add gradient accumulation

specify number accumulation steps with '--grad-acc N'.
this will simulate a bigger batch size of grad_acc*batch.

examples/finetune/finetune.cpp

@@ -1730,6 +1730,7 @@ struct train_params {
     int n_ctx;
     int n_threads;
     int n_batch;
+    int n_gradient_accumulation;
 
     bool custom_n_ctx;
 
@@ -1804,6 +1805,7 @@ struct train_params get_default_train_params() {
     params.n_ctx      =  128;
     params.n_threads  =    6;
     params.n_batch    =    8;
+    params.n_gradient_accumulation = 1;
 
     params.custom_n_ctx = false;
 
@@ -1880,6 +1882,7 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
     fprintf(stderr, "  -c N, --ctx N              Context size used during training (default %d)\n", params->n_ctx);
     fprintf(stderr, "  -t N, --threads N          Number of threads (default %d)\n", params->n_threads);
     fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
+    fprintf(stderr, "  --grad-acc N               Number of gradient accumulation steps (simulates larger batch size of batch*gradacc) (default %d)\n", params->n_gradient_accumulation);
     fprintf(stderr, "  --norm-rms-eps F           RMS-Norm epsilon value (default %f)\n", params->f_norm_rms_eps);
     fprintf(stderr, "  --rope-freq-base F         Frequency base for ROPE (default %f)\n", params->rope_freq_base);
     fprintf(stderr, "  --rope-freq-scale F        Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
@@ -2015,6 +2018,12 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
                 break;
             }
             params->n_batch = std::stoi(argv[i]);
+        } else if (arg == "--grad-acc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_gradient_accumulation = std::stoi(argv[i]);
         } else if (arg == "--norm-rms-eps") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -2299,13 +2308,14 @@ void print_duration(float fmillis) {
     printf("%02lld:%02lld:%02lld", hours, minutes, seconds);
 }
 
-void opt_callback(void * vdata, float * sched) {
+void opt_callback(void * vdata, int accum_step, float * sched) {
     struct opt_callback_data * data = (struct opt_callback_data *) vdata;
     struct train_params * params    = data->params;
     struct ggml_opt_context * opt   = data->opt;
     int n_batch = params->n_batch;
     int n_ctx = params->n_ctx;
 
+    if (accum_step == 0) {
         int64_t now = ggml_time_ms();
         if (now > data->last_time) {
             float dt = now - data->last_time;
@@ -2376,6 +2386,7 @@ void opt_callback(void * vdata, float * sched) {
         printf(">");
         // printf("improvement: %*d>", impr_plot, (int)0);
         printf("\n");
+    }
 
     if (data->shuffle_countdown < n_batch) {
         printf("%s: reshuffle samples\n", __func__);
@@ -2497,6 +2508,7 @@ int main(int argc, char ** argv) {
         opt->params.past                    = params.opt_past;
         opt->params.delta                   = params.opt_delta;
         opt->params.max_no_improvement      = params.opt_max_no_improvement;
+        opt->params.n_gradient_accumulation = params.n_gradient_accumulation;
         opt->params.adam.n_iter             = params.adam_n_iter;
         opt->params.adam.sched              = 1.0f;
         opt->params.adam.alpha              = params.adam_alpha;
@@ -2514,6 +2526,7 @@ int main(int argc, char ** argv) {
         opt->params.past                    = params.opt_past;
         opt->params.delta                   = params.opt_delta;
         opt->params.max_no_improvement      = params.opt_max_no_improvement;
+        opt->params.n_gradient_accumulation = params.n_gradient_accumulation;
         opt->params.lbfgs.n_iter            = params.lbfgs_n_iter;
     }
 

