diff --git a/common/common.cpp b/common/common.cpp index f64da2cb66bb8..c5e83cc2a9e40 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -671,7 +671,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } else if (arg == "--no-mmap") { params.use_mmap = false; } else if (arg == "--numa") { - params.numa = true; + if (++i >= argc) { + invalid_param = true; + break; + } + std::string value(argv[i]); + /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } + else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } + else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } + else { invalid_param = true; break; } } else if (arg == "--verbose-prompt") { params.verbose_prompt = true; } else if (arg == "--no-display-prompt") { @@ -935,7 +943,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -tb N, --threads-batch N\n"); printf(" number of threads to use during batch and prompt processing (default: same as --threads)\n"); printf(" -td N, --threads-draft N"); - printf(" number of threads to use during generation (default: same as --threads)"); + printf(" number of threads to use during generation (default: same as --threads)\n"); printf(" -tbd N, --threads-batch-draft N\n"); printf(" number of threads to use during batch and prompt processing (default: same as --threads-draft)\n"); printf(" -p PROMPT, --prompt PROMPT\n"); @@ -1005,7 +1013,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n"); printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks); - printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base"); + printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base\n"); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); @@ -1022,7 +1030,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { if (llama_supports_mmap()) { printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); } - printf(" --numa attempt optimizations that help on some NUMA systems\n"); + printf(" --numa TYPE attempt optimizations that help on some NUMA systems\n"); + printf(" - distribute: spread execution evenly over all nodes\n"); + printf(" - isolate: only spawn threads on CPUs on the node that execution started on\n"); + printf(" - numactl: use the CPU map provided by numactl\n"); printf(" if run without this previously, it is recommended to drop the system page cache before using this\n"); printf(" see https://github.com/ggerganov/llama.cpp/issues/1437\n"); if (llama_supports_gpu_offload()) { @@ -1689,7 +1700,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false"); fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false"); fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false"); - fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false"); fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type); fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride); fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present); diff --git a/common/common.h b/common/common.h index 9bdd45cf9f84f..74c1369953d48 100644 --- a/common/common.h +++ b/common/common.h @@ -76,6 +76,7 @@ struct gpt_params { float yarn_beta_slow = 1.0f; // YaRN high correction dim int32_t yarn_orig_ctx = 0; // YaRN original context length int32_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED; + ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED; // // sampling parameters struct llama_sampling_params sparams; @@ -134,7 +135,6 @@ struct gpt_params { bool logits_all = false; // return logits for all tokens in the batch bool use_mmap = true; // use mmap for faster loads bool use_mlock = false; // use mlock to keep model in memory - bool numa = false; // attempt optimizations that help on some NUMA systems bool verbose_prompt = false; // print prompt tokens before generation bool display_prompt = true; // print prompt before generation bool infill = false; // use infill mode diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index b52d684578ceb..55dfd97843895 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -82,7 +82,8 @@ int main(int argc, char ** argv) { // init LLM - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); // initialize the model diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 4d000534900af..d75c503d58311 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu let n_len: Int = 32 // init LLM -llama_backend_init(false) +llama_backend_init() defer { llama_backend_free() } diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index b1775e0b0e8d6..eab636692e7d1 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -50,7 +50,8 @@ int main(int argc, char ** argv) { // init LLM - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); // initialize the model diff --git a/examples/beam-search/beam-search.cpp b/examples/beam-search/beam-search.cpp index 679b382e19b4e..866c6d7a62867 100644 --- a/examples/beam-search/beam-search.cpp +++ b/examples/beam-search/beam-search.cpp @@ -119,7 +119,8 @@ int main(int argc, char ** argv) // Init LLM : //--------------------------------- - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model; llama_context * ctx; diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index b4688cf519d15..acff715e99d05 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -74,7 +74,8 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model; llama_context * ctx; diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index bc9f6fa682f96..f21bc48f3b466 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -568,7 +568,8 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model_params mparams = llama_model_params_from_gpt_params(params); diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 72fb133b4fa06..92c67b7cff5c8 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -202,7 +202,8 @@ int main(int argc, char ** argv) { std::mt19937 rng(params.seed); LOG("%s: llama backend init\n", __func__); - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model; llama_context * ctx; diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index ddb0ba064b0eb..11410f8ae7625 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -1151,8 +1151,7 @@ int main(int argc, char ** argv) { if (!params.verbose) { llama_log_set(llama_null_log_callback, NULL); } - bool numa = false; - llama_backend_init(numa); + llama_backend_init(); // initialize printer std::unique_ptr p; diff --git a/examples/llama.android/app/src/main/cpp/llama-android.cpp b/examples/llama.android/app/src/main/cpp/llama-android.cpp index d5e705dce6ca0..2beb1e0d5321d 100644 --- a/examples/llama.android/app/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/app/src/main/cpp/llama-android.cpp @@ -274,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb extern "C" JNIEXPORT void JNICALL -Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) { - llama_backend_init(numa); +Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) { + llama_backend_init(); } extern "C" diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index fc79fd3466b54..58fcf40c6fb69 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -51,7 +51,7 @@ actor LlamaContext { } static func create_context(path: String) throws -> LlamaContext { - llama_backend_init(false) + llama_backend_init() var model_params = llama_model_default_params() #if targetEnvironment(simulator) diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index bef7f7c95a7e7..e29da6cb2f9b1 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -218,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) { auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); - llama_backend_init(params->numa); + llama_backend_init(); + llama_numa_init(params->numa); llama_model_params model_params = llama_model_params_from_gpt_params(*params); diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp index e55a15a1bf054..e2551e7a494c2 100644 --- a/examples/lookahead/lookahead.cpp +++ b/examples/lookahead/lookahead.cpp @@ -54,7 +54,8 @@ int main(int argc, char ** argv) { #endif // LOG_DISABLE_LOGS // init llama.cpp - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model = NULL; llama_context * ctx = NULL; diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index 18235b8a1d31d..b53fae11045b8 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -31,7 +31,8 @@ int main(int argc, char ** argv){ #endif // LOG_DISABLE_LOGS // init llama.cpp - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model = NULL; llama_context * ctx = NULL; diff --git a/examples/main/README.md b/examples/main/README.md index c7997f66569a5..7f84e42623274 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -283,7 +283,11 @@ These options help improve the performance and memory usage of the LLaMA models. ### NUMA support -- `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root. +- `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes. +- `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node. +- `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus. + + These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root. ### Memory Float 32 diff --git a/examples/main/main.cpp b/examples/main/main.cpp index e8ab8cbae0c92..f5d2f48935eb6 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -185,7 +185,8 @@ int main(int argc, char ** argv) { } LOG("%s: llama backend init\n", __func__); - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model; llama_context * ctx; diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index d2e074d9e12b0..7d11fcd593080 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -122,7 +122,8 @@ int main(int argc, char ** argv) { #endif // LOG_DISABLE_LOGS // init llama.cpp - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model = NULL; llama_context * ctx = NULL; diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp index 5c0022832146b..e12a1cdf19a79 100644 --- a/examples/passkey/passkey.cpp +++ b/examples/passkey/passkey.cpp @@ -71,7 +71,8 @@ int main(int argc, char ** argv) { // init LLM - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); // initialize the model diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b2c131d4ce6dd..67d2d3293a327 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1809,7 +1809,8 @@ int main(int argc, char ** argv) { params.prompt = gpt_random_prompt(rng); } - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model; llama_context * ctx; diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 85f403ffc9599..4a5c504e31c9c 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -237,7 +237,7 @@ int main(int argc, char ** argv) { params.imatrix = &imatrix_data; } - llama_backend_init(false); + llama_backend_init(); // parse command line arguments const std::string fname_inp = argv[arg_idx]; diff --git a/examples/server/README.md b/examples/server/README.md index 0f7373ae86204..8e141d22d1716 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -16,6 +16,13 @@ Command line options: - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended. - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped. - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. +- `--numa STRATEGY`: Attempt one of the below optimization strategies that help on some NUMA systems +- `--numa distribute`: Spread execution evenly over all nodes +- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on +- `--numa numactl`: Use the CPU map provided by numactl +if run without this previously, it is recommended to drop the system page cache before using this +see https://github.com/ggerganov/llama.cpp/issues/1437 + - `--numa`: Attempt optimizations that help on some NUMA systems. - `--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. diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 2decd776275a9..912c750cc6223 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1855,7 +1855,10 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, { printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); } - printf(" --numa attempt optimizations that help on some NUMA systems\n"); + printf(" --numa TYPE attempt optimizations that help on some NUMA systems\n"); + printf(" - distribute: spread execution evenly over all nodes\n"); + printf(" - isolate: only spawn threads on CPUs on the node that execution started on\n"); + printf(" - numactl: use the CPU map provided my numactl\n"); if (llama_supports_gpu_offload()) { printf(" -ngl N, --n-gpu-layers N\n"); printf(" number of layers to store in VRAM\n"); @@ -2264,9 +2267,17 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, { params.use_mmap = false; } - else if (arg == "--numa") - { - params.numa = true; + else if (arg == "--numa") { + if (++i >= argc) { + invalid_param = true; + break; + } else { + std::string value(argv[i]); + /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } + else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } + else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } + else { invalid_param = true; break; } + } } else if (arg == "--embedding") { @@ -2497,7 +2508,8 @@ int main(int argc, char **argv) params.model_alias = params.model; } - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER}, {"commit", LLAMA_COMMIT}}); diff --git a/examples/simple/simple.cpp b/examples/simple/simple.cpp index 9cfde8308f18f..39e2d8ea490e3 100644 --- a/examples/simple/simple.cpp +++ b/examples/simple/simple.cpp @@ -31,7 +31,8 @@ int main(int argc, char ** argv) { // init LLM - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); // initialize the model diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index 7b3af01f339a9..3848791d475ad 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -50,7 +50,8 @@ int main(int argc, char ** argv) { #endif // LOG_DISABLE_LOGS // init llama.cpp - llama_backend_init(params.numa); + llama_backend_init(); + llama_numa_init(params.numa); llama_model * model_tgt = NULL; llama_model * model_dft = NULL; diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp index 4ff8e3fa72749..d95a9247525eb 100644 --- a/examples/tokenize/tokenize.cpp +++ b/examples/tokenize/tokenize.cpp @@ -17,7 +17,7 @@ int main(int argc, char ** argv) { const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids"; - llama_backend_init(false); + llama_backend_init(); llama_model_params model_params = llama_model_default_params(); model_params.vocab_only = true; diff --git a/ggml.c b/ggml.c index d921d82fed7d3..4e302fb7de2f4 100644 --- a/ggml.c +++ b/ggml.c @@ -1954,9 +1954,16 @@ struct ggml_numa_node { }; struct ggml_numa_nodes { + enum ggml_numa_strategy numa_strategy; struct ggml_numa_node nodes[GGML_NUMA_MAX_NODES]; uint32_t