cactus-react-native 0.2.11 → 1.0.1

package/ios/cactus.xcframework/tvos-arm64_x86_64-simulator/cactus.framework/Headers/ggml-impl.h

@@ -1,601 +0,0 @@
-#pragma once
-
-// GGML internal header
-
-#include "ggml.h"
-#include "gguf.h"
-
-#include <assert.h>
-#include <math.h>
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-#include <stdbool.h>
-#include <stdint.h>
-#include <string.h>
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-#endif
-
-#if defined(__F16C__)
-#include <immintrin.h>
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#ifndef MIN
-#    define MIN(a, b) ((a) < (b) ? (a) : (b))
-#endif
-
-#ifndef MAX
-#    define MAX(a, b) ((a) > (b) ? (a) : (b))
-#endif
-
-// required for mmap as gguf only guarantees 32-byte alignment
-#define TENSOR_ALIGNMENT 32
-
-// static_assert should be a #define, but if it's not,
-// fall back to the _Static_assert C11 keyword.
-// if C99 - static_assert is noop
-// ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef __cplusplus
-    #ifndef static_assert
-        #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-            #define static_assert(cond, msg) _Static_assert(cond, msg)
-        #else
-            #define static_assert(cond, msg) struct global_scope_noop_trick
-        #endif
-    #endif
-#endif
-
-static inline int lm_ggml_up32(int n) {
-    return (n + 31) & ~31;
-}
-
-//static inline int lm_ggml_up64(int n) {
-//    return (n + 63) & ~63;
-//}
-
-static inline int lm_ggml_up(int n, int m) {
-    // assert m is a power of 2
-    LM_GGML_ASSERT((m & (m - 1)) == 0);
-    return (n + m - 1) & ~(m - 1);
-}
-
-//
-// logging
-//
-
-LM_GGML_ATTRIBUTE_FORMAT(2, 3)
-LM_GGML_API void lm_ggml_log_internal        (enum lm_ggml_log_level level, const char * format, ...);
-LM_GGML_API void lm_ggml_log_callback_default(enum lm_ggml_log_level level, const char * text, void * user_data);
-
-#define LM_GGML_LOG(...)       lm_ggml_log_internal(LM_GGML_LOG_LEVEL_NONE , __VA_ARGS__)
-#define LM_GGML_LOG_INFO(...)  lm_ggml_log_internal(LM_GGML_LOG_LEVEL_INFO , __VA_ARGS__)
-#define LM_GGML_LOG_WARN(...)  lm_ggml_log_internal(LM_GGML_LOG_LEVEL_WARN , __VA_ARGS__)
-#define LM_GGML_LOG_ERROR(...) lm_ggml_log_internal(LM_GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
-#define LM_GGML_LOG_DEBUG(...) lm_ggml_log_internal(LM_GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
-#define LM_GGML_LOG_CONT(...)  lm_ggml_log_internal(LM_GGML_LOG_LEVEL_CONT , __VA_ARGS__)
-
-#define LM_GGML_DEBUG 0
-
-#if (LM_GGML_DEBUG >= 1)
-#define LM_GGML_PRINT_DEBUG(...) LM_GGML_LOG_DEBUG(__VA_ARGS__)
-#else
-#define LM_GGML_PRINT_DEBUG(...)
-#endif
-
-#if (LM_GGML_DEBUG >= 5)
-#define LM_GGML_PRINT_DEBUG_5(...) LM_GGML_LOG_DEBUG(__VA_ARGS__)
-#else
-#define LM_GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (LM_GGML_DEBUG >= 10)
-#define LM_GGML_PRINT_DEBUG_10(...) LM_GGML_LOG_DEBUG(__VA_ARGS__)
-#else
-#define LM_GGML_PRINT_DEBUG_10(...)
-#endif
-
-// tensor params
-
-static void lm_ggml_set_op_params(struct lm_ggml_tensor * tensor, const void * params, size_t params_size) {
-    LM_GGML_ASSERT(tensor != NULL); // silence -Warray-bounds warnings
-    assert(params_size <= LM_GGML_MAX_OP_PARAMS);
-    memcpy(tensor->op_params, params, params_size);
-}
-
-static int32_t lm_ggml_get_op_params_i32(const struct lm_ggml_tensor * tensor, uint32_t i) {
-    assert(i < LM_GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    return ((const int32_t *)(tensor->op_params))[i];
-}
-
-static float lm_ggml_get_op_params_f32(const struct lm_ggml_tensor * tensor, uint32_t i) {
-    assert(i < LM_GGML_MAX_OP_PARAMS / sizeof(float));
-    return ((const float *)(tensor->op_params))[i];
-}
-
-static void lm_ggml_set_op_params_i32(struct lm_ggml_tensor * tensor, uint32_t i, int32_t value) {
-    assert(i < LM_GGML_MAX_OP_PARAMS / sizeof(int32_t));
-    ((int32_t *)(tensor->op_params))[i] = value;
-}
-
-static void lm_ggml_set_op_params_f32(struct lm_ggml_tensor * tensor, uint32_t i, float value) {
-    assert(i < LM_GGML_MAX_OP_PARAMS / sizeof(float));
-    ((float *)(tensor->op_params))[i] = value;
-}
-
-struct lm_ggml_map_custom1_op_params {
-    lm_ggml_custom1_op_t  fun;
-    int                n_tasks;
-    void             * userdata;
-};
-
-struct lm_ggml_map_custom2_op_params {
-    lm_ggml_custom2_op_t   fun;
-    int                 n_tasks;
-    void              * userdata;
-};
-
-struct lm_ggml_map_custom3_op_params {
-    lm_ggml_custom3_op_t fun;
-    int               n_tasks;
