Threadpool: take 2 (#8672)

* Introduce ggml_compute_threadpool

- OpenMP functional: check
- Vanilla ggml functional: Check
- ggml w/threadpool functional: Check
- OpenMP no regression: No glaring problems
- Vanilla ggml no regression: No glaring problems
- ggml w/threadpool no regression: No glaring problems

* Minor fixes

* fixed use after release bug

* fixed a harmless race condition

* Fix Android bulid issue

* fix more race conditions

* fix deadlock for cases where cgraph.n_nodes == 1

and fix --poll case

* threadpool: use cpu_get_num_math to set the default number of threadpool threads

This way we avoid using E-Cores and Hyperthreaded siblings.

* bench: create fresh threadpool for each test

For benchmarking it's better to start a fresh pool for each test with the exact number of threads
needed for that test. Having larger pools is suboptimal (causes more load, etc).

* atomics: always use stdatomics with clang and use relaxed memory order when polling in ggml_barrier

This also removes sched_yield() calls from ggml_barrier() to match OpenMP behavior.

* threadpool: make polling the default to match openmp behavior

All command line args now allow for setting poll to 0 (false).

* threadpool: do not wakeup threads in already paused threadpool

* fix potential race condition in check_for_work

* threadpool: do not create two threadpools if their params are identical

* threadpool: reduce pause/resume/wakeup overhead in common cases

We now start threadpool in paused state only if we have two.
The resume is now implicit (ie new work) which allows for reduced locking and context-switch overhead.

* threadpool: add support for hybrid polling

poll params (--poll, ...) now specify "polling level", i.e. how aggresively we poll before waiting on cond.var.
poll=0 means no polling, 1 means poll for 128K rounds then wait, 2 for 256K rounds, ...

The default value of 50 (ie 50x128K rounds) seems like a decent default across modern platforms.
We can tune this further as things evolve.

* threadpool: reduce the number of barrier required

New work is now indicated with an atomic counter that is incremented for
each new graph that needs to be computed.
This removes the need for extra barrier for clearing the "new_work" and
removes the special case for trivial graphs.

* threadpool: remove special-casing for disposable threadpools

With the efficient hybrid polling there is no need to make disposable pools any different.
This simplifies the overall logic and reduces branching.

Include n_threads in debug print for disposable threadpool.

Declare pause and stop flags as atomic_bool
This doesn't actually generate any memory barriers and simply informs
the thread sanitizer that these flags can be written & read by different
threads without locking.

* threadpool: do not clear barrier counters between graphs computes (fixes race with small graphs)

This fixes the race condition with very small graphs where the main thread happens to
start a new graph while the workers are just about to exit from barriers.

* threadpool: use relaxed order for chunk sync

Full memory barrier is an overkill for this since each thread works on different chunk

* threadpool: remove abort_callback from threadpool state

* threadpool: better naming for thread/cpumask releated functions

* threadpool: consistent use of int type for n_threads params

* threadpool: add support for ggml_threadpool_params_default/init

Also removes the need for explicit mask_specified param.
all-zero cpumask means use default (usually inherited) cpu affinity mask.

* threadpool: move typedef into ggml.h

* threadpool: fix apply_priority() function name

* threadpool: fix swift wrapper errors due to n_threads int type cleanup

* threadpool: enable --cpu-mask and other threadpool related options only if threadpool is enabled

* threadpool: replace checks for compute_thread ret code with proper status check

* threadpool: simplify threadpool init logic and fix main thread affinity application

Most of the init code is now exactly the same between threadpool and openmp.

* threadpool: update threadpool resume/pause function names

* threadpool: enable openmp by default for now

* threadpool: don't forget to free workers state when omp is enabled

* threadpool: avoid updating process priority on the platforms that do not require it

On Windows we need to change overall process priority class in order to set thread priorities,
but on Linux, Mac, etc we do not need to touch the overall process settings.

* threadpool: update calling thread prio and affinity only at start/resume

This avoids extra syscalls for each graph_compute()

* llama-bench: turn threadpool params into vectors, add output headers, etc

* llama-bench: add support for cool off between tests --delay

This helps for long running tests on platforms that are thermally limited (phones, laptops, etc).
--delay (disabled by default) introduces the sleep for N seconds before starting each test.

* threadpool: move process priority setting into the apps (bench and cli)

This avoids changing the overall process priority on Windows for the apps
that use ggml/llama.cpp directy.

* threadpool: move all pause/resume logic into ggml

* threadpool: futher api cleanup and prep for future refactoring

All threadpool related functions and structs use ggml_threadpool prefix.

* threadpool: minor indent fixes

* threadpool: improve setprioty error message

* Update examples/llama-bench/llama-bench.cpp

Co-authored-by: slaren <slarengh@gmail.com>

* threadpool: fix indent in set_threadpool call

* use int32_t for n_thread type in public llama.cpp API

* threadpool: use _new and _free instead of _create and _release

* fix two more public APIs to use int32_t for n_threads

* build: set _GNU_SOURCE for Adroid

---------

Co-authored-by: Max Krasnyansky <quic_maxk@quicinc.com>
Co-authored-by: fmz <quic_fzaghlou@quic.com>
Co-authored-by: Max Krasnyansky <max.krasnyansky@gmail.com>
Co-authored-by: slaren <slarengh@gmail.com>

common/common.cpp

@@ -251,6 +251,57 @@ int32_t cpu_get_num_math() {
     return cpu_get_num_physical_cores();
 }
 
+// Helper for setting process priority
+
+#if defined(_WIN32)
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    DWORD p = NORMAL_PRIORITY_CLASS;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
+        case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
+        case GGML_SCHED_PRIO_REALTIME: p = REALTIME_PRIORITY_CLASS;     break;
+    }
+
+    if (!SetPriorityClass(GetCurrentProcess(), p)) {
+        fprintf(stderr, "warn: failed to set process priority class %d : (%d)\n", prio, (int) GetLastError());
+        return false;
+    }
+
+    return true;
+}
+
+#else // MacOS and POSIX
+#include <sys/types.h>
+#include <sys/resource.h>
+
+bool set_process_priority(enum ggml_sched_priority prio) {
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        return true;
+    }
+
+    int p = 0;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
+        case GGML_SCHED_PRIO_HIGH:     p = -10; break;
+        case GGML_SCHED_PRIO_REALTIME: p = -20; break;
+    }
+
+    if (!setpriority(PRIO_PROCESS, 0, p)) {
+        fprintf(stderr, "warn: failed to set process priority %d : %s (%d)\n", prio, strerror(errno), errno);
+        return false;
+    }
+    return true;
+}
+
+#endif
+
 //
 // CLI argument parsing
 //
@@ -277,6 +328,30 @@ void gpt_params_handle_model_default(gpt_params & params) {
     }
 }
 
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
+    int32_t n_set = 0;
+
+    if (cpuparams.n_threads < 0) {
+        // Assuming everything about cpuparams is invalid
+        if (role_model != nullptr) {
+            cpuparams = *role_model;
+        } else {
+            cpuparams.n_threads = cpu_get_num_math();
+        }
+    }
+
+    for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+        if (cpuparams.cpumask[i]) {
+            n_set++;
+        }
+    }
+
+    if (n_set && n_set < cpuparams.n_threads) {
+        // Not enough set bits, may experience performance issues.
+        fprintf(stderr, "warn: Not enough set bits in CPU mask (%d) to satisfy requested thread count: %d\n", n_set, cpuparams.n_threads);
+    }
+}
+
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
@@ -296,6 +371,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
+    postprocess_cpu_params(params.cpuparams, nullptr);
+    postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
+    postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
+
     if (params.prompt_cache_all && (params.interactive || params.interactive_first)) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
@@ -331,7 +411,7 @@ void gpt_params_parse_from_env(gpt_params & params) {
     get_env("LLAMA_ARG_MODEL_ALIAS",      params.model_alias);
     get_env("LLAMA_ARG_HF_REPO",          params.hf_repo);
     get_env("LLAMA_ARG_HF_FILE",          params.hf_file);
-    get_env("LLAMA_ARG_THREADS",          params.n_threads);
+    get_env("LLAMA_ARG_THREADS",          params.cpuparams.n_threads);
     get_env("LLAMA_ARG_CTX_SIZE",         params.n_ctx);
     get_env("LLAMA_ARG_N_PARALLEL",       params.n_parallel);
     get_env("LLAMA_ARG_BATCH",            params.n_batch);
@@ -368,6 +448,79 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     return true;
 }
 
+bool parse_cpu_range(const std::string & range, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    size_t dash_loc = range.find('-');
+    if (dash_loc == std::string::npos) {
+        fprintf(stderr, "Format of CPU range is invalid! Expected [<start>]-[<end>].\n");
+        return false;
+    }
+
+    size_t start_i;
+    size_t end_i;
+
+    if (dash_loc == 0) {
+        start_i = 0;
+    } else {
+        start_i = std::stoull(range.substr(0, dash_loc));
+        if (start_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "Start index out of bounds!\n");
+            return false;
+        }
+    }
+
+    if (dash_loc == range.length() - 1) {
+        end_i = GGML_MAX_N_THREADS - 1;
+    } else {
+        end_i = std::stoull(range.substr(dash_loc + 1));
+        if (end_i >= GGML_MAX_N_THREADS) {
+            fprintf(stderr, "End index out of bounds!\n");
+            return false;
+        }
+    }
+
+    for (size_t i = start_i; i <= end_i; i++) {
+        boolmask[i] = true;
+    }
+
+    return true;
+}
+
+bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREADS]) {
+    // Discard potential 0x prefix
+    size_t start_i = 0;
+    if (mask.length() >= 2 && mask.substr(0, 2) == "0x") {
+        start_i = 2;
+    }
+
+    size_t num_digits = mask.length() - start_i;
+    if (num_digits > 128) num_digits = 128;
+
+    size_t end_i = num_digits + start_i;
+
+    for (size_t i = start_i, n = (num_digits*4 - 1); i < end_i; i++, n-=4) {
+        char c = mask.at(i);
+        int8_t id = c;
+
+        if ((c >= '0' && c <= '9')) {
+            id -= '0';
+        } else if (c >= 'a' && c <= 'f') {
+            id -= 'a' - 10;
+        } else if (c >= 'A' && c <= 'F') {
+            id -= 'A' - 10;
+        } else {
+            fprintf(stderr, "Invalid hex character '%c' at position %d\n", c, int32_t(i));
+            return false;
+        }
+
+        boolmask[  n  ] = boolmask[  n  ] || ((id & 8) != 0);
+        boolmask[n - 1] = boolmask[n - 1] || ((id & 4) != 0);
+        boolmask[n - 2] = boolmask[n - 2] || ((id & 2) != 0);
+        boolmask[n - 3] = boolmask[n - 3] || ((id & 1) != 0);
+    }
+
+    return true;
+}
+
 #define CHECK_ARG if (++i >= argc) { invalid_param = true; return true; }
 
 bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) {
@@ -384,36 +537,142 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     }
     if (arg == "-t" || arg == "--threads") {
         CHECK_ARG
-        params.n_threads = std::stoi(argv[i]);
+        params.cpuparams.n_threads = std::stoi(argv[i]);
-        if (params.n_threads <= 0) {
+        if (params.cpuparams.n_threads <= 0) {
-            params.n_threads = std::thread::hardware_concurrency();
+            params.cpuparams.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-C" || arg == "--cpu-mask") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Cr" || arg == "--cpu-range") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio") {
+        CHECK_ARG
+        params.cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict") {
+        CHECK_ARG
+        params.cpuparams.strict_cpu = std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--poll") {
+        CHECK_ARG
+        params.cpuparams.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-tb" || arg == "--threads-batch") {
         CHECK_ARG
-        params.n_threads_batch = std::stoi(argv[i]);
+        params.cpuparams_batch.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_batch <= 0) {
+        if (params.cpuparams_batch.n_threads <= 0) {
-            params.n_threads_batch = std::thread::hardware_concurrency();
+            params.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-Cb" || arg == "--cpu-mask-batch") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "-Crb" || arg == "--cpu-range_batch") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch") {
+        params.cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch") {
+        CHECK_ARG
+        params.cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-td" || arg == "--threads-draft") {
         CHECK_ARG
-        params.n_threads_draft = std::stoi(argv[i]);
+        params.draft_cpuparams.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_draft <= 0) {
+        if (params.draft_cpuparams.n_threads <= 0) {
-            params.n_threads_draft = std::thread::hardware_concurrency();
+            params.draft_cpuparams.n_threads = std::thread::hardware_concurrency();
         }
         return true;
+    }
+        if (arg == "-Cd" || arg == "--cpu-mask-draft") {
+        CHECK_ARG
+        std::string mask = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_mask(mask, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "-Crd" || arg == "--cpu-range-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams.cpumask);
+        return true;
+    }
+    if (arg == "--prio-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-draft") {
+        params.draft_cpuparams.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-draft") {
+        CHECK_ARG
+        params.draft_cpuparams.poll = std::stoul(argv[i]);
+        return true;
     }
     if (arg == "-tbd" || arg == "--threads-batch-draft") {
         CHECK_ARG
-        params.n_threads_batch_draft = std::stoi(argv[i]);
+        params.draft_cpuparams_batch.n_threads = std::stoi(argv[i]);
-        if (params.n_threads_batch_draft <= 0) {
+        if (params.draft_cpuparams_batch.n_threads <= 0) {
-            params.n_threads_batch_draft = std::thread::hardware_concurrency();
+            params.draft_cpuparams_batch.n_threads = std::thread::hardware_concurrency();
         }
         return true;
     }
+    if (arg == "-Crbd" || arg == "--cpu-range-batch-draft") {
+        CHECK_ARG
+        std::string range = argv[i];
+        params.draft_cpuparams_batch.mask_valid = true;
+        invalid_param = !parse_cpu_range(range, params.draft_cpuparams_batch.cpumask);
+        return true;
+    }
+    if (arg == "--prio-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.priority = (enum ggml_sched_priority) std::stoul(argv[i]);
+        return true;
+    }
+    if (arg == "--cpu-strict-batch-draft") {
+        params.draft_cpuparams_batch.strict_cpu = true;
+        return true;
+    }
+    if (arg == "--poll-batch-draft") {
+        CHECK_ARG
+        params.draft_cpuparams_batch.poll = std::stoul(argv[i]);
+        return true;
+    }
     if (arg == "-p" || arg == "--prompt") {
         CHECK_ARG
         params.prompt = argv[i];
@@ -1498,11 +1757,40 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "*",           "       --no-display-prompt",    "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" });
     options.push_back({ "*",           "-co,   --color",                "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" });
     options.push_back({ "*",           "-s,    --seed SEED",            "RNG seed (default: %d, use random seed for < 0)", params.seed });
-    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.n_threads });
+    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during generation (default: %d)", params.cpuparams.n_threads });
     options.push_back({ "*",           "-tb,   --threads-batch N",      "number of threads to use during batch and prompt processing (default: same as --threads)" });
     options.push_back({ "speculative", "-td,   --threads-draft N",      "number of threads to use during generation (default: same as --threads)" });
-    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N",
+    options.push_back({ "speculative", "-tbd,  --threads-batch-draft N","number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
-                                                                        "number of threads to use during batch and prompt processing (default: same as --threads-draft)" });
+
+#ifndef GGML_USE_OPENMP
+    // these options are available only with the internal threadpool
+    options.push_back({ "*",           "-C,    --cpu-mask M",            "CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: \"\")"});
+    options.push_back({ "*",           "-Cr,   --cpu-range lo-hi",       "range of CPUs for affinity. Complements --cpu-mask"});
+    options.push_back({ "*",           "       --cpu-strict <0|1>",      "use strict CPU placement (default: %u)\n", (unsigned) params.cpuparams.strict_cpu});
+    options.push_back({ "*",           "       --priority N",            "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority});
+    options.push_back({ "*",           "       --poll <0...100>",        "use polling level to wait for work (0 - no polling, default: %u)\n", (unsigned) params.cpuparams.poll});
+
+    options.push_back({ "*",           "-Cb,   --cpu-mask-batch M",      "CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask)"});
+    options.push_back({ "*",           "-Crb,  --cpu-range-batch lo-hi", "ranges of CPUs for affinity. Complements --cpu-mask-batch"});
+    options.push_back({ "*",           "       --cpu-strict-batch <0|1>","use strict CPU placement (default: same as --cpu-strict)"});
+    options.push_back({ "*",           "       --priority-batch N",      "set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority)"});
+    options.push_back({ "*",           "       --poll-batch <0|1>",      "use polling to wait for work (default: same as --poll"});
+
+    options.push_back({ "speculative", "-Cd,   --cpu-mask-draft M",      "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)"});
+    options.push_back({ "speculative", "-Crd,  --cpu-range-draft lo-hi", "Ranges of CPUs for affinity. Complements --cpu-mask-draft"});
+    options.push_back({ "speculative", "       --cpu-strict-draft <0|1>","Use strict CPU placement for draft model (default: same as --cpu-strict)"});
+    options.push_back({ "speculative", "       --priority-draft N",      "Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: same as --priority)"});
+    options.push_back({ "speculative", "       --poll-draft <0|1>",      "Use polling to wait for draft model work (default: same as --poll])"});
+
+    options.push_back({ "speculative", "-Cbd,  --cpu-mask-batch-draft M","Draft model CPU affinity mask. Complements cpu-range-draft-batch (default: same as --cpu-mask-draft)"});
+    options.push_back({ "speculative", "-Crbd, --cpu-range-batch-draft lo-hi",
+                                                                         "Ranges of CPUs for affinity. Complements --cpu-mask-draft-batch)"});
+    options.push_back({ "speculative", "       --cpu-strict-batch-draft <0|1>",
+                                                                         "Use strict CPU placement for draft model (default: --cpu-strict-draft)"});
+    options.push_back({ "speculative", "       --priority-batch-draft N","Set draft process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: --priority-draft)"});
+    options.push_back({ "speculative", "       --poll-batch-draft <0|1>","Use polling to wait for draft model work (default: --poll-draft)"});
+#endif // GGML_USE_OPENMP
+
     options.push_back({ "speculative", "       --draft N",              "number of tokens to draft for speculative decoding (default: %d)", params.n_draft });
     options.push_back({ "speculative", "-ps,   --p-split N",            "speculative decoding split probability (default: %.1f)", (double)params.p_split });
     options.push_back({ "*",           "-lcs,  --lookup-cache-static FNAME",
@@ -1774,7 +2062,6 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     options.push_back({ "export-lora", "-m,    --model",                "model path from which to load base model (default '%s')", params.model.c_str() });
     options.push_back({ "export-lora", "       --lora FNAME",           "path to LoRA adapter  (can be repeated to use multiple adapters)" });
     options.push_back({ "export-lora", "       --lora-scaled FNAME S",  "path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)" });
-    options.push_back({ "*",           "-t,    --threads N",            "number of threads to use during computation (default: %d)", params.n_threads });
     options.push_back({ "export-lora", "-o,    --output FNAME",         "output file (default: '%s')", params.lora_outfile.c_str() });
 
     printf("usage: %s [options]\n", argv[0]);
@@ -1806,9 +2093,9 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
 std::string gpt_params_get_system_info(const gpt_params & params) {
     std::ostringstream os;
 
-    os << "system_info: n_threads = " << params.n_threads;
+    os << "system_info: n_threads = " << params.cpuparams.n_threads;
-    if (params.n_threads_batch != -1) {
+    if (params.cpuparams_batch.n_threads != -1) {
-        os << " (n_threads_batch = " << params.n_threads_batch << ")";
+        os << " (n_threads_batch = " << params.cpuparams_batch.n_threads << ")";
     }
 #if defined(_WIN32) && (_WIN32_WINNT >= 0x0601) && !defined(__MINGW64__) // windows 7 and later
     // TODO: windows + arm64 + mingw64
@@ -2332,8 +2619,9 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.n_seq_max         = params.n_parallel;
     cparams.n_batch           = params.n_batch;
     cparams.n_ubatch          = params.n_ubatch;
-    cparams.n_threads         = params.n_threads;
+    cparams.n_threads         = params.cpuparams.n_threads;
-    cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
+                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.seed              = params.seed;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
@@ -2359,6 +2647,22 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     return cparams;
 }
 
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
+    struct ggml_threadpool_params tpp;
+
+    ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults
+
+    if (params.mask_valid) {
+        std::memcpy(&tpp.cpumask, &params.cpumask, GGML_MAX_N_THREADS);
+    }
+
+    tpp.prio       = params.priority;
+    tpp.poll       = params.poll;
+    tpp.strict_cpu = params.strict_cpu;
+
+    return tpp;
+}
+
 #ifdef LLAMA_USE_CURL
 
 static bool starts_with(const std::string & str, const std::string & prefix) {
@@ -3348,7 +3652,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
-    fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);

common/common.h

@@ -67,13 +67,18 @@ enum dimre_method {
     DIMRE_METHOD_MEAN,
 };
 
+struct cpu_params {
+    int      n_threads                   = -1;
+    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
+    bool     mask_valid                  = false;   // Default: any CPU
+    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
+    bool     strict_cpu                  = false;   // Use strict CPU placement
+    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
+};
+
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = cpu_get_num_math();
-    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                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@@ -100,6 +105,11 @@ struct gpt_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    struct cpu_params cpuparams;
+    struct cpu_params cpuparams_batch;
+    struct cpu_params draft_cpuparams;
+    struct cpu_params draft_cpuparams_batch;
+
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
@@ -204,7 +214,7 @@ struct gpt_params {
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";
@@ -277,6 +287,11 @@ void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
 
 std::string gpt_params_get_system_info(const gpt_params & params);
 
+bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
+bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
+void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
+bool set_process_priority(enum ggml_sched_priority prio);
+
 //
 // String utils
 //
@@ -327,8 +342,9 @@ struct llama_init_result {
 
 struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
-struct llama_model_params   llama_model_params_from_gpt_params  (const 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);
+struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);

examples/baby-llama/baby-llama.cpp

@@ -18,7 +18,7 @@ constexpr float rms_norm_eps = 5e-6f;
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);

examples/benchmark/benchmark-matmult.cpp

@@ -21,7 +21,7 @@
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);
@@ -54,7 +54,7 @@ static void tensor_dump(const ggml_tensor * tensor, const char * name) {
 #define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
 
 struct benchmark_params_struct {
-    int32_t n_threads     = 1;
+    int     n_threads     = 1;
     int32_t n_iterations  = 10;
 };
 

examples/cvector-generator/cvector-generator.cpp

@@ -486,8 +486,8 @@ int main(int argc, char ** argv) {
     if (use_pca) {
         // run PCA
         PCA::pca_params pca_params;
-        pca_params.n_threads = params.n_threads;
+        pca_params.n_threads    = params.cpuparams.n_threads;
-        pca_params.n_batch = params.n_pca_batch;
+        pca_params.n_batch      = params.n_pca_batch;
         pca_params.n_iterations = params.n_pca_iterations;
         PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
     } else {

examples/export-lora/export-lora.cpp

@@ -410,7 +410,7 @@ int main(int argc, char ** argv) {
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
+        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
         ctx.run_merge();
     } catch (const std::exception & err) {
         fprintf(stderr, "%s\n", err.what());

examples/llama-bench/llama-bench.cpp

@@ -16,6 +16,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <thread>
 
 #include "ggml.h"
 #include "llama.h"
@@ -225,6 +226,9 @@ struct cmd_params {
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
+    std::vector<std::string> cpu_mask;
+    std::vector<bool> cpu_strict;
+    std::vector<int> poll;
     std::vector<int> n_gpu_layers;
     std::vector<std::string> rpc_servers;
     std::vector<llama_split_mode> split_mode;
@@ -236,6 +240,8 @@ struct cmd_params {
     std::vector<bool> embeddings;
     ggml_numa_strategy numa;
     int reps;
+    ggml_sched_priority prio;
+    int delay;
     bool verbose;
     output_formats output_format;
     output_formats output_format_stderr;
@@ -251,6 +257,9 @@ static const cmd_params cmd_params_defaults = {
     /* type_k               */ {GGML_TYPE_F16},
     /* type_v               */ {GGML_TYPE_F16},
     /* n_threads            */ {cpu_get_num_math()},
+    /* cpu_mask             */ {"0x0"},
+    /* cpu_strict           */ {false},
+    /* poll                 */ {50},
     /* n_gpu_layers         */ {99},
     /* rpc_servers          */ {""},
     /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
@@ -262,6 +271,8 @@ static const cmd_params cmd_params_defaults = {
     /* embeddings           */ {false},
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
+    /* prio                 */ GGML_SCHED_PRIO_NORMAL,
+    /* delay                */ 0,
     /* verbose              */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
@@ -281,6 +292,9 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -ctk, --cache-type-k <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv, --cache-type-v <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -C, --cpu-mask <hex,hex>            (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str());
+    printf("  --cpu-strict <0|1>                  (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
+    printf("  --poll <0...100>                    (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     printf("  -rpc, --rpc <rpc_servers>           (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
     printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
@@ -292,6 +306,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
+    printf("  --prio <0|1|2|3>                    (default: %d)\n", cmd_params_defaults.prio);
+    printf("  --delay <0...N> (seconds)           (default: %d)\n", cmd_params_defaults.delay);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -oe, --output-err <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
@@ -338,6 +354,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.output_format_stderr = cmd_params_defaults.output_format_stderr;
     params.reps = cmd_params_defaults.reps;
     params.numa = cmd_params_defaults.numa;
+    params.prio = cmd_params_defaults.prio;
+    params.delay = cmd_params_defaults.delay;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -433,6 +451,27 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
+        } else if (arg == "-C" || arg == "--cpu-mask") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<std::string>(argv[i], split_delim);
+            params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end());
+        } else if (arg == "--cpu-strict") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<bool>(argv[i], split_delim);
+            params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end());
+        } else if (arg == "--poll") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = string_split<int>(argv[i], split_delim);
+            params.poll.insert(params.poll.end(), p.begin(), p.end());
         } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -541,6 +580,18 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.reps = std::stoi(argv[i]);
+        } else if (arg == "--prio") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.prio = (enum ggml_sched_priority) std::stoi(argv[i]);
+        } else if (arg == "--delay") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.delay = std::stoi(argv[i]);
         } else if (arg == "-o" || arg == "--output") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -585,6 +636,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
     if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
+    if (params.cpu_mask.empty())     { params.cpu_mask  = cmd_params_defaults.cpu_mask;  }
+    if (params.cpu_strict.empty())   { params.cpu_strict = cmd_params_defaults.cpu_strict; }
+    if (params.poll.empty())         { params.poll = cmd_params_defaults.poll; }
 
     return params;
 }
@@ -598,6 +652,9 @@ struct cmd_params_instance {
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
+    std::string cpu_mask;
+    bool cpu_strict;
+    int poll;
     int n_gpu_layers;
     std::string rpc_servers;
     llama_split_mode split_mode;
@@ -667,7 +724,10 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & tv : params.type_v)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & fa : params.flash_attn)
-    for (const auto & nt : params.n_threads) {
+    for (const auto & nt : params.n_threads)
+    for (const auto & cm : params.cpu_mask)
+    for (const auto & cs : params.cpu_strict)
+    for (const auto & pl : params.poll) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
                 continue;
@@ -681,6 +741,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -707,6 +770,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -733,6 +799,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
+                /* .cpu_mask     = */ cm,
+                /* .cpu_strict   = */ cs,
+                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -769,6 +838,9 @@ struct test {
     int n_batch;
     int n_ubatch;
     int n_threads;
+    std::string cpu_mask;
+    bool cpu_strict;
+    int poll;
     bool has_rpc;
     ggml_type type_k;
     ggml_type type_v;
@@ -795,6 +867,9 @@ struct test {
         n_batch = inst.n_batch;
         n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
+        cpu_mask = inst.cpu_mask;
+        cpu_strict = inst.cpu_strict;
+        poll = inst.poll;
         has_rpc = !inst.rpc_servers.empty();
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -872,13 +947,14 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
-            "n_threads", "type_k", "type_v",
+            "n_threads", "cpu_mask", "cpu_strict", "poll",
+            "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload", "flash_attn",
             "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
-            "avg_ts", "stddev_ts"
+            "avg_ts", "stddev_ts",
         };
         return fields;
     }
@@ -887,7 +963,7 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" ||
+            field == "n_threads" || field == "poll" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -896,6 +972,7 @@ struct test {
         }
         if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
+            field == "cpu_strict" ||
             field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
@@ -928,7 +1005,8 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_ubatch),
-            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
+            ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
             tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
@@ -1067,7 +1145,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 16;
+            return 20;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1149,6 +1227,15 @@ struct markdown_printer : public printer {
         if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
             fields.emplace_back("n_threads");
         }
+        if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) {
+            fields.emplace_back("cpu_mask");
+        }
+        if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) {
+            fields.emplace_back("cpu_strict");
+        }
+        if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) {
+            fields.emplace_back("poll");
+        }
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
@@ -1383,6 +1470,8 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    set_process_priority(params.prio);
+
     // initialize printer
     std::unique_ptr<printer> p = create_printer(params.output_format);
     std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
@@ -1428,6 +1517,28 @@ int main(int argc, char ** argv) {
 
         llama_kv_cache_clear(ctx);
 
+        // cool off before the test
+        if (params.delay) {
+            std::this_thread::sleep_for(std::chrono::seconds(params.delay));
+        }
+
+        struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads);
+        if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) {
+            LOG_TEE("%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str());
+            exit(1);
+        }
+        tpp.strict_cpu = t.cpu_strict;
+        tpp.poll       = t.poll;
+        tpp.prio       = params.prio;
+
+        struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
+        if (!threadpool) {
+            LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+            exit(1);
+        }
+
+        llama_attach_threadpool(ctx, threadpool, NULL);
+
         // warmup run
         if (t.n_prompt > 0) {
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
@@ -1466,6 +1577,8 @@ int main(int argc, char ** argv) {
         llama_print_timings(ctx);
 
         llama_free(ctx);
+
+        ggml_threadpool_free(threadpool);
     }
 
     llama_free_model(lmodel);

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -71,8 +71,8 @@ actor LlamaContext {
         var ctx_params = llama_context_default_params()
         ctx_params.seed  = 1234
         ctx_params.n_ctx = 2048
-        ctx_params.n_threads       = UInt32(n_threads)
+        ctx_params.n_threads       = Int32(n_threads)
-        ctx_params.n_threads_batch = UInt32(n_threads)
+        ctx_params.n_threads_batch = Int32(n_threads)
 
         let context = llama_new_context_with_model(model, ctx_params)
         guard let context else {

examples/llava/llava-cli.cpp

@@ -129,14 +129,14 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
         if (!params->image.empty()) {
             LOG_TEE("using base64 encoded image instead of command line image path\n");
         }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
         if (!embed) {
             LOG_TEE("%s: can't load image from prompt\n", __func__);
             return NULL;
         }
         params->prompt = remove_image_from_prompt(prompt);
     } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str());
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
         if (!embed) {
             fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
             return NULL;

examples/llava/minicpmv-cli.cpp

@@ -180,7 +180,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
 
 static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
     auto ctx_clip = clip_init_context(params);
-    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
+    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
     if (!embeds) {
         std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
         return NULL;

examples/main/main.cpp

@@ -221,6 +221,40 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    LOG("%s: llama threadpool init = n_threads = %d\n",
+        __func__,
+        (int) params.cpuparams.n_threads
+    );
+    struct ggml_threadpool_params tpp_batch =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
+    struct ggml_threadpool_params tpp =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams);
+
+    set_process_priority(params.cpuparams.priority);
+
+    struct ggml_threadpool * threadpool_batch = NULL;
+    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+        threadpool_batch = ggml_threadpool_new(&tpp_batch);
+        if (!threadpool_batch) {
+            LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
+            exit(1);
+        }
+
+        // Start the non-batch threadpool in the paused state
+        tpp.paused = true;
+    }
+
+    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+    if (!threadpool) {
+        LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+        exit(1);
+    }
+
+    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
+    if (ctx_guidance) {
+        llama_attach_threadpool(ctx_guidance, threadpool, threadpool_batch);
+    }
+
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
     LOG("n_ctx: %d\n", n_ctx);
@@ -989,6 +1023,9 @@ int main(int argc, char ** argv) {
     llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
+    ggml_threadpool_free(threadpool);
+    ggml_threadpool_free(threadpool_batch);
+
 #ifndef LOG_DISABLE_LOGS
     LOG_TEE("Log end\n");
 #endif // LOG_DISABLE_LOGS

examples/server/server.cpp

@@ -2534,8 +2534,8 @@ int main(int argc, char ** argv) {
     });
 
     LOG_INFO("system info", {
-        {"n_threads",       params.n_threads},
+        {"n_threads",       params.cpuparams.n_threads},
-        {"n_threads_batch", params.n_threads_batch},
+        {"n_threads_batch", params.cpuparams_batch.n_threads},
         {"total_threads",   std::thread::hardware_concurrency()},
         {"system_info",     llama_print_system_info()},
     });

examples/speculative/speculative.cpp

@@ -73,10 +73,11 @@ int main(int argc, char ** argv) {
     // load the draft model
     params.model = params.model_draft;
     params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.n_threads_draft > 0) {
+    if (params.draft_cpuparams.n_threads > 0) {
-        params.n_threads = params.n_threads_draft;
+        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
     }
-    params.n_threads_batch = params.n_threads_batch_draft;
+
+    params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
     llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
     model_dft = llama_init_dft.model;
     ctx_dft = llama_init_dft.context;

ggml/include/ggml-alloc.h

@@ -7,8 +7,8 @@ extern "C" {
 #endif
 
 typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct      ggml_backend_buffer * ggml_backend_buffer_t;
-typedef struct ggml_backend * ggml_backend_t;
+typedef struct             ggml_backend * ggml_backend_t;
 
 // Tensor allocator
 struct ggml_tallocr {

ggml/include/ggml-backend.h

@@ -103,6 +103,7 @@ extern "C" {
 
     GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
     GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
     GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
     // Create a backend buffer from an existing pointer

ggml/include/ggml.h

@@ -231,6 +231,8 @@
 #define GGML_MAX_SRC            10
 #ifndef GGML_MAX_NAME
 #define GGML_MAX_NAME           64
+#define GGML_MAX_N_THREADS      512
+
 #endif
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
@@ -628,6 +630,29 @@ extern "C" {
     // If it returns true, the computation is aborted
     typedef bool (*ggml_abort_callback)(void * data);
 
+    // Scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // Threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct  ggml_threadpool * ggml_threadpool_t;
+
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -635,6 +660,7 @@ extern "C" {
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
 
         int n_threads;
+        struct ggml_threadpool * threadpool;
 
         // abort ggml_graph_compute when true
         ggml_abort_callback abort_callback;
@@ -2057,10 +2083,23 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
+    GGML_API struct ggml_threadpool_params   ggml_threadpool_params_default(int n_threads);
+    GGML_API void                            ggml_threadpool_params_init  (struct ggml_threadpool_params *p, int n_threads);
+    GGML_API bool                            ggml_threadpool_params_match (const struct ggml_threadpool_params *p0, const struct ggml_threadpool_params *p1);
+    GGML_API struct ggml_threadpool*         ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_API void                            ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API int                             ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API void                            ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                            ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan   (const struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    GGML_API struct ggml_cplan ggml_graph_plan(
-    GGML_API enum ggml_status  ggml_graph_compute(      struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
+    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
     GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);

ggml/src/CMakeLists.txt

@@ -1247,7 +1247,7 @@ endif()
 
 # Data types, macros and functions related to controlling CPU affinity and
 # some memory allocation are available on Linux through GNU extensions in libc
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
     add_compile_definitions(_GNU_SOURCE)
 endif()
 

ggml/src/ggml-backend.c

@@ -722,9 +722,11 @@ ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
 #endif
 
 struct ggml_backend_cpu_context {
-    int n_threads;
+    int                 n_threads;
-    void * work_data;
+    ggml_threadpool_t   threadpool;
-    size_t work_size;
+
+    void *              work_data;
+    size_t              work_size;
 
     ggml_abort_callback abort_callback;
     void *              abort_callback_data;
@@ -759,7 +761,7 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg
 
     struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
 
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
     cpu_plan->cgraph = *cgraph; // FIXME: deep copy
 
     if (cpu_plan->cplan.work_size > 0) {
@@ -796,7 +798,7 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backe
 GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
 
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
 
     if (cpu_ctx->work_size < cplan.work_size) {
         free(cpu_ctx->work_data);
@@ -873,6 +875,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     }
 
     ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
     ctx->work_data           = NULL;
     ctx->work_size           = 0;
     ctx->abort_callback      = NULL;
@@ -903,6 +906,18 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
     ctx->n_threads = n_threads;
 }
 
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
 void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
     GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
 

include/llama.h

@@ -304,8 +304,8 @@ extern "C" {
         uint32_t n_batch;           // logical maximum batch size that can be submitted to llama_decode
         uint32_t n_ubatch;          // physical maximum batch size
         uint32_t n_seq_max;         // max number of sequences (i.e. distinct states for recurrent models)
-        uint32_t n_threads;         // number of threads to use for generation
+        int32_t  n_threads;         // number of threads to use for generation
-        uint32_t n_threads_batch;   // number of threads to use for batch processing
+        int32_t  n_threads_batch;   // number of threads to use for batch processing
 
         enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
         enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
@@ -428,6 +428,13 @@ extern "C" {
     //optional:
     LLAMA_API void llama_numa_init(enum ggml_numa_strategy numa);
 
+    // Optional: an auto threadpool gets created in ggml if not passed explicitly
+    LLAMA_API void llama_attach_threadpool(
+               struct   llama_context * ctx,
+            ggml_threadpool_t   threadpool,
+            ggml_threadpool_t   threadpool_batch);
+    LLAMA_API void llama_detach_threadpool(struct llama_context * ctx);
+
     // Call once at the end of the program - currently only used for MPI
     LLAMA_API void llama_backend_free(void);
 
@@ -837,13 +844,13 @@ extern "C" {
     // Set the number of threads used for decoding
     // n_threads is the number of threads used for generation (single token)
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
-    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
+    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch);
 
     // Get the number of threads used for generation of a single token.
-    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
+    LLAMA_API int32_t llama_n_threads(struct llama_context * ctx);
 
     // Get the number of threads used for prompt and batch processing (multiple token).
-    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
+    LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
     // Set whether the model is in embeddings mode or not
     // If true, embeddings will be returned but logits will not

src/llama.cpp

@@ -2373,8 +2373,8 @@ struct llama_cparams {
     uint32_t n_batch;
     uint32_t n_ubatch;
     uint32_t n_seq_max;
-    uint32_t n_threads;       // number of threads to use for generation
+    int      n_threads;       // number of threads to use for generation
-    uint32_t n_threads_batch; // number of threads to use for batch processing
+    int      n_threads_batch; // number of threads to use for batch processing
 
     float rope_freq_base;
     float rope_freq_scale;
@@ -3091,6 +3091,9 @@ struct llama_context {
 #endif
     ggml_backend_t backend_cpu = nullptr;
 
+    ggml_threadpool_t threadpool       = nullptr;
+    ggml_threadpool_t threadpool_batch = nullptr;
+
     bool has_evaluated_once = false;
 
     int64_t t_start_us;
@@ -15494,9 +15497,10 @@ static void llama_output_reorder(struct llama_context * ctx) {
 }
 
 static void llama_graph_compute(
-        llama_context & lctx,
+          llama_context & lctx,
-          ggml_cgraph * gf,
+            ggml_cgraph * gf,
-                  int   n_threads) {
+                    int   n_threads,
+        ggml_threadpool * threadpool) {
 #ifdef GGML_USE_METAL
     if (ggml_backend_is_metal(lctx.backend_metal)) {
         ggml_backend_metal_set_n_cb(lctx.backend_metal, n_threads);
@@ -15505,6 +15509,7 @@ static void llama_graph_compute(
 
     if (lctx.backend_cpu != nullptr) {
         ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
+        ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
         ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
     }
 #ifdef GGML_USE_BLAS
@@ -15625,6 +15630,8 @@ static int llama_decode_internal(
         }
 
         int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+        ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch;
+
         GGML_ASSERT(n_threads > 0);
 
         // non-causal masks do not use the KV cache
@@ -15686,7 +15693,7 @@ static int llama_decode_internal(
 
         llama_set_inputs(lctx, ubatch);
 
-        llama_graph_compute(lctx, gf, n_threads);
+        llama_graph_compute(lctx, gf, n_threads, threadpool);
 
         // update the kv ring buffer
         {
@@ -15863,7 +15870,9 @@ static int llama_encode_internal(
     lctx.inp_embd_enc = NULL;
     lctx.n_outputs = n_tokens;
 
-    const int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+    int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+    ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch;
+
     GGML_ASSERT(n_threads > 0);
 
     ggml_backend_sched_reset(lctx.sched);
@@ -15895,7 +15904,7 @@ static int llama_encode_internal(
 
     llama_set_inputs(lctx, ubatch);
 
-    llama_graph_compute(lctx, gf, n_threads);
+    llama_graph_compute(lctx, gf, n_threads, threadpool);
 
     // extract embeddings
     if (embd) {
@@ -16177,7 +16186,7 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
 
     ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids);
 
-    llama_graph_compute(lctx, gf, lctx.cparams.n_threads);
+    llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool);
 #endif
 
     //const int64_t t_end = ggml_time_us();
@@ -16203,7 +16212,7 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) {
 
             llama_set_k_shift(lctx);
 
-            llama_graph_compute(lctx, gf, lctx.cparams.n_threads);
+            llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool);
 
             need_reserve = true;
         }
@@ -17451,6 +17460,19 @@ void llama_numa_init(enum ggml_numa_strategy numa) {
     }
 }
 
+void llama_attach_threadpool(
+             struct llama_context * ctx,
+        ggml_threadpool_t   threadpool,
+        ggml_threadpool_t   threadpool_batch) {
+    ctx->threadpool       = threadpool;
+    ctx->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool;
+}
+
+void llama_detach_threadpool(struct llama_context * ctx) {
+    ctx->threadpool       = nullptr;
+    ctx->threadpool_batch = nullptr;
+}
+
 void llama_backend_free(void) {
     ggml_quantize_free();
 }
@@ -19367,16 +19389,16 @@ size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepa
     }
 }
 
-void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch) {
+void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) {
     ctx->cparams.n_threads       = n_threads;
     ctx->cparams.n_threads_batch = n_threads_batch;
 }
 
-uint32_t llama_n_threads(struct llama_context * ctx) {
+int32_t llama_n_threads(struct llama_context * ctx) {
     return ctx->cparams.n_threads;
 }
 
-uint32_t llama_n_threads_batch(struct llama_context * ctx) {
+int32_t llama_n_threads_batch(struct llama_context * ctx) {
     return ctx->cparams.n_threads_batch;
 }
 

tests/test-rope.cpp

@@ -113,7 +113,7 @@ static struct ggml_tensor * get_random_tensor_f32(
 }
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);