CUDA full GPU acceleration, KV cache in VRAM (#1827)

* Fixed CUDA RoPE

* ggml_cuda_mul_mat_vec_p021

* ggml_cuda_scale

* ggml_cuda_diag_mask_inf

* ggml_is_permuted

* ggml_cuda_cpy

* flatten rows for ggml_cuda_op

* Added a --low-vram option

* Fixed Windows performance

* Fixed LLAMA_CUDA_DMMV_Y > 1 for WizardLM

examples/CMakeLists.txt

@@ -37,6 +37,7 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
     add_subdirectory(baby-llama)
+    add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -79,34 +79,39 @@ struct ggml_tensor * randomize_tensor_normal(
         int ndims,
         const int64_t ne[],
         struct random_normal_distribution * rnd) {
+    float scale = 1.0; // xavier
     switch (ndims) {
         case 1:
+            scale /= sqrtf(ne[0]);
             for (int i0 = 0; i0 < ne[0]; i0++) {
-                ((float *)tensor->data)[i0] = frand_normal(rnd);
+                ((float *)tensor->data)[i0] = scale * frand_normal(rnd);
             }
             break;
         case 2:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i1 = 0; i1 < ne[1]; i1++) {
                 for (int i0 = 0; i0 < ne[0]; i0++) {
-                    ((float *)tensor->data)[i1*ne[0] + i0] = frand_normal(rnd);
+                    ((float *)tensor->data)[i1*ne[0] + i0] = scale * frand_normal(rnd);
                 }
             }
             break;
         case 3:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i2 = 0; i2 < ne[2]; i2++) {
                 for (int i1 = 0; i1 < ne[1]; i1++) {
                     for (int i0 = 0; i0 < ne[0]; i0++) {
-                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                        ((float *)tensor->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                     }
                 }
             }
             break;
         case 4:
+            scale /= sqrtf(ne[0]+ne[1]);
             for (int i3 = 0; i3 < ne[3]; i3++) {
                 for (int i2 = 0; i2 < ne[2]; i2++) {
                     for (int i1 = 0; i1 < ne[1]; i1++) {
                         for (int i0 = 0; i0 < ne[0]; i0++) {
-                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand_normal(rnd);
+                            ((float *)tensor->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = scale * frand_normal(rnd);
                         }
                     }
                 }
@@ -148,8 +153,8 @@ struct llama_hparams_lora {
     uint32_t n_rot   = 64;
     uint32_t n_lora  = 64;
 
-    bool operator!=(const llama_hparams & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+    bool operator!=(const llama_hparams_lora & other) const {
+        return memcmp(this, &other, sizeof(llama_hparams_lora)) != 0;
     }
 };
 

examples/common.cpp

@@ -331,6 +331,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
 #else
       fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--low-vram" || arg == "-lv") {
+#ifdef GGML_USE_CUBLAS
+            params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
@@ -479,6 +485,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
@@ -528,6 +535,7 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     lparams.n_gpu_layers = params.n_gpu_layers;
     lparams.main_gpu     = params.main_gpu;
     memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.low_vram     = params.low_vram;
     lparams.seed         = params.seed;
     lparams.f16_kv       = params.memory_f16;
     lparams.use_mmap     = params.use_mmap;

examples/common.h

@@ -21,15 +21,16 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed                           = -1;   // RNG seed
-    int32_t n_threads                      = get_num_physical_cores();
-    int32_t n_predict                      = -1;   // new tokens to predict
-    int32_t n_ctx                          = 512;  // context size
-    int32_t n_batch                        = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep                         = 0;    // number of tokens to keep from initial prompt
-    int32_t n_gpu_layers                   = 0;    // number of layers to store in VRAM
-    int32_t main_gpu                       = 0;    // the GPU that is used for scratch and small tensors
+    int32_t seed                            = -1;  // RNG seed
+    int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_predict                       = -1;  // new tokens to predict
+    int32_t n_ctx                           = 512; // context size
+    int32_t n_batch                         = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
+    int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
+    int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
+    bool    low_vram                        = 0;   // if true, reduce VRAM usage at the cost of performance
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens

examples/main/README.md

@@ -288,5 +288,6 @@ These options provide extra functionality and customization when running the LLa
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/server/README.md

