add gradient clipping to AdamW

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

@@ -2430,7 +2430,8 @@ void write_opt_context(struct llama_file * file, struct ggml_opt_context * opt)
     GGML_ASSERT(opt->nx   >= 0);
     GGML_ASSERT(opt->iter >= 0);
     file->write_u32(version);
-    file->write_raw(&opt->params, sizeof(opt->params));
+    file->write_u32(opt->params.past);
+    file->write_u32(opt->params.lbfgs.m);
     file->write_raw(&opt->nx,     sizeof(opt->nx));
     file->write_raw(&opt->iter,   sizeof(opt->iter));
     file->write_u32((uint32_t)  opt->just_initialized);
@@ -2469,8 +2470,43 @@ void write_opt_context(struct llama_file * file, struct ggml_opt_context * opt)
     }
 }
 
+struct ggml_opt_params_v0 {
+    enum ggml_opt_type type;
+    int n_threads;
+    int past;
+    float delta;
+    int max_no_improvement;
+    bool print_forward_graph;
+    bool print_backward_graph;
+    struct {
+        int n_iter;
+        float sched;
+        float decay;
+        float alpha;
+        float beta1;
+        float beta2;
+        float eps;
+        float eps_f;
+        float eps_g;
+    } adam;
+    struct {
+        int m;
+        int n_iter;
+        int max_linesearch;
+        float eps;
+        float ftol;
+        float wolfe;
+        float min_step;
+        float max_step;
+        enum ggml_linesearch linesearch;
+    } lbfgs;
+};
+
 void read_opt_context_v0(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) {
-    file->read_raw(&opt->params, sizeof(opt->params));
+    ggml_opt_params_v0 pv0;
+    file->read_raw(&pv0, sizeof(pv0));
+    opt->params.past = pv0.past;
+    opt->params.lbfgs.m = pv0.lbfgs.m;
     file->read_raw(&opt->nx, sizeof(opt->nx));
     ggml_opt_init(ctx, opt, opt->params, opt->nx);
 
@@ -2516,7 +2552,8 @@ void read_opt_context_v0(struct llama_file * file, struct ggml_context * ctx, st
 }
 
 void read_opt_context_v1(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) {
-    file->read_raw(&opt->params, sizeof(opt->params));
+    opt->params.past    = (int) file->read_u32();
+    opt->params.lbfgs.m = (int) file->read_u32();
     file->read_raw(&opt->nx,     sizeof(opt->nx));
     ggml_opt_init(ctx, opt, opt->params, opt->nx);
 
@@ -2558,6 +2595,7 @@ void read_opt_context_v1(struct llama_file * file, struct ggml_context * ctx, st
 
 void read_opt_context(struct llama_file * file, struct ggml_context * ctx, struct ggml_opt_context * opt) {
     uint32_t version = file->read_u32();
+    printf("%s: opt context version %u\n", __func__, version);
     switch (version) {
         case 0:
             {
@@ -2569,7 +2607,7 @@ void read_opt_context(struct llama_file * file, struct ggml_context * ctx, struc
             } break;
         default:
             {
-                fprintf(stderr, "%s: unknown version %ud\n", __func__, version);
+                fprintf(stderr, "%s: unknown version %u\n", __func__, version);
             }
     }
 }
@@ -2783,6 +2821,9 @@ struct train_params {
     int   adam_n_iter;
     float adam_alpha;
     float adam_decay;
+    float adam_beta1;
+    float adam_beta2;
+    float adam_gclip;
 
     int mem_model_gb;
     int mem_compute_gb;
@@ -2830,6 +2871,9 @@ struct train_params get_default_train_params() {
     params.adam_n_iter       = 16;
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
+    params.adam_beta1        = 0.9f;
+    params.adam_beta2        = 0.999f;
+    params.adam_gclip        = 1.0f;
 
