Merge remote-tracking branch 'upstream/concedo'

CMakeLists.txt

@@ -46,7 +46,7 @@ set(LLAMA_CUDA_DMMV_X "32" CACHE STRING "llama: x stride for dmmv CUDA kern
 set(LLAMA_CUDA_DMMV_Y "1" CACHE STRING "llama: y block size for dmmv CUDA kernels")
 option(LLAMA_CUDA_DMMV_F16 "llama: use 16 bit floats for dmmv CUDA kernels" OFF)
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
-option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF)
+option(LLAMA_HIPBLAS "llama: use hipBLAS" ON)
 option(LLAMA_K_QUANTS "llama: use k-quants" ON)
 
 
@@ -339,4 +339,3 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME "koboldcpp_cublas")
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 target_link_libraries(${TARGET} PUBLIC ggml ggml_v1 ggml_v2 common2 gpttype_adapter ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-

Makefile

@@ -42,7 +42,7 @@ endif
 
 # keep standard at C11 and C++11
 CFLAGS = -I. -I./include -I./include/CL -I./otherarch -I./otherarch/tools -Ofast -DNDEBUG -std=c11 -fPIC -DGGML_USE_K_QUANTS
-CXXFLAGS = -I. -I./examples -I./include -I./include/CL -I./otherarch -I./otherarch/tools -O3 -DNDEBUG -std=c++11 -fPIC
+CXXFLAGS = -I. -I./examples -I./include -I./include/CL -I./otherarch -I./otherarch/tools -O3 -DNDEBUG -std=c++11 -fPIC -DGGML_USE_K_QUANTS
 LDFLAGS =
 
 # these are used on windows, to build some libraries with extra old device compatibility
@@ -53,7 +53,11 @@ NONECFLAGS =
 OPENBLAS_FLAGS = -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 CLBLAST_FLAGS = -DGGML_USE_CLBLAST
 FAILSAFE_FLAGS = -DUSE_FAILSAFE
-CUBLAS_FLAGS = -DGGML_USE_CUBLAS
+ifdef LLAMA_CUBLAS
+ CUBLAS_FLAGS = -DGGML_USE_CUBLAS
+else
+ CUBLAS_FLAGS =
+endif
 CUBLASLD_FLAGS =
 CUBLAS_OBJS =
 

examples/common.cpp

@@ -110,7 +110,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
  invalid_param = true;
  break;
  }
- params.seed = std::stoi(argv[i]);
+ params.seed = std::stoul(argv[i]);
  } else if (arg == "-t" || arg == "--threads") {
  if (++i >= argc) {
  invalid_param = true;

examples/common.h

@@ -22,7 +22,7 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
- int32_t seed  = -1; // RNG seed
+ uint32_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

examples/embedding/embedding.cpp

@@ -24,11 +24,11 @@ int main(int argc, char ** argv) {
 
  fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
- if (params.seed < 0) {
+ if (params.seed == LLAMA_DEFAULT_SEED) {
  params.seed = time(NULL);
  }
 
- fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+ fprintf(stderr, "%s: seed = %u\n", __func__, params.seed);
 
  std::mt19937 rng(params.seed);
  if (params.random_prompt) {

examples/main/README.md

@@ -242,7 +242,7 @@ Example usage: `--logit-bias 29905-inf`
 
 ### RNG Seed
 
-- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+- `-s SEED, --seed SEED`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
 The RNG seed is used to initialize the random number generator that influences the text generation process. By setting a specific seed value, you can obtain consistent and reproducible results across multiple runs with the same input and settings. This can be helpful for testing, debugging, or comparing the effects of different options on the generated text to see when they diverge. If the seed is set to a value less than 0, a random seed will be used, which will result in different outputs on each run.
 

examples/main/main.cpp

@@ -94,11 +94,11 @@ int main(int argc, char ** argv) {
 
  fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
- if (params.seed < 0) {
+ if (params.seed == LLAMA_DEFAULT_SEED) {
  params.seed = time(NULL);
  }
 
- fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+ fprintf(stderr, "%s: seed = %u\n", __func__, params.seed);
 
  std::mt19937 rng(params.seed);
  if (params.random_prompt) {

examples/perplexity/perplexity.cpp

@@ -136,11 +136,11 @@ int main(int argc, char ** argv) {
 
  fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
 
- if (params.seed < 0) {
+ if (params.seed == LLAMA_DEFAULT_SEED) {
  params.seed = time(NULL);
  }
 
- fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+ fprintf(stderr, "%s: seed = %u\n", __func__, params.seed);
 
  std::mt19937 rng(params.seed);
  if (params.random_prompt) {

examples/server/README.md

@@ -152,7 +152,7 @@ node .
 
