Realm

tu_nlp_byte_moe — Byte × sparse MoE Transformer

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

Experimental

The byte-tokenizer × sparse MoE 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 → Attention + MoE(E=4,K=2,DFF=16) → LN → Linear
Parameters28,068
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 (64 nodes for this model).

Measured — Surface Laptop 5 (ADL GT2)

Byte × sparse MoE Transformer training loss

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

Loss init→finalAccuracyProgram nodesWall ms/stepToken/s
5.687→0.19992.6%649.86103.9K

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_moe

Regenerate the training-loss image from the repository root:

cargo run --release -p sdk --example tu_nlp_byte_moe -- --plot sdk/rs/tutorials/ml/nlp/tu_nlp_byte_moe/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 MoE Transformer · all-position LM (autograd)║
╚══════════════════════════════════════════════════════════════════╝
tokenizer: raw bytes · vocab=256
Task: dense next-byte prediction at every position
Model: Byte + position Embedding → Attention + MoE(E=4,K=2,DFF=16) → LN → Linear
params: 28068 Optimizer: AdamW(lr=0.01)
300/300 |██████████| 3.08s · 10.26 ms/step · 6.24K sample/s · 99.76K byte token/s · 99.76K byte/s · loss: 0.1991
Summary:
loss: initial 5.686717 · final 0.199083 · mean 0.805950
Wall: 10.26 ms/step · 6.23K sample/s · 99.76K byte token/s · 99.76K byte/s
GPU (byte_moe_training_step): mean 9.265 ms/step · p50 7.190 · p95 16.982 · min 6.014 · max 19.805 · 6.91K sample/s · 110.52K byte token/s · 110.52K byte/s · wall-GPU gap 10%
run: 3.08s · 300 steps · batch 64 · sequence 16 byte token/sample
Evaluation:
Random-loss baseline ln(256) = 5.5452
byte-token accuracy: 92.6%
post-update evaluation loss: 0.193275
Generation:
prompt: "to be"
generated: "to be that is the question whether tis nobler in the mind to suffer the slings and ar"
program: 64 nodes · 1 recording · 299 cache hits · 0 uploads
checkpoint: optimizer step, accuracy, and exact generation restored

Source