@@ -289,6 +289,7 @@ Test();
 -   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
 -   `--embedding`: Enable the embedding mode. **Completion function doesn't work in this mode**.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`;
 -   `--port`: Set the port to listen. Default: `8080`.

examples/server/server.cpp

@@ -405,6 +405,7 @@ void server_print_usage(int /*argc*/, char **argv, const gpt_params &params)
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
   fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+  fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
   fprintf(stderr, "  -m FNAME, --model FNAME\n");
   fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -537,6 +538,14 @@ bool server_params_parse(int argc, char **argv, server_params &sparams, gpt_para
       }
 #else
       fprintf(stderr, "WARNING: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+    }
+    else if (arg == "--low-vram" || arg == "-lv")
+    {
+#ifdef GGML_USE_CUBLAS
+      params.low_vram = true;
+#else
+      fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
     }
     else if (arg == "--main-gpu" || arg == "-mg")

examples/train-text-from-scratch/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET train-text-from-scratch)
+add_executable(${TARGET} train-text-from-scratch.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/train-text-from-scratch/README.md

@@ -0,0 +1,22 @@
+# train-text-from-scratch
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://github.com/brunoklein99/deep-learning-notes/blob/master/shakespeare.txt
+
+# train
+./bin/train-text-from-scratch \
+        --vocab-model ../models/ggml-vocab.bin \
+        --ctx 64 --embd 256 --head 8 --layer 16 \
+        --checkpoint-in  chk-shakespeare-256x16.bin \
+        --checkpoint-out chk-shakespeare-256x16.bin \
+        --model-out ggml-shakespeare-256x16-f32.bin \
+        --train-data "shakespeare.txt" \
+        -t 6 -b 16 -n 32 --seed 1 --adam-iter 16 \
+        --print-details-interval 0 --predict 16 --use-flash
+
+# predict
+./bin/main -m ggml-shakespeare-256x16-f32.bin
+```

ggml-cuda.h

@@ -28,8 +28,10 @@ void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
 
 void   ggml_cuda_free_data(struct ggml_tensor * tensor);
 void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
+void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
 void   ggml_cuda_set_main_device(int main_device);
 void   ggml_cuda_set_scratch_size(size_t scratch_size);
+void   ggml_cuda_free_scratch(void);
 bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
 
 #ifdef  __cplusplus

ggml.h

@@ -296,6 +296,7 @@ extern "C" {
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
+        GGML_OP_REPEAT_BACK,
         GGML_OP_ABS,
         GGML_OP_SGN,
         GGML_OP_NEG,
@@ -309,6 +310,7 @@ extern "C" {
         GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
+        GGML_OP_OUT_PROD,
 
         GGML_OP_SCALE,
         GGML_OP_SET,
@@ -324,6 +326,7 @@ extern "C" {
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
+        GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
@@ -333,10 +336,14 @@ extern "C" {
 
         GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_BACK,
 
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_CROSS_ENTROPY_LOSS,
+        GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+
         GGML_OP_COUNT,
     };
 
@@ -478,6 +485,7 @@ extern "C" {
 
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
@@ -574,6 +582,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_acc(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -645,6 +658,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_repeat_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -698,14 +716,22 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // A: m rows, n columns
-    // B: p rows, n columns (i.e. we transpose it internally)
+    // A: n columns, m rows
+    // B: n columns, p rows  (i.e. we transpose it internally)
     // result is m columns, p rows
     GGML_API struct ggml_tensor * ggml_mul_mat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // A: m columns, n rows,
+    // B: p columns, n rows,
+    // result is m columns, p rows
+    GGML_API struct ggml_tensor * ggml_out_prod(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     //
     // operations on tensors without backpropagation
     //
@@ -916,6 +942,17 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_soft_max_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
@@ -982,6 +1019,14 @@ extern "C" {
             struct ggml_tensor  * v,
             bool                  masked);
 
