llama : remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD

common/common.cpp

@@ -637,11 +637,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             const std::regex regex{R"([,/]+)"};
             std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
             std::vector<std::string> split_arg{it, {}};
-            if (split_arg.size() >= LLAMA_MAX_DEVICES) {
+            if (split_arg.size() >= llama_max_devices()) {
                 invalid_param = true;
                 break;
             }
-            for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+            for (size_t i = 0; i < llama_max_devices(); ++i) {
                 if (i < split_arg.size()) {
                     params.tensor_split[i] = std::stof(split_arg[i]);
                 } else {
@@ -1651,7 +1651,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
 
-    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
+    const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);

common/common.h

@@ -43,40 +43,40 @@ extern char const *LLAMA_BUILD_TARGET;
 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_draft                 = -1;
-    int32_t n_threads_batch                 = -1;    // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft           = -1;
-    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_draft                         = 8;     // number of tokens to draft during speculative decoding
-    int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
-    int32_t n_parallel                      = 1;     // number of parallel sequences to decode
-    int32_t n_sequences                     = 1;     // number of sequences to decode
-    float   p_accept                        = 0.5f;  // speculative decoding accept probability
-    float   p_split                         = 0.1f;  // speculative decoding split probability
-    int32_t n_gpu_layers                    = -1;    // number of layers to store in VRAM (-1 - use default)
-    int32_t n_gpu_layers_draft              = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
-    llama_split_mode split_mode             = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
-    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
-    int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
-    int32_t grp_attn_n                      = 1;     // group-attention factor
-    int32_t grp_attn_w                      = 512;   // group-attention width
-    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base                  = 0.0f;  // RoPE base frequency
-    float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
-    float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor                = 1.0f;  // YaRN magnitude scaling factor
-    float   yarn_beta_fast                  = 32.0f; // YaRN low correction dim
-    float   yarn_beta_slow                  = 1.0f;  // YaRN high correction dim
-    int32_t yarn_orig_ctx                   = 0;     // YaRN original context length
-    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
-                                                                              //       pinging @cebtenzzre
+    uint32_t seed                 = -1;    // RNG seed
+
+    int32_t n_threads             = get_num_physical_cores();
+    int32_t n_threads_draft       = -1;
+    int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_threads_batch_draft = -1;
+    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_draft               = 8;     // number of tokens to draft during speculative decoding
+    int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
+    int32_t n_parallel            = 1;     // number of parallel sequences to decode
+    int32_t n_sequences           = 1;     // number of sequences to decode
+    float   p_accept              = 0.5f;  // speculative decoding accept probability
+    float   p_split               = 0.1f;  // speculative decoding split probability
+    int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
+    int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
+    llama_split_mode split_mode   = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
+    int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
+    int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
+    int32_t grp_attn_n            = 1;     // group-attention factor
+    int32_t grp_attn_w            = 512;   // group-attention width
+    int32_t n_print               = -1;    // print token count every n tokens (-1 = disabled)
+    float   rope_freq_base        = 0.0f;  // RoPE base frequency
+    float   rope_freq_scale       = 0.0f;  // RoPE frequency scaling factor
+    float   yarn_ext_factor       = -1.0f; // YaRN extrapolation mix factor
+    float   yarn_attn_factor      = 1.0f;  // YaRN magnitude scaling factor
+    float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
+    float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
+    int32_t yarn_orig_ctx         = 0;     // YaRN original context length
+    int8_t  rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
+                                                                    //       pinging @cebtenzzre
 
     // // sampling parameters
     struct llama_sampling_params sparams;

examples/batched-bench/batched-bench.cpp

@@ -88,7 +88,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
+    const std::vector<float> t_split(llama_max_devices(), 0.0f);
 
     model_params.n_gpu_layers = n_gpu_layers;
     model_params.tensor_split = t_split.data();

examples/llama-bench/llama-bench.cpp

@@ -160,7 +160,7 @@ struct cmd_params {
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
     std::vector<bool> mul_mat_q;
-    std::vector<std::array<float, LLAMA_MAX_DEVICES>> tensor_split;
+    std::vector<std::vector<float>> tensor_split;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -179,7 +179,7 @@ static const cmd_params cmd_params_defaults = {
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
     /* mul_mat_q     */ {true},
-    /* tensor_split  */ {{}},
+    /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -380,10 +380,10 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 const std::regex regex{R"([;/]+)"};
                 std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
                 std::vector<std::string> split_arg{it, {}};
-                GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+                GGML_ASSERT(split_arg.size() <= llama_max_devices());
 
-                std::array<float, LLAMA_MAX_DEVICES> tensor_split;
-                for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
+                std::vector<float> tensor_split(llama_max_devices());
+                for (size_t i = 0; i < llama_max_devices(); ++i) {
                     if (i < split_arg.size()) {
                         tensor_split[i] = std::stof(split_arg[i]);
                     } else {
@@ -459,7 +459,7 @@ struct cmd_params_instance {
     int main_gpu;
     bool no_kv_offload;
     bool mul_mat_q;
-    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+    std::vector<float> tensor_split;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -582,7 +582,7 @@ struct test {
     int main_gpu;
     bool no_kv_offload;
     bool mul_mat_q;
-    std::array<float, LLAMA_MAX_DEVICES> tensor_split;
+    std::vector<float> tensor_split;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -704,7 +704,7 @@ struct test {
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
         int max_nonzero = 0;
-        for (int i = 0; i < LLAMA_MAX_DEVICES; i++) {
+        for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
                 max_nonzero = i;
             }

llama.cpp

@@ -10090,8 +10090,16 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
     return result;
 }
 
-int32_t llama_max_devices(void) {
-    return LLAMA_MAX_DEVICES;
+size_t llama_max_devices(void) {
+#if defined(GGML_USE_METAL)
+    return 1;
+#elif defined(GGML_USE_CUDA)
+    return GGML_CUDA_MAX_DEVICES;
+#elif defined(GGML_USE_SYCL)
+    return GGML_SYCL_MAX_DEVICES;
+#else
+    return 1;
+#endif
 }
 
 bool llama_mmap_supported(void) {

llama.h

@@ -5,13 +5,10 @@
 #include "ggml-backend.h"
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
-#define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES
 #elif defined(GGML_USE_SYCL)
 #include "ggml-sycl.h"
-#define LLAMA_MAX_DEVICES GGML_SYCL_MAX_DEVICES
-#else
-#define LLAMA_MAX_DEVICES 1
-#endif // GGML_USE_CUBLAS
+#endif
+
 #include <stddef.h>
 #include <stdint.h>
 #include <stdio.h>
@@ -50,7 +47,7 @@
 #define LLAMA_SESSION_VERSION 4
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_VULKAN) || \
-    defined(GGML_USE_SYCL) || defined(GGML_USE_KOMPUTE)
+    defined(GGML_USE_SYCL)   || defined(GGML_USE_KOMPUTE)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
 #define LLAMA_SUPPORTS_GPU_OFFLOAD
 #endif
@@ -201,7 +198,7 @@ extern "C" {
         // LLAMA_SPLIT_LAYER: ignored
         int32_t main_gpu;
 
-        // proportion of the model (layers or rows) to offload to each GPU, size: LLAMA_MAX_DEVICES
+        // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
@@ -338,7 +335,8 @@ extern "C" {
 
     LLAMA_API int64_t llama_time_us(void);
 
-    LLAMA_API int32_t  llama_max_devices(void);
+    LLAMA_API size_t llama_max_devices(void);
+
     LLAMA_API bool llama_mmap_supported (void);
     LLAMA_API bool llama_mlock_supported(void);