  `mirostat_eta`: Set the Mirostat learning rate, parameter eta (default: 0.1).
 
- `seed`: Set the random number generator (RNG) seed (default: -1, < 0 = random seed).
+ `seed`: Set the random number generator (RNG) seed (default: -1, -1 = random seed).
 
  `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 

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

@@ -2768,7 +2768,7 @@ void train_print_usage(int /*argc*/, char ** argv, const struct train_params * p
  fprintf(stderr, " --checkpoint-in FNAME path from which to load training checkpoint (default '%s')\n", params->fn_checkpoint_in);
  fprintf(stderr, " --checkpoint-out FNAME path to save training checkpoint (default '%s')\n", params->fn_checkpoint_out);
  fprintf(stderr, " --model-out FNAME path to save ggml model (default '%s')\n", params->fn_model_out);
- fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)\n");
+ fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for -1)\n");
  fprintf(stderr, " -c N, --ctx N Context size used during training (default %d)\n", params->n_ctx);
  fprintf(stderr, " --embd N Embedding size used for new models (default %d)\n", params->n_embd);
  fprintf(stderr, " --mult N Mult size used for new models, influences feedforward size. (default %d)\n", params->n_mult);
@@ -3034,10 +3034,10 @@ int main(int argc, char ** argv) {
  return 1;
  }
 
- if (params.seed < 0) {
+ if (params.seed == LLAMA_DEFAULT_SEED) {
  params.seed = time(NULL);
  }
- printf("%s: seed: %d\n", __func__, params.seed);
+ printf("%s: seed: %u\n", __func__, params.seed);
  srand(params.seed);
 
  struct llama_context_params llama_params = llama_context_default_params();

koboldcpp.py

@@ -236,7 +236,7 @@ def utfprint(str):
 maxhordelen = 256
 modelbusy = False
 defaultport = 5001
-KcppVersion = "1.33"
+KcppVersion = "1.34"
 showdebug = True
 
 class ServerRequestHandler(http.server.SimpleHTTPRequestHandler):

llama.cpp

@@ -12,7 +12,8 @@
 #include "ggml.h"
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#elif defined(GGML_USE_CLBLAST)
+#endif
+#if defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
 #endif
 
@@ -66,6 +67,7 @@ enum e_model {
  MODEL_65B,
 };
 
+static const size_t kB = 1024;
 static const size_t MB = 1024*1024;
 
 // computed for n_ctx == 2048
@@ -129,6 +131,34 @@ static const std::map<e_model, size_t> & MEM_REQ_EVAL()
  return k_sizes;
 }
 
+// amount of VRAM needed per batch size to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
+{
+ static std::map<e_model, size_t> k_sizes = {
+ { MODEL_3B, 512ull * kB },
+ { MODEL_7B, 512ull * kB },
+ { MODEL_13B, 640ull * kB },
+ { MODEL_30B, 768ull * kB },
+ { MODEL_65B, 1536ull * kB },
+ };
+ return k_sizes;
+}
+
+// amount of VRAM needed per batch size and context to hold temporary results
+// the values for 3b and 65b are not derived from testing but instead chosen conservatively
+static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
+{
+ static std::map<e_model, size_t> k_sizes = {
+ { MODEL_3B, 128ull },
+ { MODEL_7B, 128ull },
+ { MODEL_13B, 160ull },
+ { MODEL_30B, 208ull },
+ { MODEL_65B, 416ull },
+ };
+ return k_sizes;
+}
+
 // default hparams (LLaMA 7B)
 struct llama_hparams {
  uint32_t n_vocab = 32000;
@@ -777,7 +807,7 @@ static bool kv_cache_init(
 