     params.mem_model_gb   = 2;
     params.mem_compute_gb = 24;
@@ -2877,6 +2921,9 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
     fprintf(stderr, "  --adam-iter N              Maximum number of Adam optimization iterations for each batch (default %d)\n", params->adam_n_iter);
     fprintf(stderr, "  --adam-alpha N             Adam learning rate alpha (default %f)\n", params->adam_alpha);
     fprintf(stderr, "  --adam-decay N             AdamW weight decay. Values greater zero enable AdamW instead of regular Adam. (default %f)\n", params->adam_decay);
+    fprintf(stderr, "  --adam-beta1 N             AdamW beta1 in interval [0,1). How much to smooth the first moment of gradients. (default %f)\n", params->adam_beta1);
+    fprintf(stderr, "  --adam-beta2 N             AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
+    fprintf(stderr, "  --adam-gclip N             AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
     fprintf(stderr, "  --mem-model N              Memory to allocate for model and cache in gigabytes. (default %d)\n", params->mem_model_gb);
     fprintf(stderr, "  --mem-compute N            Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute_gb);
     fprintf(stderr, "  --mem-compute0 N           Memory to allocate for compute in gigabytes. (default %d)\n", params->mem_compute0_gb);
@@ -3066,6 +3113,24 @@ bool train_params_parse(int argc, char ** argv, struct train_params * params) {
                 break;
             }
             params->adam_decay = std::stof(argv[i]);
+        } else if (arg == "--adam-beta1") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->adam_beta1 = std::stof(argv[i]);
+        } else if (arg == "--adam-beta2") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->adam_beta2 = std::stof(argv[i]);
+        } else if (arg == "--adam-gclip") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->adam_gclip = std::stof(argv[i]);
         } else if (arg == "--mem-model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -3212,6 +3277,9 @@ int main(int argc, char ** argv) {
     opt_params_adam.adam.sched  = 1.0f;
     opt_params_adam.adam.alpha  = params.adam_alpha;
     opt_params_adam.adam.decay  = params.adam_decay;
+    opt_params_adam.adam.beta1  = params.adam_beta1;
+    opt_params_adam.adam.beta2  = params.adam_beta2;
+    opt_params_adam.adam.gclip  = params.adam_gclip;
 
     opt_params_lbfgs.print_forward_graph = false;
     opt_params_lbfgs.print_backward_graph = false;

ggml.c

@@ -17354,6 +17354,7 @@ static enum ggml_opt_result ggml_opt_adam(
     const float beta1 = params.adam.beta1;
     const float beta2 = params.adam.beta2;
     const float eps   = params.adam.eps;
+    const float gclip = params.adam.gclip;
 
     float * m  = opt->adam.m->data;  // first moment
     float * v  = opt->adam.v->data;  // second moment
@@ -17404,16 +17405,34 @@ static enum ggml_opt_result ggml_opt_adam(
         UNUSED(t_start_cpu);
 
         {
+            float gnorm = 1.0f;
+            if (gclip > 0.0f) {
+                // gradient clipping
+                ggml_float sum = 0.0;
+                for (int p = 0; p < np; ++p) {
+                    const int64_t ne = ggml_nelements(ps[p]);
+                    for (int64_t j = 0; j < ne; ++j) {
+                        float g = ggml_get_f32_1d(ps[p]->grad, j);
+                        sum += g*g;
+                    }
+                }
+                ggml_float norm = sqrt(sum);
+                if (norm > (ggml_float) gclip) {
+                    gnorm = (float) ((ggml_float) gclip / norm);
+                }
+            }
+            const float beta1h = alpha/(1.0f - powf(beta1, opt->iter));
+            const float beta2h =  1.0f/(1.0f - powf(beta2, opt->iter));
             int64_t i = 0;
             for (int p = 0; p < np; ++p) {
-                const int64_t ne = ggml_nelements(ps[p]) ;
+                const int64_t ne = ggml_nelements(ps[p]);
                 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);
+                    float g = ggml_get_f32_1d(ps[p]->grad, j)*gnorm;
                     m[i] = m[i]*beta1 +   g*(1.0f - beta1);
                     v[i] = v[i]*beta2 + g*g*(1.0f - beta2);
-                    float mh = m[i]*alpha/(1.0f - powf(beta1, opt->iter));
+                    float mh = m[i]*beta1h;
-                    float vh = v[i]*1.0f /(1.0f - powf(beta2, opt->iter));
+                    float vh = v[i]*beta2h;
                     vh = sqrtf(vh) + eps;
                     x  = x*(1.0f - decay) - mh/vh;
                     ggml_set_f32_1d(ps[p], j, x);
@@ -17902,6 +17921,7 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                         .eps    = 1e-8f,
                         .eps_f  = 1e-5f,
                         .eps_g  = 1e-3f,
+                        .gclip  = 0.0f,
                     },
                 };
             } break;

ggml.h

@@ -1509,6 +1509,7 @@ extern "C" {
             float eps;   // epsilon for numerical stability
             float eps_f; // epsilon for convergence test
             float eps_g; // epsilon for convergence test
+            float gclip; // gradient clipping
         } adam;
 
         // LBFGS parameters