Realm

tu_nlp_bpe_mamba3 — BPE × Mamba-3

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

Experimental

The BPE-tokenizer × Mamba-3 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 Embedding → Mamba-3(32,state=32,expand=2) + residual → Linear
Parameters29,960
OptimizerAdamW(lr=0.003) — lower rate for the Mamba-3/Empyrealm family

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 (34 nodes for this model).

Measured — Surface Laptop 5 (ADL GT2)

BPE × Mamba-3 training loss

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

Loss init→finalAccuracyProgram nodesWall ms/stepToken/s
5.775→0.02098.9%3412.2383.7K

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

1prompt: "to be"
2generated: "to beion wher tis nobler in the mind to suffer the slings and arrows of outrageous fo"
3

Run

cargo run --release -p sdk --example tu_nlp_bpe_mamba3

Regenerate the training-loss image from the repository root:

cargo run --release -p sdk --example tu_nlp_bpe_mamba3 -- --plot sdk/rs/tutorials/ml/nlp/tu_nlp_bpe_mamba3/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 Mamba-3 · 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 Embedding → Mamba-3(32,state=32,expand=2) + residual → Linear
params: 29960 Optimizer: AdamW(lr=0.003)
300/300 |██████████| 3.33s · 11.09 ms/step · 5.77K sample/s · 92.31K BPE token/s · 256.20K byte/s · loss: 0.0200
Summary:
loss: initial 5.774934 · final 0.020031 · mean 0.748239
Wall: 11.09 ms/step · 5.77K sample/s · 92.31K BPE token/s · 256.20K byte/s
GPU (bpe_mamba3_training_step): mean 9.971 ms/step · p50 7.788 · p95 16.138 · min 7.191 · max 19.297 · 6.42K sample/s · 102.70K BPE token/s · 285.04K byte/s · wall-GPU gap 10%
run: 3.33s · 300 steps · batch 64 · sequence 16 BPE token/sample
Evaluation:
Random-loss baseline ln(320) = 5.7683
final batch: 2.771 byte/token · 0.0104 bits/byte
BPE-token accuracy: 98.9%
post-update evaluation loss: 0.019415
Generation:
prompt: "to be"
generated: "to beion wher tis nobler in the mind to suffer the slings and arrows of outrageous fo"
program: 34 nodes · 1 recording · 299 cache hits · 0 uploads

Source