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1299,8 +1299,9 @@ struct train_params {
     int n_ff;
 
     int n_threads;
-    int n_batch;
     int n_examples;
+    int n_batch;
+    int n_gradient_accumulation;
 
     float f_norm_rms_eps;
     float rope_freq_base;
@@ -1362,8 +1363,9 @@ struct train_params get_default_train_params() {
     params.n_ff       =  768;
 
     params.n_threads  =    6;
-    params.n_batch    =    8;
     params.n_examples =    1;
+    params.n_batch    =    8;
+    params.n_gradient_accumulation = 1;
 
     params.f_norm_rms_eps  = 1e-5f;
     params.rope_freq_base  = 10000.0f;
@@ -1428,8 +1430,9 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
     fprintf(stderr, "  --rope-freq-base F         Frequency base for ROPE (default %f)\n", params->rope_freq_base);
     fprintf(stderr, "  --rope-freq-scale F        Frequency scale for ROPE (default %f)\n", params->rope_freq_scale);
     fprintf(stderr, "  -t N, --threads N          Number of threads (default %d)\n", params->n_threads);
-    fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
     fprintf(stderr, "  -n N, --examples N         Number of examples to train (default %d)\n", params->n_examples);
+    fprintf(stderr, "  -b N, --batch N            Parallel batch size (default %d)\n", params->n_batch);
+    fprintf(stderr, "  --grad-acc N               Number of gradient accumulation steps (simulates larger batch size of batch*gradacc) (default %d)\n", params->n_gradient_accumulation);
     fprintf(stderr, "  --print-info-interval N    Print infos during training each N examples (default %d)\n", params->print_info_interval);
     fprintf(stderr, "  --samples-after-nl         Training samples start after newlines. (default %s)\n", params->samples_start_after_nl ? "on" : "off");
     fprintf(stderr, "  --use-lbfgs                Use LBFGS optimizer instead of default Adam\n");
@@ -1591,6 +1594,12 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
                 break;
             }
             params->n_batch = std::stoi(argv[i]);
+        } else if (arg == "--grad-acc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_gradient_accumulation = std::stoi(argv[i]);
         } else if (arg == "-n" || arg == "--examples") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -1779,13 +1788,14 @@ struct opt_callback_data {
     struct ggml_tensor *      target_probs;
 };
 
-void opt_callback(void * vdata, float * sched) {
+void opt_callback(void * vdata, int accum_step, float * sched) {
     struct opt_callback_data * data = (struct opt_callback_data *) vdata;
     struct train_params * params    = data->params;
     struct ggml_opt_context * opt   = data->opt;
     int n_batch = params->n_batch;
     int n_ctx = params->n_ctx;
 
+    if (accum_step == 0) {
         const bool save_now = (params->save_every > 0) && (opt->iter - data->last_save_iter >= params->save_every);
         if (save_now) {
             int new_iters = opt->iter - data->last_save_iter;
@@ -1819,6 +1829,8 @@ void opt_callback(void * vdata, float * sched) {
         int impr_plot = std::isnan(opt->loss_after) ? 0 : -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
         printf("%s: iter=%*d, sched=%f loss0=%f loss=%f | improvement: %*d>\n", __func__, 6, opt->iter, *sched, opt->loss_before, opt->loss_after, impr_plot, (int)0);
 
