Method encode

subprojects/llama.cpp/src/llama-context.cpp:1181–1344

}

int llama_context::encode(const llama_batch & batch_inp) {
    GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT

    if (batch_inp.n_tokens == 0) {
        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

    const auto & hparams = model.hparams;

    const int64_t n_embd  = hparams.n_embd_inp();
    const int64_t n_vocab = model.vocab.n_tokens();

    // note: during encode, we always pass the full sequence starting from pos = 0
    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
        return -1;
    }

    const uint32_t n_tokens = balloc->get_n_tokens();

    // [TAG_NO_CACHE_PAD]
    // TODO: add new split mode where we pad the input sequences so that ubatch.equal_seqs == true
    const llama_ubatch ubatch = balloc->split_simple(n_tokens);

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

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

    // TODO: this clear of the buffer can easily be forgotten - need something better
    embd_seq.clear();

    sched_reserve();

    n_queued_tokens += 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 (uint32_t i = 0; i < n_tokens; ++i) {
        output_ids[i] = i;
    }

    n_outputs = n_tokens;

    const auto causal_attn_org = cparams.causal_attn;

    // always use non-causal attention for encoder graphs
    // TODO: this is a tmp solution until we have a proper way to support enc-dec models
    //       ref: https://github.com/ggml-org/llama.cpp/pull/12181#issuecomment-2730451223
    cparams.causal_attn = false;