Realm

tu_nlp_bpe_moe — BPE × sparse MoE Transformer

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

Experimental

The BPE-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 320-token learned BPE vocabulary (BpeTokenizer/BpeSampler: 256 bytes + 64 merges) — merges collapse the corpus into short predictable runs, which is why BPE rows reach ~0.02 loss
ModelBPE + position Embedding → Attention + MoE(E=4,K=2,DFF=16) → LN → Linear
Parameters32,228
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 (62 nodes for this model).

Measured — Surface Laptop 5 (ADL GT2)

BPE × sparse MoE Transformer training loss

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

Loss init→finalAccuracyProgram nodesWall ms/stepToken/s
5.844→0.02098.8%6210.4398.2K

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

1prompt: "to be"
2generated: "to bee or not to be that is the question whether tis nobler in the mind to suffer the"
3

Run

cargo run --release -p sdk --example tu_nlp_bpe_moe

Regenerate the training-loss image from the repository root:

cargo run --release -p sdk --example tu_nlp_bpe_moe -- --plot sdk/rs/tutorials/ml/nlp/tu_nlp_bpe_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 — BPE MoE Transformer · all-position LM (autograd)║
╚══════════════════════════════════════════════════════════════════╝
tokenizer: byte BPE · vocab=320 (256 bytes + 64 merges)
compression: 576 bytes → 210 tokens · 2.743 byte/token · 63.5% fewer positions
context coverage: 16 BPE tokens ≈ 43.9 source bytes at corpus average
Task: dense next-token prediction at every position
Model: BPE + position Embedding → Attention + MoE(E=4,K=2,DFF=16) → LN → Linear
params: 32228 Optimizer: AdamW(lr=0.01)
300/300 |██████████| 3.01s · 10.02 ms/step · 6.39K sample/s · 102.21K BPE token/s · 283.69K byte/s · loss: 0.0204
Summary:
loss: initial 5.844332 · final 0.020420 · mean 0.232860
Wall: 10.02 ms/step · 6.39K sample/s · 102.21K BPE token/s · 283.68K byte/s
GPU (bpe_moe_training_step): mean 8.785 ms/step · p50 6.953 · p95 14.975 · min 5.817 · max 15.641 · 7.29K sample/s · 116.57K BPE token/s · 323.53K byte/s · wall-GPU gap 12%
run: 3.01s · 300 steps · batch 64 · sequence 16 BPE token/sample
Evaluation:
Random-loss baseline ln(320) = 5.7683
final batch: 2.771 byte/token · 0.0106 bits/byte
BPE-token accuracy: 98.8%
post-update evaluation loss: 0.019962
Generation:
prompt: "to be"
generated: "to bee or not to be that is the question whether tis nobler in the mind to suffer the"
program: 62 nodes · 1 recording · 299 cache hits · 0 uploads

Source