Method encode

smallthinker/src/llama-context.cpp:704–872

}

int llama_context::encode(llama_batch & inp_batch) {
    if (inp_batch.n_tokens == 0) {
        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

    // temporary allocate memory for the input batch if needed
    // note: during encode, we always pass the full sequence starting from pos = 0
    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);

    const llama_batch & batch = batch_allocr.batch;
    const int32_t n_tokens = batch.n_tokens;

    const auto & hparams = model.hparams;

    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT

    // TODO: move the validation to the llama_batch_allocr
    if (batch.token) {
        for (int32_t i = 0; i < n_tokens; ++i) {
            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                return -1;
            }

            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
                LLAMA_LOG_ERROR("%s: invalid seq_id[%d] = %d > %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
                throw -1;
            }
        }
    }

    // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");

    if (t_compute_start_us == 0) {
        t_compute_start_us = ggml_time_us();
    }

    embd_seq.clear();

    n_queued_tokens += n_tokens;

    const int64_t n_embd = hparams.n_embd;

    llama_sbatch sbatch = llama_sbatch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);

    const llama_ubatch ubatch = sbatch.split_simple(n_tokens);

    // reserve output buffer
    if (output_reserve(n_tokens) < n_tokens) {
        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
        return -2;
    };

    for (int32_t i = 0; i < n_tokens; ++i) {
        output_ids[i] = i;
    }