---
license: cc-by-nc-sa-4.0
base_model: stanford-star/rt-j
tags:
- relational-deep-learning
- relational-databases
- tabular
- tabular-classification
- tabular-regression
- foundation-model
- in-context-learning
- quantized
- int8
- relational-transformer
datasets:
- stanford-star/the-join
- stanford-star/relbench
---

# RT-J — int8 (Q8_0-style) quantized checkpoints

Int8 quantizations of [stanford-star/rt-j](https://huggingface.co/stanford-star/rt-j),
the Relational Transformer foundation model for in-context learning over
relational databases. Produced and consumed by the
[RelativeDB](https://github.com/RelativeDB) native inference engine
(`cpp/rt_quantize --type q8`). Weights stay **quantized-resident** at
inference: the engine's CPU (Accelerate / portable SIMD) and Metal/MPS
kernels dequantize inside the GEMM, so DRAM weight traffic is the int8
payload.

Sibling repos: [rt-j-int4](https://huggingface.co/RelativeDB/rt-j-int4) ·
[rt-j-fp16](https://huggingface.co/RelativeDB/rt-j-fp16)

| File | Task head | Size |
|---|---|---|
| `classification/model.q8.safetensors` | classification / ranking (logits — apply sigmoid) | 86 MB |
| `regression/model.q8.safetensors` | regression / forecasting (normalized values) | 86 MB |

(vs. 171 MB bf16 upstream, 342 MB fp32 in memory.)

## Quantization format

Q8_0-style, plain safetensors — no custom container:

- Every **transformer-block projection** (`wq/wk/wv/wg`, `wo`, `ffn.w1/w2/w3`
  across all 12 blocks — ~99% of parameters) is stored as an `I8` tensor with
  a per-output-row symmetric fp32 scale in a `<name>.q_scale` companion
  tensor: `W[o,i] ≈ q[o,i] * scale[o]`, `scale[o] = max|W[o,:]| / 127`.
- The value/col-name **encoders, decoder head, norms, biases and mask
  embeddings stay fp32** (like llama.cpp's embedding/output layers):
  input-side quantization error would propagate through all 12 blocks for a
  negligible size win.

Any safetensors reader can load these files; dequantization is one line per
row. The RelativeDB engine keeps the int8 payload resident and dequantizes
64-row tiles (CPU) or in-register while staging GEMM tiles (Metal); the CUDA
backend is fp32-only for now.

## Accuracy

Measured on the RelativeDB golden batch (B=5, S=16 churn example) against the
PyTorch reference, identical on CPU and Metal/MPS:

| Metric | fp32 | int8 |
|---|---|---|
| `yhat` max abs. error vs. torch | 3.9e-3 | 1.1e-2 |
| `yhat` mean abs. error | 5.1e-4 | 3.5e-3 |
| target-score sign / ranking | preserved | preserved |

Worst per-row mean relative weight error: ≈5%.

## Usage (RelativeDB native engine)

```bash
# direct path
./build/rt_test testdata classification/model.q8.safetensors --quantized --device mps

# via the Java / Python / Rust bindings: place the .q8 file next to the fp32
# checkpoint (or point at a directory containing it) and opt in with
export RELATIVEDB_RT_QUANTIZED=q8   # (or 1/true; q4 and f16 select siblings)
```

The C ABI (`rt_model_load`) accepts these files directly — the format is
auto-detected from the tensor dtypes.

## Reproduce

```bash
cmake -B build -S cpp && cmake --build build -j
./build/rt_quantize <rt-j>/classification/model.safetensors classification/model.q8.safetensors
./build/rt_quantize <rt-j>/regression/model.safetensors regression/model.q8.safetensors
```

## License & attribution

Derivative of [stanford-star/rt-j](https://huggingface.co/stanford-star/rt-j)
(Stanford STAR lab), redistributed under the same
**CC-BY-NC-SA-4.0** license. Architecture and training details are described
in the upstream model card; only the weight storage format differs here.