+    }
+
     if (data->shuffle_countdown < n_batch) {
         printf("%s: reshuffle samples\n", __func__);
         shuffle_ints(data->samples_data, data->samples_data + data->samples_size);
@@ -1923,6 +1935,7 @@ int main(int argc, char ** argv) {
     opt_params_adam.past                    = params.opt_past;
     opt_params_adam.delta                   = params.opt_delta;
     opt_params_adam.max_no_improvement      = params.opt_max_no_improvement;
+    opt_params_adam.n_gradient_accumulation = params.n_gradient_accumulation;
     opt_params_adam.adam.n_iter             = params.adam_n_iter;
     opt_params_adam.adam.sched              = 1.0f;
     opt_params_adam.adam.alpha              = params.adam_alpha;
@@ -1936,9 +1949,10 @@ int main(int argc, char ** argv) {
     opt_params_lbfgs.print_forward_graph     = false;
     opt_params_lbfgs.print_backward_graph    = false;
     opt_params_lbfgs.n_threads               = params.n_threads;
-    opt_params_adam.past                  = params.opt_past;
+    opt_params_lbfgs.past                    = params.opt_past;
-    opt_params_adam.delta                 = params.opt_delta;
+    opt_params_lbfgs.delta                   = params.opt_delta;
-    opt_params_adam.max_no_improvement    = params.opt_max_no_improvement;
+    opt_params_lbfgs.max_no_improvement      = params.opt_max_no_improvement;
+    opt_params_lbfgs.n_gradient_accumulation = params.n_gradient_accumulation;
     opt_params_lbfgs.lbfgs.n_iter            = params.lbfgs_n_iter;
 
     opt->ctx = model.ctx;

ggml.c

@@ -19112,7 +19112,7 @@ static void ggml_opt_get_params(int np, struct ggml_tensor * const ps[], float *
 }
 
 static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g) {
-    int i = 0;
+    int64_t i = 0;
     for (int p = 0; p < np; ++p) {
         const int64_t ne = ggml_nelements(ps[p]) ;
         // TODO: add function to get all elements at once
@@ -19122,6 +19122,17 @@ static void ggml_opt_get_grad(int np, struct ggml_tensor * const ps[], float * g
     }
 }
 
+static void ggml_opt_acc_grad(int np, struct ggml_tensor * const ps[], float * g, float scale) {
+    int64_t i = 0;
+    for (int p = 0; p < np; ++p) {
+        const int64_t ne = ggml_nelements(ps[p]) ;
+        // TODO: add function to get all elements at once
+        for (int64_t j = 0; j < ne; ++j) {
+            g[i++] += ggml_get_f32_1d(ps[p]->grad, j) * scale;
+        }
+    }
+}
+
 //
 // ADAM
 //
@@ -19170,26 +19181,37 @@ static enum ggml_opt_result ggml_opt_adam(
     const float eps   = params.adam.eps;
     const float gclip = params.adam.gclip;
     const int decay_min_ndim = params.adam.decay_min_ndim;
+    const int n_accum = MAX(1, params.n_gradient_accumulation);
+    const float accum_norm = 1.0f / (float) n_accum;
 
+    float * g  = opt->adam.g->data;  // gradients
     float * m  = opt->adam.m->data;  // first moment
     float * v  = opt->adam.v->data;  // second moment
 
     float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values
 
-    if (callback) {
-        callback(callback_data, &sched);
-    }
-
-    // compute the function value
-    // ggml_graph_reset  (gf);
-    ggml_set_f32      (f->grad, 1.0f);
-
     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;
-    ggml_graph_compute(gb, &cplan);
 