+    GGML_API struct ggml_tensor * ggml_flash_attn_back(
+           struct ggml_context * ctx,
+           struct ggml_tensor  * q,
+           struct ggml_tensor  * k,
+           struct ggml_tensor  * v,
+           struct ggml_tensor  * d,
+           bool                  masked);
+
     GGML_API struct ggml_tensor * ggml_flash_ff(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1005,6 +1050,19 @@ extern "C" {
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    // loss function
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b);
+
+    GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+            struct ggml_tensor          * c);
+
     //
     // automatic differentiation
     //
@@ -1099,6 +1157,8 @@ extern "C" {
         struct {
             int n_iter;
 
+            float sched; // schedule multiplier (fixed, decay or warmup)
+            float decay; // weight decay for AdamW, use 0.0f to disable
             float alpha; // learning rate
             float beta1;
             float beta2;
@@ -1123,6 +1183,49 @@ extern "C" {
         } lbfgs;
     };
 
+    struct ggml_opt_context {
+        struct ggml_context * ctx;
+        struct ggml_opt_params params;
+
+        int iter;
+        int64_t nx; // number of parameter elements
+
+        bool just_initialized;
+
+        struct {
+            struct ggml_tensor * x;  // view of the parameters
+            struct ggml_tensor * g1; // gradient
+            struct ggml_tensor * g2; // gradient squared
+            struct ggml_tensor * m;  // first moment
+            struct ggml_tensor * v;  // second moment
+            struct ggml_tensor * mh; // first moment hat
+            struct ggml_tensor * vh; // second moment hat
+            struct ggml_tensor * pf; // past function values
+            float fx_best;
+            float fx_prev;
+            int n_no_improvement;
+        } adam;
+
+        struct {
+            struct ggml_tensor * x;    // current parameters
+            struct ggml_tensor * xp;   // previous parameters
+            struct ggml_tensor * g;    // current gradient
+            struct ggml_tensor * gp;   // previous gradient
+            struct ggml_tensor * d;    // search direction
+            struct ggml_tensor * pf;   // past function values
+            struct ggml_tensor * lmal; // the L-BFGS memory alpha
+            struct ggml_tensor * lmys; // the L-BFGS memory ys
+            struct ggml_tensor * lms;  // the L-BFGS memory s
+            struct ggml_tensor * lmy;  // the L-BFGS memory y
+            float fx_best;
+            float step;
+            int j;
+            int k;
+            int end;
+            int n_no_improvement;
+        } lbfgs;
+    };
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -1131,6 +1234,27 @@ extern "C" {
             struct ggml_opt_params params,
             struct ggml_tensor * f);
 
+    // initialize optimizer context
+    GGML_API void ggml_opt_init(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_opt_params params,
+            int64_t nx);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f);
+
+    // continue optimizing the function defined by the tensor f
+    GGML_API enum ggml_opt_result ggml_opt_resume_g(
+            struct ggml_context * ctx,
+            struct ggml_opt_context * opt,
+            struct ggml_tensor * f,
+            struct ggml_cgraph * gf,
+            struct ggml_cgraph * gb);
+
     //
     // quantization
     //

llama.cpp

@@ -165,6 +165,11 @@ struct llama_kv_cache {
         if (ctx) {
             ggml_free(ctx);
         }
+
+#ifdef GGML_USE_CUBLAS
+        ggml_cuda_free_data(k);
+        ggml_cuda_free_data(v);
+#endif // GGML_USE_CUBLAS
     }
 };
 