 struct llama_context_params llama_context_default_params() {
  struct llama_context_params result = {
- /*.seed =*/ -1,
+ /*.seed =*/ LLAMA_DEFAULT_SEED,
  /*.n_ctx =*/ 512,
  /*.n_batch =*/ 512,
  /*.gpu_layers =*/ 0,
@@ -1113,11 +1143,14 @@ static void llama_model_load_internal(
  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;
+ const size_t vram_scratch_base = VRAM_REQ_SCRATCH_BASE().at(model.type);
+ const size_t vram_scratch_per_context = VRAM_REQ_SCRATCH_PER_CONTEXT().at(model.type);
+ vram_scratch = n_batch * (vram_scratch_base + n_ctx * vram_scratch_per_context);
  ggml_cuda_set_scratch_size(vram_scratch);
  if (n_gpu_layers > 0) {
- fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",
- __func__, vram_scratch / MB);
+ fprintf(stderr, "%s: allocating batch_size x (%zd kB + n_ctx x %zd B) = %zd MB VRAM for the scratch buffer\n",
+ __func__, vram_scratch_base / kB, vram_scratch_per_context,
+ (vram_scratch + MB - 1) / MB); // round up
  }
  }
 #endif // GGML_USE_CUBLAS
@@ -2540,7 +2573,7 @@ struct llama_context * llama_new_context_with_model(
 
  llama_context * ctx = new llama_context(*model, model->vocab);
 
- if (params.seed < 0) {
+ if (params.seed == LLAMA_DEFAULT_SEED) {
  params.seed = time(NULL);
  }
 
@@ -2974,8 +3007,8 @@ int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
 
 #define LLAMA_MAX_RNG_STATE (64*1024)
 
-void llama_set_rng_seed(struct llama_context * ctx, int seed) {
- if (seed < 0) {
+void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed) {
+ if (seed == LLAMA_DEFAULT_SEED) {
  seed = time(NULL);
  }
  ctx->rng.seed(seed);

llama.h

@@ -46,6 +46,8 @@
 #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION 1
 
+#define LLAMA_DEFAULT_SEED 0xFFFFFFFF
+
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
@@ -81,11 +83,11 @@ extern "C" {
  typedef void (*llama_progress_callback)(float progress, void *ctx);
 
  struct llama_context_params {
- int seed;  // RNG seed, -1 for random
- int n_ctx;  // text context
- int n_batch;  // prompt processing batch size
- int n_gpu_layers;  // number of layers to store in VRAM
- int main_gpu;  // the GPU that is used for scratch and small tensors
+ uint32_t seed; // RNG seed, -1 for random
+ int32_t n_ctx; // text context
+ int32_t n_batch; // prompt processing batch size
+ int32_t n_gpu_layers; // number of layers to store in VRAM
+ int32_t 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
  // called with a progress value between 0 and 1, pass NULL to disable
  llama_progress_callback progress_callback;
@@ -196,7 +198,7 @@ extern "C" {
  LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
 
  // Sets the current rng seed.
- LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
+ LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
 
  // Returns the maximum size in bytes of the state (rng, logits, embedding
  // and kv_cache) - will often be smaller after compacting tokens

otherarch/gpt2_v3.cpp

@@ -18,10 +18,12 @@
 
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#elif defined(GGML_USE_CLBLAST)
+#endif
+#if defined(GGML_USE_CLBLAST)
 #include "ggml-opencl.h"
 #endif
 
+
 // load the model's weights from a file
 ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, FileFormat file_format, int gpulayers) {
  printf("%s: loading model from '%s'\n", __func__, fname.c_str());

otherarch/llama_v2.cpp

@@ -9,13 +9,15 @@
 #include "llama_v2.h"
 
 #include "ggml_v2.h"
+
 #ifdef GGML_USE_CUBLAS
 #include "ggml_v2-cuda.h"
 #endif
 #if defined(GGML_USE_CLBLAST)
 #include "ggml_v2-opencl.h"
 #endif
 
+
 #include <array>
 #include <ctime>
 #include <cinttypes>