requested= int8
audio quantized_int8= True
audio quantized= True
audio trainable= 0
audio quant_classes= {'QLinear': 128}

Focused forward smokes:

image_forward_ok (1, 254, 512) True
audio_forward_ok (1, 36, 512) True

Full Multimodal CUDA Smoke

ARBModel(enable_image=True, enable_audio=True, enable_vq=True, enable_graph=True, enable_memory_modules=True, enable_moe=True):

image_quantized_int8 True
audio_quantized_int8 True
full_multimodal_cuda_train_smoke_ok logits=(1, 8, 297), targets=(1, 7), indices=(1, 298), loss=17.4929

This exercised:

int8 DINO/ViT sidecar
int8 Moonshine sidecar
ternary image/audio projections
multimodal VQ bridge
graph path with Triton aggregation/gather kernels
MoE path with Triton dense combine kernel
output router / byte head
backward and _ternary_update_memory()

Full model audit with image/audio enabled:

logical ternary weights: 42,669,632
ternary training state: 55.72 MB
trainable float params: 0 tensors, 0.00 MB
frozen float params: 433 tensors, 318.80 MB
float buffers: 406 tensors, 0.00 MB

The frozen float params belong to imported sidecars. Their compute modules were verified as Quanto int8 wrappers where supported by the active environment.

pig-vae Status

pig-vae/model.safetensors is present locally, and the loader now applies the same quantization metadata/freeze path as the vision/audio sidecars.

Runtime verification is blocked in this Python environment because diffusers is not installed:

RuntimeError: pig-vae requires the optional diffusers dependency.

The system Python is externally managed, so I did not force-install packages into it. To verify pig-vae in this checkout, create/use a project venv and install:

pip install -e .[diffusers]

Then run:

python - <<'PY'
from arbitor.encoders.pig_vae import load_vae
vae = load_vae(device='cpu', quantize='int8')
print(vae.vae._arb_quantized_int8)
PY

Kernel Coverage

Current Triton-backed full-system paths:

packed ternary linear forward/backward/update
packed ternary embedding forward/backward/update
ternary RMSNorm
E residual update
T_accum update
Graph edge weighting + target aggregation
Graph VQ-index gather + residual add
MoE dense route combine
VideoHead denoise update

Remaining non-fused control loops:

VideoHead diffusion/halting loop
Graph hop loop
MoE ACT iteration loop

Those loops control repeated computation and halting. They are supported by Triton kernels internally, but they are not yet persistent monolithic kernels.

Verification

python -m py_compile arbitor/components.py arbitor/sequencers.py arbitor/encoders/audio.py arbitor/encoders/pig_vae.py arbitor/main.py arbitor/vq.py arbitor/kernel/ternary_scale.py arbitor/kernel/ternary_audit.py
python -m pytest -q testing/test_tscale.py -k "cuda_triton_correctness_update_E or cuda_triton_tscale_path": 2 passed
full multimodal CUDA smoke passed with image/audio sidecars enabled