Python tutorial

Character GRU language model

The public oa.core, oa.ml, and oa.runtime bindings drive the same all-position character task as the C++ tutorial.

oa.mlNative Vulkan execution300 steps · batch 64

Contract	Value
Imports	oa.core, oa.ml, oa.runtime through the public oa package
Model	OaEmbedding → OaGru → OaLinear
Autograd	ml.GradientTape and ml.OaAdamW
Validation	Loss, accuracy, generation, and checkpoint round-trip

Model

Tutorialnlpchargruag.py

class CharGruLM:
    def __init__(self):
        self.embed = ml.OaEmbedding(27, nlp.D_MODEL)
        self.gru = ml.OaGru(nlp.D_MODEL, nlp.HIDDEN_DIM, 1)
        self.head = ml.OaLinear(nlp.HIDDEN_DIM, 27)
        for parameter in self.parameters():
            parameter.Data.SetRequiresGrad(True)

    def forward(self, tokens):
        batch, sequence = tokens.Size(0), tokens.Size(1)
        embedded = self.embed.Forward(tokens).Reshape(
            [batch, sequence, nlp.D_MODEL]
        )
        hidden = self.gru.Forward(embedded).Reshape(
            [batch * sequence, nlp.HIDDEN_DIM]
        )
        return self.head.Forward(hidden)

Train and evaluate

Tutorialnlpchargruag.py

runtime.OaInitComputeEngine()
model = CharGruLM()
sampler = nlp.NlpAllPositionSampler(
    nlp.CORPUS, nlp.BATCH, encode=nlp.char_encode
)

optimizer, initial_loss, final_loss, x, y = nlp.train_all_position(
    model.forward,
    model.parameters(),
    sampler,
    lr=0.01,
    timer_name="char_gru_allpos_step",
)

accuracy = nlp.accuracy_all_positions(model.forward, x, y, 27)
generated = nlp.generate_greedy(
    model.forward, "hello", 48, 27, encode=nlp.char_encode
)

No Python tensor fork

Python objects wrap the same OaMatrix and module implementations used by C++.

Shared test scaffolding

_nlp_common.py owns the corpus, sampler, metrics, greedy generation, and training loop.

Explicit runtime lifetime

The tutorial initializes OA once, records GPU work through native calls, and synchronizes only at deliberate boundaries.

Generated reference

The Python API pages list only symbols currently registered by Source/Python.

Run

Terminal

OA_PYTHON_BUILD_DIR=Build/Release PYTHONPATH=Tutorial/Py Build/PythonVenv/bin/python Tutorial/Py/TutorialNlpCharGruAg.py

View current source