Function ggml_opt_lbfgs

ggml.c:18606–18872

}

static enum ggml_opt_result ggml_opt_lbfgs(
        struct ggml_context * ctx,
        struct ggml_opt_context * opt,
        struct ggml_opt_params params,
        struct ggml_tensor * f,
        struct ggml_cgraph * gf,
        struct ggml_cgraph * gb,
        ggml_opt_callback callback,
        void * callback_data) {
    if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE ||
        params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
        if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) {
            return GGML_OPT_INVALID_WOLFE;
        }
    }

    const int m = params.lbfgs.m;

    // these will store the parameters we want to optimize
    struct ggml_tensor * ps[GGML_MAX_PARAMS];

    int np = 0;
    int nx = 0;
    for (int i = 0; i < gf->n_nodes; ++i) {
        if (gf->nodes[i]->is_param) {
            GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);

            GGML_ASSERT(np < GGML_MAX_PARAMS);

            ps[np++] = gf->nodes[i];
            nx += ggml_nelements(gf->nodes[i]);
        }
    }

    if ((opt->params.type != params.type) || (opt->nx != nx) || (opt->params.past != params.past) || (opt->params.lbfgs.m != params.lbfgs.m)) {
        int iter = opt->iter;
        ggml_opt_init(ctx, opt, params, nx);
        opt->iter = iter;
    }

    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
    cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;

    float * x  = opt->lbfgs.x->data;  // current parameters
    float * xp = opt->lbfgs.xp->data; // previous parameters
    float * g  = opt->lbfgs.g->data;  // current gradient
    float * gp = opt->lbfgs.gp->data; // previous gradient
    float * d  = opt->lbfgs.d->data;  // search direction

    float * pf = params.past > 0 ? opt->lbfgs.pf->data : NULL; // past function values

    const int n_accum = MAX(1, params.n_gradient_accumulation);
    const float accum_norm = 1.0f / (float) n_accum;

    float fx    = 0.0f; // cost function value
    float xnorm = 0.0f; // ||x||
    float gnorm = 0.0f; // ||g||