@@ -210,6 +215,7 @@ struct llama_model {
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cuda_free_data(tensors_by_name[i].second);
         }
+        ggml_cuda_free_scratch();
 #elif defined(GGML_USE_CLBLAST)
         for (size_t i = 0; i < tensors_by_name.size(); ++i) {
             ggml_cl_free_data(tensors_by_name[i].second);
@@ -867,7 +873,8 @@ static bool kv_cache_init(
         const struct llama_hparams & hparams,
              struct llama_kv_cache & cache,
                          ggml_type   wtype,
-                               int   n_ctx) {
+                               int   n_ctx,
+                               int   n_gpu_layers) {
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
@@ -893,6 +900,15 @@ static bool kv_cache_init(
     ggml_set_name(cache.k, "cache_k");
     ggml_set_name(cache.v, "cache_v");
 
+#ifdef GGML_USE_CUBLAS
+    if (n_gpu_layers > n_layer + 1) {
+        ggml_cuda_assign_buffers_no_scratch(cache.v);
+    }
+    if (n_gpu_layers > n_layer + 2) {
+        ggml_cuda_assign_buffers_no_scratch(cache.k);
+    }
+#endif // GGML_USE_CUBLAS
+
     return true;
 }
 
@@ -903,6 +919,7 @@ struct llama_context_params llama_context_default_params() {
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ {0},
+        /*.low_vram                    =*/ false,
         /*.seed                        =*/ -1,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
@@ -1011,6 +1028,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1036,6 +1054,12 @@ static void llama_model_load_internal(
             case 40: model.type = e_model::MODEL_13B; break;
             case 60: model.type = e_model::MODEL_30B; break;
             case 80: model.type = e_model::MODEL_65B; break;
+            default:
+                {
+                    if (hparams.n_layer < 32) {
+                        model.type = e_model::MODEL_7B;
+                    }
+                } break;
         }
 
         hparams.n_ctx = n_ctx;
@@ -1131,18 +1155,34 @@ static void llama_model_load_internal(
         ml->ggml_ctx = ctx;
 
         model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
-        model.norm           = ml->get_tensor("norm.weight",           {n_embd},          GGML_BACKEND_CPU);
 
         // "output" tensor
         {
+            ggml_backend backend_norm;
             ggml_backend backend_output;
             if (n_gpu_layers > int(n_layer)) { // NOLINT
+                // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
+                // on Windows however this is detrimental unless everything is on the GPU
+#ifndef _WIN32
+                backend_norm = low_vram ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#else
+                backend_norm = low_vram || n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+#endif // _WIN32
+
                 backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT;
             } else {
+                backend_norm = GGML_BACKEND_CPU;
                 backend_output = GGML_BACKEND_CPU;
             }
 
+            model.norm   = ml->get_tensor("norm.weight",   {n_embd},          backend_norm);
             model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
+            if (backend_norm == GGML_BACKEND_GPU) {
+                vram_weights += ggml_nbytes(model.norm);
+            }
+            if (backend_output == GGML_BACKEND_GPU_SPLIT) {
+                vram_weights += ggml_nbytes(model.output);
+            }
         }
 
         const int i_gpu_start = n_layer - n_gpu_layers;
@@ -1200,23 +1240,49 @@ static void llama_model_load_internal(
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
 
         (void) vram_scratch;
+        (void) n_batch;
 #ifdef GGML_USE_CUBLAS
-        vram_scratch = n_batch * MB;
-        ggml_cuda_set_scratch_size(vram_scratch);
-        if (n_gpu_layers > 0) {
-            fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
-                    __func__, vram_scratch / MB);
+        if (low_vram) {
+            fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
+            ggml_cuda_set_scratch_size(0); // disable scratch
+        } else {
+            vram_scratch = n_batch * MB;
+            ggml_cuda_set_scratch_size(vram_scratch);
+            if (n_gpu_layers > 0) {
+                fprintf(stderr, "%s: allocating batch_size x 1 MB = %ld MB VRAM for the scratch buffer\n",
+                        __func__, vram_scratch / MB);
+            }
         }
 #endif // GGML_USE_CUBLAS
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
         const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
 