-    opt->adam.fx_prev = ggml_get_f32_1d(f, 0);
+
+    // compute the function value
+
+    float fx = 0;
+    ggml_set_zero(opt->adam.g);
+    for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+        if (callback) {
+            callback(callback_data, accum_step, &sched);
+        }
+        // ggml_graph_reset  (gf);
+        ggml_set_f32      (f->grad, 1.0f);
+        ggml_graph_compute(gb, &cplan);
+        ggml_opt_acc_grad(np, ps, g, accum_norm);
+        fx += ggml_get_f32_1d(f, 0);
+    }
+    fx *= accum_norm;
+
+    opt->adam.fx_prev = fx;
     opt->adam.fx_best = opt->adam.fx_prev;
     if (pf) {
         pf[opt->iter % params.past] = opt->adam.fx_prev;
@@ -19234,12 +19256,8 @@ static enum ggml_opt_result ggml_opt_adam(
             if (gclip > 0.0f) {
                 // gradient clipping
                 ggml_float sum = 0.0;
-                for (int p = 0; p < np; ++p) {
+                for (int64_t i = 0; i < nx; ++i) {
-                    const int64_t ne = ggml_nelements(ps[p]);
+                    sum += (ggml_float)(g[i]*g[i]);
-                    for (int64_t j = 0; j < ne; ++j) {
-                        float g = ggml_get_f32_1d(ps[p]->grad, j);
-                        sum += (ggml_float)(g*g);
-                    }
                 }
                 ggml_float norm = sqrt(sum);
                 if (norm > (ggml_float) gclip) {
@@ -19254,9 +19272,9 @@ static enum ggml_opt_result ggml_opt_adam(
                 const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
                 for (int64_t j = 0; j < ne; ++j) {
                     float x  = ggml_get_f32_1d(ps[p], j);
-                    float g = ggml_get_f32_1d(ps[p]->grad, j)*gnorm;
+                    float g_ = g[i]*gnorm;
-                    m[i] = m[i]*beta1 +   g*(1.0f - beta1);
+                    m[i] = m[i]*beta1 +    g_*(1.0f - beta1);
-                    v[i] = v[i]*beta2 + g*g*(1.0f - beta2);
+                    v[i] = v[i]*beta2 + g_*g_*(1.0f - beta2);
                     float mh = m[i]*beta1h;
                     float vh = v[i]*beta2h;
                     vh = sqrtf(vh) + eps;
@@ -19267,16 +19285,20 @@ static enum ggml_opt_result ggml_opt_adam(
             }
         }
 
+        fx = 0;
+        ggml_set_zero(opt->adam.g);
+        for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
             if (callback) {
-            callback(callback_data, &sched);
+                callback(callback_data, accum_step, &sched);
             }
-
             // ggml_graph_reset  (gf);
             ggml_set_f32      (f->grad, 1.0f);
-
             ggml_graph_compute(gb, &cplan);
+            ggml_opt_acc_grad(np, ps, g, accum_norm);
+            fx += ggml_get_f32_1d(f, 0);
+        }
+        fx *= accum_norm;
 
-        const float fx = ggml_get_f32_1d(f, 0);
         opt->loss_after = fx;
 
 
@@ -19373,6 +19395,9 @@ static enum ggml_opt_result linesearch_backtracking(
     const float dec = 0.5f;
     const float inc = 2.1f;
 
+    const int n_accum = MAX(1, params->n_gradient_accumulation);
+    const float accum_norm = 1.0f / (float) n_accum;
+
     if (*step <= 0.f) {
         return GGML_LINESEARCH_INVALID_PARAMETERS;
     }
@@ -19390,12 +19415,6 @@ static enum ggml_opt_result linesearch_backtracking(
     dgtest = params->lbfgs.ftol*dginit;
 
     while (true) {
-        if (callback) {
-            // LBFG-S does not support learning rate -> ignore learning schedule
-            float sched = 0;
-            callback(callback_data, &sched);
-        }
-
         ggml_vec_cpy_f32(nx, x, xp);
         ggml_vec_mad_f32(nx, x, d, *step);
 
@@ -19403,14 +19422,22 @@ static enum ggml_opt_result linesearch_backtracking(
         {
             ggml_opt_set_params(np, ps, x);
 
+            *fx = 0;
+            memset(g, 0, sizeof(float)*nx);
+            for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+                if (callback) {
+                    // LBFG-S does not support learning rate -> ignore learning schedule
+                    float sched = 0;
+                    callback(callback_data, accum_step, &sched);
+                }
                 // ggml_graph_reset  (gf);
                 ggml_set_f32      (f->grad, 1.0f);
-
                 ggml_graph_compute(gb, cplan);
+                ggml_opt_acc_grad(np, ps, g, accum_norm);
+                *fx += ggml_get_f32_1d(f, 0);
+            }
+            *fx *= accum_norm;
 
-            ggml_opt_get_grad(np, ps, g);
-
-            *fx = ggml_get_f32_1d(f, 0);
         }
 
