Function sam3_build_text_encoder_graph

sam3.cpp:4040–4068

Build the full text encoder computation graph. token_ids: [L] int32 tensor (BPE token IDs, padded to ctx_len with 0s). Must be marked as input by caller; data uploaded after alloc. Returns: text_features tensor [text_out_dim, L] = [256, L]. Also creates the causal mask internally (marked as input).

// Returns: text_features tensor [text_out_dim, L] = [256, L].
// Also creates the causal mask internally (marked as input).
static struct ggml_tensor* sam3_build_text_encoder_graph(struct ggml_context* ctx,
                                                         struct ggml_tensor* token_ids,
                                                         const sam3_model& model) {
    const auto& hp = model.hparams;
    const auto& enc = model.text_enc;
    const int L = hp.text_ctx_len;  // 32

    auto* x = ggml_get_rows(ctx, enc.token_embed_w, token_ids);
    ggml_set_name(x, "text_token_embed");

    x = ggml_add(ctx, x, enc.pos_embed);
    ggml_set_name(x, "text_after_pos_embed");

    auto* causal_mask = sam3_build_causal_mask(ctx, L);

    for (int i = 0; i < hp.text_layers; ++i) {
        x = sam3_text_block_forward(ctx, x, enc.blocks[i], hp, causal_mask, i);
    }

    x = sam3_layer_norm(ctx, x, enc.ln_final_w, enc.ln_final_b);
    ggml_set_name(x, "text_final_ln");

    // Resizer: project 1024 → 256
    x = ggml_mul_mat(ctx, enc.resizer_w, x);
    x = ggml_add(ctx, x, enc.resizer_b);
    ggml_set_name(x, "text_features_2d");

    return x;
}

/*****************************************************************************
** SAM2 — Hiera Backbone Graph Building