Realm

tu_nlp_byte_transformer — Byte × Transformer

Source-owned OA v0.12.0 SDK tutorial. Recorded results qualify the named device and driver.

Experimental

The byte-tokenizer × Transformer cell of the — same 576-character corpus, same 300-step recipe and captured training lifecycle. This entry selects the tokenizer and model shown below. See for the full walkthrough.

This cell

Tokenizerthe 256-token packed-U8 byte vocabulary (ByteSampler) — one token per source byte
ModelByte + position Embedding → Transformer(32,64) → LN → Linear(32→256)
Parameters25,760
OptimizerAdamW(lr=0.01) — the canonical recipe

The loop is a prepare/record pair per step — GradientTape → cross_entropy → backward → AdamW — driven by oa::ml::ItTraining, so the whole run is one captured GPU program replayed 299 times (46 nodes for this model).

Measured — Surface Laptop 5 (ADL GT2)

Byte × Transformer training loss

Canonical 300-step training loss, recorded by this tutorial.

Loss init→finalAccuracyProgram nodesWall ms/stepToken/s
5.657→0.20192.9%467.85130.5K

Reference generation on the default recipe (asserted by the tutorial):

1prompt: "to be"
2generated: "to be that is the question whether tis nobler in the mind to suffer the slings and ar"
3

Run

cargo run --release -p sdk --example tu_nlp_byte_transformer

Regenerate the training-loss image from the repository root:

cargo run --release -p sdk --example tu_nlp_byte_transformer -- --plot sdk/rs/tutorials/ml/nlp/tu_nlp_byte_transformer/asset/training-curve.png

On the default recipe the tutorial self-tests: loss regime, accuracy floor, exact reference generation, and .oam checkpoint round-trip.

Terminal output

Recorded canonical 300-step run on Intel Iris Xe (ADL GT2), Mesa 26.2.7. Startup logs, intermediate progress updates and the plot-save path are omitted. Timings vary by run; these are not new benchmark baselines.

╔══════════════════════════════════════════════════════════════════╗
║ OA Tutorial — Byte Transformer · all-position LM (autograd) ║
╚══════════════════════════════════════════════════════════════════╝
tokenizer: raw bytes · vocab=256
Task: dense next-byte prediction at every position
Model: Byte + position Embedding → Transformer(32,64) → LN → Linear(32→256)
params: 25760 Optimizer: AdamW(lr=0.01)
300/300 |██████████| 2.14s · 7.13 ms/step · 8.98K sample/s · 143.68K byte token/s · 143.68K byte/s · loss: 0.2005
Summary:
loss: initial 5.656755 · final 0.200502 · mean 0.756808
Wall: 7.13 ms/step · 8.98K sample/s · 143.67K byte token/s · 143.67K byte/s
GPU (byte_transformer_training_step): mean 6.310 ms/step · p50 4.861 · p95 13.006 · min 4.053 · max 16.328 · 10.14K sample/s · 162.28K byte token/s · 162.28K byte/s · wall-GPU gap 11%
run: 2.14s · 300 steps · batch 64 · sequence 16 byte token/sample
Evaluation:
Random-loss baseline ln(256) = 5.5452
byte-token accuracy: 92.9%
post-update evaluation loss: 0.190566
Generation:
prompt: "to be"
generated: "to be that is the question whether tis nobler in the mind to suffer the slings and ar"
program: 46 nodes · 1 recording · 299 cache hits · 0 uploads
checkpoint: optimizer step, accuracy, and exact generation restored

Source