Skip to main content

ClikaRT::nn::QEmbedding

class

Header: ClikaRT/nn/qembedding.h
Inherits: ClikaRT::nn::Module

Quantized embedding-table module: gathers rows of a packed table by token index and decodes them to float.

The table is [V, E] (one row per vocabulary entry); forward(ids) returns ids.shape ++ [E]. The codes are never densely materialized: the gather decodes the rows it reads. A dense table is Embedding's.

auto emb = ClikaRT::nn::QEmbedding::make(weight);   // a QTensor
auto h   = emb->forward(token_ids);                  // [B, S] -> [B, S, E]

Static member functions​

make(QTensor, QEmbeddingOptions)​

static std::shared_ptr<QEmbedding> make(QTensor weight, QEmbeddingOptions options = {})

Defaults: every option at its default (QEmbeddingOptions).

Declared in ClikaRT/nn/qembedding.h, line 81

make(int64_t, int64_t, QEmbeddingOptions)​

static std::shared_ptr<QEmbedding> make(
    std::int64_t num_embeddings,
    std::int64_t embedding_dim,
    QEmbeddingOptions options = {}
)

Defaults: every option at its default (QEmbeddingOptions).

Declared in ClikaRT/nn/qembedding.h, line 95

Member functions​

set_weights()​

void set_weights(QTensor weight)

Bind the weight slot: a quantized payload whose logical shape is the declared [num_embeddings, embedding_dim]; the declared placement wins; the codes are never cast. Re-binding drops the built lookup; the next forward rebuilds. Raises ClikaRT::Error on a geometry mismatch or a weight without a scheme.

Declared in ClikaRT/nn/qembedding.h, line 107

~QEmbedding()​

~QEmbedding() override

Declared in ClikaRT/nn/qembedding.h, line 111

to_impl(StreamOrDevice)​

virtual Result<void> to_impl(StreamOrDevice where) override

Move to a placement (Device / Stream): the packed table moves with the module.

Declared in ClikaRT/nn/qembedding.h, line 115

to_impl(DataType)​

virtual Result<void> to_impl(DataType dtype) override

The weight stays quantized at rest: a dtype move is Unsupported (decode explicitly, ops::dequantize, when a dense copy is wanted).

Declared in ClikaRT/nn/qembedding.h, line 118

forward()​

Tensor forward(Tensor ids) const

forward_impl, unwrapped: raises ClikaRT::Error on failure.

Declared in ClikaRT/nn/qembedding.h, line 127

operator()()​

Tensor operator()(Tensor ids) const

Same as forward.

Declared in ClikaRT/nn/qembedding.h, line 131

num_embeddings()​

std::int64_t num_embeddings() const noexcept

The declared row count.

Declared in ClikaRT/nn/qembedding.h, line 136

embedding_dim()​

std::int64_t embedding_dim() const noexcept

The declared row length.

Declared in ClikaRT/nn/qembedding.h, line 138

options()​

const QEmbeddingOptions& options() const noexcept

The settings this module was made with.

Declared in ClikaRT/nn/qembedding.h, line 140

Protected member functions​

initialize_impl()​

virtual Result<void> initialize_impl() override

The pack hook (Module::initialize_impl): builds the lookup now (idempotent, thread-safe). The weight slot must be bound; a still-declared slot is a clean error. After the build the registry slot reads back from the packed table.

Declared in ClikaRT/nn/qembedding.h, line 152