-        fprintf(stderr, "%s: offloading %d layers to GPU\n", __func__, n_gpu);
+        fprintf(stderr, "%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
         if (n_gpu_layers > (int) hparams.n_layer) {
-            fprintf(stderr, "%s: offloading output layer to GPU\n", __func__);
+            fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__);
         }
+        size_t vram_kv_cache = 0;
+        if (n_gpu_layers > (int) hparams.n_layer + 1) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        if (n_gpu_layers > (int) hparams.n_layer + 2) {
+            if (low_vram) {
+                fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
+            } else {
+                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+            }
+        }
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+        fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n",
+                __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3);
         fprintf(stderr, "%s: total VRAM used: %zu MB\n",
-                __func__, (vram_weights + vram_scratch + MB - 1) / MB); // round up
+                __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up
 #else
         (void) n_gpu_layers;
 #endif
@@ -1227,6 +1293,7 @@ static void llama_model_load_internal(
         model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
     }
 
+    (void) tensor_split;
 #if defined(GGML_USE_CUBLAS)
     {
         ggml_cuda_set_tensor_split(tensor_split);
@@ -1254,6 +1321,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         float * tensor_split,
+        bool low_vram,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1261,7 +1329,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, memory_type,
+        llama_model_load_internal(fname, lctx, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1337,12 +1405,33 @@ static bool llama_eval_internal(
     const int i_gpu_start = n_layer - n_gpu_layers;
     (void) i_gpu_start;
 
+    // offload functions set the tensor output backend to GPU
+    // tensors are GPU-accelerated if any input or the output has been offloaded
+    //
+    // with the low VRAM option VRAM scratch is disabled in llama_load_model_internal
+    // in that case ggml_cuda_assign_buffers has no effect
+    offload_func_t offload_func_nr = llama_nop; // nr = non-repeating
+    offload_func_t offload_func_kq = llama_nop;
+    offload_func_t offload_func_v  = llama_nop;
+
+#ifdef GGML_USE_CUBLAS
+        if (n_gpu_layers > n_layer) {
+            offload_func_nr = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 1) {
+            offload_func_v  = ggml_cuda_assign_buffers;
+        }
+        if (n_gpu_layers > n_layer + 2) {
+            offload_func_kq = ggml_cuda_assign_buffers;
+        }
+#endif // GGML_USE_CUBLAS
+
     for (int il = 0; il < n_layer; ++il) {
         offload_func_t offload_func = llama_nop;
 
 #ifdef GGML_USE_CUBLAS
         if (il >= i_gpu_start) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
+            offload_func = ggml_cuda_assign_buffers;
         }
 #endif // GGML_USE_CUBLAS
 
@@ -1365,31 +1454,42 @@ static bool llama_eval_internal(
         // self-attention
         {
             // compute Q and K and RoPE them
-            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            // offload_func(tmpq);
-            ggml_set_name(tmpq, "tmpq");
-
             struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            // offload_func(tmpk);
+            offload_func_kq(tmpk);
             ggml_set_name(tmpk, "tmpk");
 
+            struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            offload_func_kq(tmpq);
+            ggml_set_name(tmpq, "tmpq");
+
             struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Kcur);
             ggml_set_name(Kcur, "Kcur");
 
             struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
             // store key and value to memory
             {
                 // compute the transposed [N, n_embd] V matrix
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
+
+                struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                offload_func_v(tmpv);
+                ggml_set_name(tmpv, "tmpv");
+
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N));
+                offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
                 struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                offload_func_kq(k);
                 ggml_set_name(k, "k");
+
                 struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
                         (   n_ctx)*ggml_element_size(kv_self.v),
                         (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                offload_func_v(v);
                 ggml_set_name(v, "v");
 
                 // important: storing RoPE-ed version of K in the KV cache!
@@ -1401,6 +1501,7 @@ static bool llama_eval_internal(
                 ggml_permute(ctx0,
                         Qcur,
                         0, 2, 1, 3);
+            offload_func_kq(Q);
             ggml_set_name(Q, "Q");
 
             struct ggml_tensor * K =
@@ -1409,10 +1510,12 @@ static bool llama_eval_internal(
                             ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
                             n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
+            offload_func_kq(K);
             ggml_set_name(K, "K");
 
             // K * Q
             struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
@@ -1421,14 +1524,17 @@ static bool llama_eval_internal(
 
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
+            offload_func_kq(KQ_scaled);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
             struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+            offload_func_kq(KQ_masked);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+            offload_func_v(KQ_soft_max);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
             // split cached V into n_head heads
@@ -1438,10 +1544,12 @@ static bool llama_eval_internal(
                         n_ctx*ggml_element_size(kv_self.v),
                         n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
                         il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
+            offload_func_v(V);
             ggml_set_name(V, "V");
 
 #if 1
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+            offload_func_v(KQV);
             ggml_set_name(KQV, "KQV");
 #else
             // make V contiguous in memory to speed up the matmul, however we waste time on the copy
@@ -1453,12 +1561,14 @@ static bool llama_eval_internal(
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            offload_func_v(KQV_merged);
             ggml_set_name(KQV_merged, "KQV_merged");
 
             // cur = KQV_merged.contiguous().view(n_embd, N)
             cur = ggml_cpy(ctx0,
                     KQV_merged,
                     ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+            offload_func_v(cur);
             ggml_set_name(cur, "KQV_merged_contiguous");
 
             // projection (no bias)
@@ -1470,7 +1580,6 @@ static bool llama_eval_internal(
         }
 
         lctx.use_buf(ctx0, 1);
-        //ggml_cuda_set_scratch(1);
 
         struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
         offload_func(inpFF);
@@ -1528,32 +1637,24 @@ static bool llama_eval_internal(
     }
 
     lctx.use_buf(ctx0, 0);
-    //ggml_cuda_set_scratch(0);
 
     // used at the end to optionally extract the embeddings
     struct ggml_tensor * embeddings = NULL;
 
-    offload_func_t offload_func = llama_nop;
-
-#ifdef GGML_USE_CUBLAS
-        if (n_gpu_layers > n_layer) {
-            offload_func = ggml_cuda_assign_buffers; // sets the output backend to GPU
-        }
-#endif // GGML_USE_CUBLAS
 
     // norm
     {
         cur = ggml_rms_norm(ctx0, inpL);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_inpL");
 
         cur = ggml_rms_norm(ctx0, cur);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "rms_norm_after");
 
         // cur = cur*norm(broadcasted)
         cur = ggml_mul(ctx0, cur, model.norm);
-        offload_func(cur);
+        offload_func_nr(cur);
         ggml_set_name(cur, "result_norm");
 
         embeddings = cur;
@@ -2161,6 +2262,10 @@ llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_tok
         return -log2f(candidate.p) > *mu;
     }));
 
+    if (candidates->size == 0) {
+        candidates->size = 1;
+    }
+
     // Normalize the probabilities of the remaining words
     llama_sample_softmax(ctx, candidates);
 
@@ -2540,8 +2645,8 @@ struct llama_context * llama_init_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, memory_type, params.use_mmap, params.use_mlock,
+    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_batch, params.n_gpu_layers, params.main_gpu,
+                params.tensor_split, params.low_vram, memory_type, params.use_mmap, params.use_mlock,
                 params.vocab_only, params.progress_callback, params.progress_callback_user_data)) {
         fprintf(stderr, "%s: failed to load model\n", __func__);
         llama_free(ctx);
@@ -2550,7 +2655,7 @@ struct llama_context * llama_init_from_file(
 