n_nodes; uint32_t total_cpus; // hardware threads on system + uint32_t current_node; // node on which main process is execting +#ifdef __linux__ + cpu_set_t cpuset; // cpuset from numactl +#else + uint32_t cpuset; // no NUMA support outside of Linux at this time. Use a portable datatype +#endif }; // @@ -1990,7 +1997,22 @@ inline static void ggml_critical_section_end(void) { atomic_fetch_sub(&g_state_barrier, 1); } -void ggml_numa_init(void) { +#ifdef __linux__ +static cpu_set_t ggml_get_numa_affinity(void) { + cpu_set_t cpuset; + pthread_t thread; + thread = pthread_self(); + CPU_ZERO(&cpuset); + pthread_getaffinity_np(thread, sizeof(cpu_set_t), &cpuset); + return cpuset; +} +#else +static uint32_t ggml_get_numa_affinity(void) { + return 0; // no NUMA support +} +#endif + +void ggml_numa_init(enum ggml_numa_strategy numa_flag) { if (g_state.numa.n_nodes > 0) { fprintf(stderr, "ggml_numa_init: NUMA already initialized\n"); @@ -2002,6 +2024,13 @@ void ggml_numa_init(void) { char path[256]; int rv; + // set numa scheme + g_state.numa.numa_strategy = numa_flag; + + GGML_PRINT_DEBUG("numa strategy %u\n",g_state.numa.numa_strategy); + + g_state.numa.cpuset = ggml_get_numa_affinity(); + // enumerate nodes while (g_state.numa.n_nodes < GGML_NUMA_MAX_NODES) { rv = snprintf(path, sizeof(path), "/sys/devices/system/node/node%u", g_state.numa.n_nodes); @@ -2020,11 +2049,17 @@ void ggml_numa_init(void) { GGML_PRINT_DEBUG("found %u numa nodes, %u CPUs\n", g_state.numa.n_nodes, g_state.numa.total_cpus); - if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1) { + // figure out which node we're on + uint current_cpu; + int getcpu_ret = getcpu(¤t_cpu, &g_state.numa.current_node); + + if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) { g_state.numa.n_nodes = 0; return; } + GGML_PRINT_DEBUG("found our process on numa node %u, CPU %u\n", g_state.numa.current_node, current_cpu); + for (uint32_t n = 0; n < g_state.numa.n_nodes; ++n) { struct ggml_numa_node * node = &g_state.numa.nodes[n]; GGML_PRINT_DEBUG("CPUs on node %u:", n); @@ -16638,26 +16673,46 @@ typedef pthread_t ggml_thread_t; // Android's libc implementation "bionic" does not support setting affinity #if defined(__linux__) && !defined(__BIONIC__) -static void set_numa_thread_affinity(int thread_n, int n_threads) { +static void set_numa_thread_affinity(int thread_n) { if (!ggml_is_numa()) { return; } - // run thread on node_num thread_n / (threads per node) - const int node_num = thread_n / ((n_threads + g_state.numa.n_nodes - 1) / g_state.numa.n_nodes); - struct ggml_numa_node * node = &g_state.numa.nodes[node_num]; + int node_num; + int rv; size_t setsize = CPU_ALLOC_SIZE(g_state.numa.total_cpus); + switch(g_state.numa.numa_strategy) { + case GGML_NUMA_STRATEGY_DISTRIBUTE: + // run thread on node_num thread_n / (threads per node) + node_num = thread_n % g_state.numa.n_nodes; + break; + case GGML_NUMA_STRATEGY_ISOLATE: + // run thread on current_node + node_num = g_state.numa.current_node; + break; + case GGML_NUMA_STRATEGY_NUMACTL: + // use the cpuset that numactl gave us + rv = pthread_setaffinity_np(pthread_self(), setsize, &g_state.numa.cpuset); + if (rv) { + fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",strerror(rv)); + } + return; + default: + return; + } + + struct ggml_numa_node * node = &g_state.numa.nodes[node_num]; + cpu_set_t * cpus = CPU_ALLOC(g_state.numa.total_cpus); CPU_ZERO_S(setsize, cpus); for (size_t i = 0; i < node->n_cpus; ++i) { CPU_SET_S(node->cpus[i], setsize, cpus); } - int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus); + rv = pthread_setaffinity_np(pthread_self(), setsize, cpus); if (rv) { - fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", - strerror(rv)); + fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv)); } CPU_FREE(cpus); @@ -16678,8 +16733,7 @@ static void clear_numa_thread_affinity(void) { int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus); if (rv) { - fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", - strerror(rv)); + fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv)); } CPU_FREE(cpus); @@ -16687,7 +16741,7 @@ static void clear_numa_thread_affinity(void) { #else // TODO: Windows etc. // (the linux implementation may also work on BSD, someone should test) -static void set_numa_thread_affinity(int thread_n, int n_threads) { UNUSED(thread_n); UNUSED(n_threads); } +static void set_numa_thread_affinity(int thread_n) { UNUSED(thread_n); } static void clear_numa_thread_affinity(void) {} #endif @@ -16987,7 +17041,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { const int n_threads = state->shared->n_threads; - set_numa_thread_affinity(state->ith, n_threads); + set_numa_thread_affinity(state->ith); int node_n = -1; int task_phase = GGML_TASK_FINALIZE; diff --git a/ggml.h b/ggml.h index 