         ++count;
@@ -19512,6 +19539,9 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
     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||
@@ -19525,24 +19555,25 @@ static enum ggml_opt_result ggml_opt_lbfgs(
     float * lm_s     = opt->lbfgs.lms->data;
     float * lm_y     = opt->lbfgs.lmy->data;
 
-    if (callback) {
-        // LBFG-S does not support learning rate -> ignore learning schedule
-        float sched = 0;
-        callback(callback_data, &sched);
-    }
-
     // evaluate the function value and its gradient
     {
         ggml_opt_set_params(np, ps, x);
 
+        fx = 0;
+        memset(g, 0, sizeof(float)*nx);
+        for (int accum_step = 0; accum_step < n_accum; ++accum_step) {
+            if (callback) {
+                // LBFG-S does not support learning rate -> ignore learning schedule
+                float sched = 0;
+                callback(callback_data, accum_step, &sched);
+            }
             // ggml_graph_reset  (gf);
             ggml_set_f32      (f->grad, 1.0f);
-
             ggml_graph_compute(gb, &cplan);
-
+            ggml_opt_acc_grad(np, ps, g, accum_norm);
-        ggml_opt_get_grad(np, ps, g);
+            fx += ggml_get_f32_1d(f, 0);
-
+        }
-        fx = ggml_get_f32_1d(f, 0);
+        fx *= accum_norm;
 
         opt->loss_before = fx;
         opt->loss_after  = fx;
@@ -19729,6 +19760,8 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                     .print_forward_graph  = true,
                     .print_backward_graph = true,
 
+                    .n_gradient_accumulation = 1,
+
                     .adam = {
                         .n_iter = 10000,
                         .sched  = 1.000f,
@@ -19757,6 +19790,8 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                     .print_forward_graph  = true,
                     .print_backward_graph = true,
 
+                    .n_gradient_accumulation = 1,
+
                     .lbfgs = {
                         .m              = 6,
                         .n_iter         = 100,
@@ -19790,7 +19825,7 @@ GGML_API void ggml_opt_init(
     if (opt->ctx == NULL) {
         struct ggml_init_params ctx_opt_params;
         if (opt->params.type == GGML_OPT_ADAM) {
-            ctx_opt_params.mem_size = GGML_MEM_ALIGN*2 + ggml_tensor_overhead()*2 + ggml_type_size(GGML_TYPE_F32)*nx*2;
+            ctx_opt_params.mem_size = GGML_MEM_ALIGN*3 + ggml_tensor_overhead()*3 + ggml_type_size(GGML_TYPE_F32)*nx*3;
             if (opt->params.past > 0) {
                 ctx_opt_params.mem_size += GGML_MEM_ALIGN + ggml_tensor_overhead() + ggml_type_size(GGML_TYPE_F32)*opt->params.past;
             }
@@ -19808,6 +19843,7 @@ GGML_API void ggml_opt_init(
     switch (opt->params.type) {
         case GGML_OPT_ADAM:
             {
+                opt->adam.g  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->adam.m  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->adam.v  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->adam.pf = params.past > 0

ggml.h

@@ -1708,7 +1708,7 @@ extern "C" {
         GGML_LINESEARCH_INVALID_PARAMETERS,
     };
 
-    typedef void (*ggml_opt_callback)(void * data, float * sched);
+    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched);
 
     // optimization parameters
     //
@@ -1739,6 +1739,8 @@ extern "C" {
         bool print_forward_graph;
         bool print_backward_graph;
 
+        int n_gradient_accumulation;
+
         // ADAM parameters
         struct {
             int n_iter;
@@ -1784,6 +1786,7 @@ extern "C" {
         float loss_after;
 
         struct {
+            struct ggml_tensor * g;  // current gradient
             struct ggml_tensor * m;  // first moment
             struct ggml_tensor * v;  // second moment
             struct ggml_tensor * pf; // past function values