     // reserve memory for context buffers
     if (!params.vocab_only) {
-        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
+        if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx, params.n_gpu_layers)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
             return nullptr;
@@ -3287,6 +3392,19 @@ int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
+int llama_get_vocab(
+        const struct llama_context * ctx,
+        const char * * strings,
+        float  * scores,
+        int capacity) {
+    int n = std::min(capacity, (int) ctx->vocab.id_to_token.size());
+    for (int i = 0; i<n; ++i) {
+        strings[i] = ctx->vocab.id_to_token[i].tok.c_str();
+        scores[i]  = ctx->vocab.id_to_token[i].score;
+    }
+    return n;
+}
+
 float * llama_get_logits(struct llama_context * ctx) {
     return ctx->logits.data();
 }

llama.h

@@ -77,6 +77,7 @@ extern "C" {
         int n_gpu_layers;                      // number of layers to store in VRAM
         int main_gpu;                          // the GPU that is used for scratch and small tensors
         float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        bool low_vram;                         // if true, reduce VRAM usage at the cost of performance
         int seed;                              // RNG seed, -1 for random
 
         bool f16_kv;     // use fp16 for KV cache
@@ -220,6 +221,14 @@ extern "C" {
     LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
     LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
+    // Get the vocabulary as output parameters.
+    // Returns number of results.
+    LLAMA_API int llama_get_vocab(
+            const struct llama_context * ctx,
+                          const char * * strings,
+                                 float * scores,
+                                   int   capacity);
+
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
     // Can be mutated in order to change the probabilities of the next token

tests/test-grad0.c

@@ -5,7 +5,7 @@
 #include <stdlib.h>
 #include <assert.h>
 
-#define MAX_NARGS 2
+#define MAX_NARGS 3
 
 #undef MIN
 #undef MAX
@@ -1090,6 +1090,25 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // cross_entropy_loss
+        {
+            const int nargs = 1;
+
+            int64_t ne2[4];
+            get_random_dims(ne2, 4);
+
+            for (int ndims = 1; ndims <= 3; ++ndims) {
+                x[0] = get_random_tensor(ctx0, ndims, ne2, -1.0f, 1.0f);
+                x[1] = get_random_tensor(ctx0, ndims, ne2, 0.0f, 1.0f);
+                ggml_set_param(ctx0, x[0]);
+
+                struct ggml_tensor * f = ggml_sum(ctx0, ggml_cross_entropy_loss(ctx0, x[0], x[1]));
+
+                check_gradient("cross_entropy_loss", ctx0, x, f, ndims, nargs, 1e-1f, 1e-2f, INFINITY);
+                // finite differences regularly fails!
+            }
+        }
+
         // rope
         {
             const int nargs = 1;
@@ -1124,6 +1143,45 @@ int main(int argc, const char ** argv) {
             }
         }
 
+        // flash_attn
+        {
+            const int nargs = 3;
+
+            int64_t ne2[4];
+
+            get_random_dims(ne2, 4);
+            int64_t D = ne2[0];
+            int64_t N = ne2[1];
+            int64_t M = ne2[2] + N;
+            int64_t B = ne2[3];
+
+            for (int masked = 0; masked <= 1; ++masked) {
+                for (int ndims = 2; ndims <= 4; ++ndims) {
+                    int64_t neq[4] = { D, N, B, ne[3] };
+                    int64_t nek[4] = { D, M, B, ne[3] };
+                    int64_t nev[4] = { M, D, B, ne[3] };
+                    if (ndims == 2) {
+                        neq[2] = 1; neq[3] = 1;
+                        nek[2] = 1; nek[3] = 1;
+                        nev[2] = 1; nev[3] = 1;
+                    } else if (ndims == 3) {
+                        neq[3] = 1;
+                        nek[3] = 1;
+                        nev[3] = 1;
+                    }
+                    x[0] = get_random_tensor(ctx0, ndims, neq, -0.1250f, 0.1250f);
+                    x[1] = get_random_tensor(ctx0, ndims, nek, -0.1250f, 0.1250f);
+                    x[2] = get_random_tensor(ctx0, ndims, nev, -0.1250f, 0.1250f);
+                    ggml_set_param(ctx0, x[0]);
+                    ggml_set_param(ctx0, x[1]);
+                    ggml_set_param(ctx0, x[2]);
+
+                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+
+                    check_gradient("flash_attn", ctx0, x, f, ndims, nargs, 1.5e-4f, INFINITY, 3.5f);
+                }
+            }
+        }
         ggml_free(ctx0);
     }