01cecc1e1845f..270018185f397 100644 --- a/ggml.h +++ b/ggml.h @@ -658,6 +658,16 @@ extern "C" { void * wdata; }; + // numa strategies + enum ggml_numa_strategy { + GGML_NUMA_STRATEGY_DISABLED = 0, + GGML_NUMA_STRATEGY_DISTRIBUTE = 1, + GGML_NUMA_STRATEGY_ISOLATE = 2, + GGML_NUMA_STRATEGY_NUMACTL = 3, + GGML_NUMA_STRATEGY_MIRROR = 4, + GGML_NUMA_STRATEGY_COUNT + }; + // misc GGML_API void ggml_time_init(void); // call this once at the beginning of the program @@ -668,7 +678,7 @@ extern "C" { GGML_API void ggml_print_backtrace(void); - GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems + GGML_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node GGML_API void ggml_print_object (const struct ggml_object * obj); diff --git a/llama.cpp b/llama.cpp index aceb9c25a4e46..08e7b02b4cc1d 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1034,7 +1034,7 @@ struct llama_mmap { int fd = fileno(file->fp); int flags = MAP_SHARED; // prefetch/readahead impairs performance on NUMA systems - if (numa) { prefetch = 0; } + if (numa) { prefetch = 0; } #ifdef __linux__ // advise the kernel to read the file sequentially (increases readahead) if (posix_fadvise(fd, 0, 0, POSIX_FADV_SEQUENTIAL)) { @@ -11182,7 +11182,7 @@ bool llama_mlock_supported(void) { return llama_supports_mlock(); } -void llama_backend_init(bool numa) { +void llama_backend_init(void) { ggml_time_init(); // needed to initialize f16 tables @@ -11192,15 +11192,17 @@ void llama_backend_init(bool numa) { ggml_free(ctx); } - if (numa) { - ggml_numa_init(); - } - #ifdef GGML_USE_MPI ggml_mpi_backend_init(); #endif } +void llama_numa_init(enum ggml_numa_strategy numa) { + if (numa != GGML_NUMA_STRATEGY_DISABLED) { + ggml_numa_init(numa); + } +} + void llama_backend_free(void) { #ifdef GGML_USE_MPI ggml_mpi_backend_free(); diff --git a/llama.h b/llama.h index 4a26bd61988df..f4ec6ea6394a3 100644 --- a/llama.h +++ b/llama.h @@ -312,7 +312,10 @@ extern "C" { // Initialize the llama + ggml backend // If numa is true, use NUMA optimizations // Call once at the start of the program - LLAMA_API void llama_backend_init(bool numa); + LLAMA_API void llama_backend_init(void); + + //optional: + LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa); // Call once at the end of the program - currently only used for MPI LLAMA_API void llama_backend_free(void); diff --git a/tests/test-autorelease.cpp b/tests/test-autorelease.cpp index 36a23c0bb790c..57fa000114d5d 100644 --- a/tests/test-autorelease.cpp +++ b/tests/test-autorelease.cpp @@ -12,7 +12,7 @@ int main(int argc, char ** argv) { auto * model_path = get_model_or_exit(argc, argv); std::thread([&model_path]() { - llama_backend_init(false); + llama_backend_init(); auto * model = llama_load_model_from_file(model_path, llama_model_default_params()); auto * ctx = llama_new_context_with_model(model, llama_context_default_params()); llama_free(ctx); diff --git a/tests/test-model-load-cancel.cpp b/tests/test-model-load-cancel.cpp index 7ea4bbaccd8d1..858535c3c4020 100644 --- a/tests/test-model-load-cancel.cpp +++ b/tests/test-model-load-cancel.cpp @@ -14,7 +14,7 @@ int main(int argc, char *argv[] ) { fprintf(stderr, "using '%s'\n", model_path); fclose(file); - llama_backend_init(false); + llama_backend_init(); auto params = llama_model_params{}; params.use_mmap = false; params.progress_callback = [](float progress, void * ctx){ diff --git a/tests/test-tokenizer-0-falcon.cpp b/tests/test-tokenizer-0-falcon.cpp index a4e9d2b912728..472b0b3a8f436 100644 --- a/tests/test-tokenizer-0-falcon.cpp +++ b/tests/test-tokenizer-0-falcon.cpp @@ -61,7 +61,7 @@ int main(int argc, char **argv) { llama_model * model; llama_context * ctx; - llama_backend_init(false); + llama_backend_init(); // load the vocab { diff --git a/tests/test-tokenizer-0-llama.cpp b/tests/test-tokenizer-0-llama.cpp index 39c8d188c9086..0a16cd7eb404b 100644 --- a/tests/test-tokenizer-0-llama.cpp +++ b/tests/test-tokenizer-0-llama.cpp @@ -60,7 +60,7 @@ int main(int argc, char **argv) { llama_model * model; llama_context * ctx; - llama_backend_init(false); + llama_backend_init(); // load the vocab { diff --git a/tests/test-tokenizer-1-bpe.cpp b/tests/test-tokenizer-1-bpe.cpp index 3bb6295613fa6..3596ce55af2ce 100644 --- a/tests/test-tokenizer-1-bpe.cpp +++ b/tests/test-tokenizer-1-bpe.cpp @@ -25,7 +25,7 @@ int main(int argc, char **argv) { llama_model * model; llama_context * ctx; - llama_backend_init(false); + llama_backend_init(); // load the vocab { diff --git a/tests/test-tokenizer-1-llama.cpp b/tests/test-tokenizer-1-llama.cpp index b0d814a417fb6..9333f8686fa1c 100644 --- a/tests/test-tokenizer-1-llama.cpp +++ b/tests/test-tokenizer-1-llama.cpp @@ -25,7 +25,7 @@ int main(int argc, char **argv) { llama_model * model; llama_context * ctx; - llama_backend_init(false); + llama_backend_init(); // load the vocab {