llama.cpp : split llama_context_params into model and context params

common/common.cpp

@@ -129,6 +129,15 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads <= 0) {
                 params.n_threads = std::thread::hardware_concurrency();
             }
+        } else if (arg == "-tb" || arg == "--threads-batch") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_threads_batch = std::stoi(argv[i]);
+            if (params.n_threads_batch <= 0) {
+                params.n_threads_batch = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -425,12 +434,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.mul_mat_q = false;
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\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;
@@ -606,7 +609,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        (can be specified more than once for multiple prompts).\n");
     printf("  --color               colorise output to distinguish prompt and user input from generations\n");
     printf("  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    printf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
+    printf("  -tb N, --threads-batch N\n");
+    printf("                        number of threads to use during batch evaluation and prompt processing (default: same as --threads)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
     printf("                        prompt to start generation with (default: empty)\n");
     printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
@@ -621,7 +626,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -f FNAME, --file FNAME\n");
     printf("                        prompt file to start generation.\n");
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
-    printf("  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
@@ -678,7 +683,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -ts SPLIT --tensor-split SPLIT\n");
     printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    printf("  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
 #ifdef GGML_USE_CUBLAS
     printf("  -nommq, --no-mul-mat-q\n");
     printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
@@ -722,40 +726,50 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
 // Model utils
 //
 
-struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
-    auto lparams = llama_context_default_params();
+struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params) {
+    auto mparams = llama_model_default_params();
 
-    lparams.n_ctx           = params.n_ctx;
-    lparams.n_batch         = params.n_batch;
     if (params.n_gpu_layers != -1) {
-        lparams.n_gpu_layers = params.n_gpu_layers;
+        mparams.n_gpu_layers = params.n_gpu_layers;
     }
-    lparams.main_gpu        = params.main_gpu;
-    lparams.tensor_split    = params.tensor_split;
-    lparams.low_vram        = params.low_vram;
-    lparams.mul_mat_q       = params.mul_mat_q;
-    lparams.seed            = params.seed;
-    lparams.f16_kv          = params.memory_f16;
-    lparams.use_mmap        = params.use_mmap;
-    lparams.use_mlock       = params.use_mlock;
-    lparams.logits_all      = params.perplexity;
-    lparams.embedding       = params.embedding;
-    lparams.rope_freq_base  = params.rope_freq_base;
-    lparams.rope_freq_scale = params.rope_freq_scale;
-
-    return lparams;
+    mparams.main_gpu        = params.main_gpu;
+    mparams.tensor_split    = params.tensor_split;
+    mparams.use_mmap        = params.use_mmap;
+    mparams.use_mlock       = params.use_mlock;
+
+    return mparams;
+}
+
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
+    auto cparams = llama_context_default_params();
+
+    cparams.n_ctx           = params.n_ctx;
+    cparams.n_batch         = params.n_batch;
+    cparams.n_threads       = params.n_threads;
+    cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q       = params.mul_mat_q;
+    cparams.seed            = params.seed;
+    cparams.f16_kv          = params.memory_f16;
+    cparams.logits_all      = params.perplexity;
+    cparams.embedding       = params.embedding;
+    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale;
+
+    return cparams;
 }
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
-    auto lparams = llama_context_params_from_gpt_params(params);
+    auto mparams = llama_model_params_from_gpt_params(params);
 
-    llama_model * model  = llama_load_model_from_file(params.model.c_str(), lparams);
+    llama_model * model  = llama_load_model_from_file(params.model.c_str(), mparams);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return std::make_tuple(nullptr, nullptr);
     }
 
-    llama_context * lctx = llama_new_context_with_model(model, lparams);
+    auto cparams = llama_context_params_from_gpt_params(params);
+
+    llama_context * lctx = llama_new_context_with_model(model, cparams);
     if (lctx == NULL) {
         fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
         llama_free_model(model);
@@ -783,7 +797,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         LOG("warming up the model with an empty run\n");
 
         const std::vector<llama_token> tmp = { llama_token_bos(lctx), llama_token_eos(lctx), };
-        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, params.n_threads);
+        llama_eval(lctx, tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0);
         llama_reset_timings(lctx);
     }
 
@@ -1213,7 +1227,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
-    fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
     fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);

common/common.h

@@ -36,6 +36,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
+    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
     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)
@@ -48,8 +49,8 @@ struct gpt_params {
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
     int32_t n_probs                         = 0;    // if greater than 0, output the probabilities of top n_probs tokens.
     int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
-    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
-    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
 
     // sampling parameters
     int32_t top_k             = 40;    // <= 0 to use vocab size
@@ -93,7 +94,6 @@ struct gpt_params {
     bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
@@ -131,6 +131,7 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 //
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
+struct llama_model_params   llama_model_params_from_gpt_params(const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
 //

examples/beam-search/beam-search.cpp

@@ -159,7 +159,7 @@ int main(int argc, char ** argv)
     std::cout << std::flush;
 
     int n_past = llama_get_kv_cache_token_count(ctx);
-    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past, params.n_threads))
+    if (llama_eval(ctx, tokens_list.data(), tokens_list.size(), n_past))
     {
         fprintf(stderr, "%s : failed to eval prompt.\n" , __func__ );
         return 1;
@@ -169,7 +169,7 @@ int main(int argc, char ** argv)
     beam_search_callback_data callback_data{ctx, {}};
     size_t const beam_width = static_cast<size_t>(params.n_beams);
     int const n_predict = 256;
-    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict, params.n_threads);
+    llama_beam_search(ctx, beam_search_callback, &callback_data, beam_width, n_past, n_predict);
 
     std::cout << "\n\n";
     for (llama_token const token_id : callback_data.response) {

examples/embd-input/embd-input-lib.cpp

@@ -80,7 +80,7 @@ bool eval_float(void * model, float * input, int N){
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past, params.n_threads)) {
+        if (llama_eval_embd(ctx, (input+i*n_emb), n_eval, n_past)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }
@@ -101,7 +101,7 @@ bool eval_tokens(void * model, std::vector<llama_token> tokens) {
         if (n_eval > params.n_batch) {
             n_eval = params.n_batch;
         }
-        if (llama_eval(ctx, &tokens[i], n_eval, n_past, params.n_threads)) {
+        if (llama_eval(ctx, &tokens[i], n_eval, n_past)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return false;
         }

examples/embedding/embedding.cpp

@@ -43,9 +43,11 @@ int main(int argc, char ** argv) {
     }
 
     const int n_ctx_train = llama_n_ctx_train(ctx);
-    if (params.n_ctx > n_ctx_train) {
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
         fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
+                __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
@@ -70,15 +72,15 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)params.n_ctx) {
+    if (embd_inp.size() > (size_t)n_ctx) {
         fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), params.n_ctx);
+                __func__, embd_inp.size(), n_ctx);
         return 1;
     }
 
     while (!embd_inp.empty()) {
         int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
+        if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past)) {
             fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }