Synchronize with llama.cpp upstream

AMD GPU support on Windows is still in a broken state.

Author: jart

llama.cpp/README.llamafile

@@ -9,8 +9,8 @@ LICENSE
 ORIGIN
 
   https://github.com/ggerganov/llama.cpp/pull/4406/
-  67fac4b95fcccfda8ab965e9ba4992a9ddf3a25f
-  2024-04-10
+  b8109bc0139f15a5b321909f47510b89dca47ffc
+  2024-04-21
 
 LOCAL MODIFICATIONS
 

llama.cpp/common.cpp

@@ -1,16 +1,15 @@
 // -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-
 // vi: set et ft=c++ ts=4 sts=4 sw=4 fenc=utf-8 :vi
+
 #include "common.h"
+#include "json.h"
+#include "json-schema-to-grammar.h"
 #include "llama.h"
-#include "ggml-cuda.h"
-#include "ggml-metal.h"
 
 #include <algorithm>
 #include <cassert>
 #include <cmath>
-#include <cerrno>
 #include <cstring>
-#include <climits>
 #include <ctime>
 #include <fstream>
 #include <iterator>
@@ -74,6 +73,8 @@
 #define LLAMA_CURL_MAX_HEADER_LENGTH 256
 #endif // LLAMA_USE_CURL
 
+using json = nlohmann::ordered_json;
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -110,6 +111,79 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+#include <pthread.h>
+
+static void cpuid(unsigned leaf, unsigned subleaf,
+                  unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
+    __asm__("movq\t%%rbx,%%rsi\n\t"
+            "cpuid\n\t"
+            "xchgq\t%%rbx,%%rsi"
+            : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
+            : "0"(leaf), "2"(subleaf));
+}
+
+static int pin_cpu(int cpu) {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(cpu, &mask);
+    return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
+}
+
+static bool is_hybrid_cpu(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(7, 0, &eax, &ebx, &ecx, &edx);
+    return !!(edx & (1u << 15));
+}
+
+static bool is_running_on_efficiency_core(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
+    int intel_atom = 0x20;
+    int core_type = (eax & 0xff000000u) >> 24;
+    return core_type == intel_atom;
+}
+
+static int count_math_cpus(int cpu_count) {
+    int result = 0;
+    for (int cpu = 0; cpu < cpu_count; ++cpu) {
+        if (pin_cpu(cpu)) {
+            return -1;
+        }
+        if (is_running_on_efficiency_core()) {
+            continue; // efficiency cores harm lockstep threading
+        }
+        ++cpu; // hyperthreading isn't useful for linear algebra
+        ++result;
+    }
+    return result;
+}
+
+#endif // __x86_64__ && __linux__
+
+/**
+ * Returns number of CPUs on system that are useful for math.
+ */
+int get_math_cpu_count() {
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+    int cpu_count = sysconf(_SC_NPROCESSORS_ONLN);
+    if (cpu_count < 1) {
+        return get_num_physical_cores();
+    }
+    if (is_hybrid_cpu()) {
+        cpu_set_t affinity;
+        if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
+            int result = count_math_cpus(cpu_count);
+            pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
+            if (result > 0) {
+                return result;
+            }
+        }
+    }
+#endif
+    return get_num_physical_cores();
+}
+
 void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -1167,6 +1241,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         );
         return true;
     }
+    if (arg == "-j" || arg == "--json-schema") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
+        return true;
+    }
     if (arg == "--override-kv") {
         if (++i >= argc) {
             invalid_param = true;
@@ -1374,6 +1456,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
     printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
     printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  -j SCHEMA, --json-schema SCHEMA\n");
+    printf("                        JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object.\n");
+    printf("                        For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead\n");
     printf("  --cfg-negative-prompt PROMPT\n");
     printf("                        negative prompt to use for guidance. (default: empty)\n");
     printf("  --cfg-negative-prompt-file FNAME\n");
@@ -1766,6 +1851,8 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.defrag_thold      = params.defrag_thold;
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
@@ -2213,7 +2300,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
-    {
+    if (params.warmup) {
         LOG("warming up the model with an empty run\n");
 
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };

llama.cpp/common.h

@@ -5,6 +5,7 @@
 
 #pragma once
 
+#include "llamafile/log.h"
 #include "llama.h"
 
 #include "sampling.h"
@@ -43,6 +44,7 @@ extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
+int get_math_cpu_count();
 int32_t get_num_physical_cores();
 
 //
@@ -52,7 +54,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = llamafile_get_math_cpu_count();
+    int32_t n_threads             = get_math_cpu_count();
     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;
@@ -84,6 +86,9 @@ struct gpt_params {
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    ggml_backend_sched_eval_callback cb_eval = nullptr;
+    void * cb_eval_user_data                 = nullptr;
+
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
     llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
@@ -160,6 +165,7 @@ struct gpt_params {
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
+    bool warmup            = true;  // warmup run
 
     std::string cache_type_k = "f16"; // KV cache data type for the K
     std::string cache_type_v = "f16"; // KV cache data type for the V

llama.cpp/ggml-backend.c

@@ -2,29 +2,22 @@
 // vi: set et ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi
 
 #include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
 
 #include <assert.h>
 #include <limits.h>
 #include <stdarg.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include <stdlib.h>
-
-#include "ggml-alloc.h"
-#include "ggml-cuda.h"
-#include "ggml-impl.h"
-#include "ggml-metal.h"
-#include "llamafile/log.h"
 
 #ifndef NDEBUG
 #define NDEBUG // [jart] delete printf debugging
 #endif
 
-
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
-
 // backend buffer type
 
 const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
@@ -2107,6 +2100,8 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
     return true;
 }
 
+#include "llamafile/log.h"
+
 GGML_CALL static void system_exit(int rc) {
     exit(rc);
 }