Realm

tu_nlp_char_moe_transformer — Character × sparse MoE Transformer

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

Experimental

The character-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 27-token character vocabulary (CharSampler) — accuracy is comparable within Char rows only
Modeltoken+position embedding → attention → sparse MoE(E=4,K=2,DFF=16) → LayerNorm → Linear
Parameters13,183
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)

Character × sparse MoE Transformer training loss

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

Loss init→finalAccuracyProgram nodesWall ms/stepToken/s
3.514→0.19092.6%628.41121.8K

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_char_moe_transformer
cargo run --release -p sdk --example tu_nlp_char_moe_transformer -- --plot sdk/rs/tutorials/ml/nlp/tu_nlp_char_moe_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 — Char MoE Transformer · all-position LM ║
╚══════════════════════════════════════════════════════════════════╝
tokenizer: character · vocab=27 (a-z + space)
Task: dense next-character prediction at every position
Model: token+position embedding → attention → sparse MoE(E=4, K=2, DFF=16) → LayerNorm → Linear
params: 13183 Optimizer: AdamW(lr=0.01)
training: 300 steps · batch=64 · sequence=16 character tokens
300/300 |██████████| 2.32s · 7.74 ms/step · 8.27K sample/s · 132.37K token/s · loss: 0.1904
Summary:
loss: initial 3.514067 · final 0.190367 · mean 0.506875
Wall: 7.74 ms/step · 8.27K sample/s · 132.37K token/s
GPU (char_moe_transformer_training_step): mean 6.636 ms/step · p50 5.095 · p95 13.775 · min 4.599 · max 15.515 · 9.64K sample/s · 154.31K token/s · wall-GPU gap 14%
run: 2.32s · 300 steps · batch 64 · sequence 16 token/sample
Evaluation:
Random-loss baseline ln(27) = 3.2958
loss: evaluation 0.183673 · last training 0.190367
character-token accuracy: 92.6%
Generation:
prompt: "to be"
generated: "to be that is the question whether tis nobler in the mind to suffer the slings and ar"
program: 62 nodes · 1 recording · 299 cache hits · 0 uploads

Source