-    void            * userdata;
-};
-
-struct lm_ggml_custom_op_params {
-    lm_ggml_custom_op_t fun;
-    int              n_tasks;
-    void           * userdata;
-};
-
-// bitset
-
-typedef uint32_t lm_ggml_bitset_t;
-
-static_assert(sizeof(lm_ggml_bitset_t) == 4, "bitset_t constants must be updated");
-#define BITSET_SHR 5 // log2(sizeof(lm_ggml_bitset_t)*8)
-#define BITSET_MASK (sizeof(lm_ggml_bitset_t)*8 - 1)
-
-static size_t lm_ggml_bitset_size(size_t n) {
-    return (n + BITSET_MASK) >> BITSET_SHR;
-}
-
-static inline bool lm_ggml_bitset_get(const lm_ggml_bitset_t * bitset, size_t i) {
-    return !!(bitset[i >> BITSET_SHR] & (1u << (i & BITSET_MASK)));
-}
-
-static inline void lm_ggml_bitset_set(lm_ggml_bitset_t * bitset, size_t i) {
-    bitset[i >> BITSET_SHR] |= (1u << (i & BITSET_MASK));
-}
-
-static inline void lm_ggml_bitset_clear(lm_ggml_bitset_t * bitset, size_t i) {
-    bitset[i >> BITSET_SHR] &= ~(1u << (i & BITSET_MASK));
-}
-
-// hash set
-
-#define LM_GGML_HASHSET_FULL ((size_t)-1)
-#define LM_GGML_HASHSET_ALREADY_EXISTS ((size_t)-2)
-
-struct lm_ggml_hash_set {
-    size_t size;
-    lm_ggml_bitset_t * used;       // whether or not the keys are in use i.e. set
-    struct lm_ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if lm_ggml_bitset_get(used, i)
-};
-
-struct lm_ggml_hash_set lm_ggml_hash_set_new(size_t size);
-void                 lm_ggml_hash_set_free(struct lm_ggml_hash_set * hash_set);
-
-// returns the minimum size for a hash set that can hold min_sz elements
-size_t lm_ggml_hash_size(size_t min_sz);
-
-// remove all elements from the hash set
-void lm_ggml_hash_set_reset(struct lm_ggml_hash_set * hash_set);
-
-// returns true if key is in the hash set
-static bool lm_ggml_hash_contains(const struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key);
-
-// returns LM_GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t lm_ggml_hash_find(const struct lm_ggml_hash_set * hash_set, const struct lm_ggml_tensor * key);
-
-// returns LM_GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
-static size_t lm_ggml_hash_insert(struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key);
-
-// return index, asserts if table is full
-static size_t lm_ggml_hash_find_or_insert(struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key);
-
-// hash function for lm_ggml_tensor
-static inline size_t lm_ggml_hash(const struct lm_ggml_tensor * p) {
-    // the last 4 bits are always zero due to alignment
-    return (size_t)(uintptr_t)p >> 4;
-}
-
-static size_t lm_ggml_hash_find(const struct lm_ggml_hash_set * hash_set, const struct lm_ggml_tensor * key) {
-    size_t h = lm_ggml_hash(key) % hash_set->size;
-
-    // linear probing
-    size_t i = h;
-    while (lm_ggml_bitset_get(hash_set->used, i) && hash_set->keys[i] != key) {
-        i = (i + 1) % hash_set->size;
-        if (i == h) {
-            // visited all hash table entries -> not found
-            return LM_GGML_HASHSET_FULL;
-        }
-    }
-    return i;
-}
-
-static bool lm_ggml_hash_contains(const struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key) {
-    size_t i = lm_ggml_hash_find(hash_set, key);
-    return i != LM_GGML_HASHSET_FULL && lm_ggml_bitset_get(hash_set->used, i);
-}
-
-static size_t lm_ggml_hash_insert(struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key) {
-    size_t h = lm_ggml_hash(key) % hash_set->size;
-
-    // linear probing
-    size_t i = h;
-    do {
-        if (!lm_ggml_bitset_get(hash_set->used, i)) {
-            lm_ggml_bitset_set(hash_set->used, i);
-            hash_set->keys[i] = key;
-            return i;
-        }
-        if (hash_set->keys[i] == key) {
-            return LM_GGML_HASHSET_ALREADY_EXISTS;
-        }
-        i = (i + 1) % hash_set->size;
-    } while (i != h);
-
-    // visited all hash table entries -> not found
-    LM_GGML_ABORT("fatal error");
-}
-
-static size_t lm_ggml_hash_find_or_insert(struct lm_ggml_hash_set * hash_set, struct lm_ggml_tensor * key) {
-    size_t h = lm_ggml_hash(key) % hash_set->size;
-
-    // linear probing
-    size_t i = h;
-    do {
-        if (!lm_ggml_bitset_get(hash_set->used, i)) {
-            lm_ggml_bitset_set(hash_set->used, i);
-            hash_set->keys[i] = key;
-            return i;
-        }
-        if (hash_set->keys[i] == key) {
-            return i;
-        }
-        i = (i + 1) % hash_set->size;
-    } while (i != h);
-
-    // visited all hash table entries -> not found
-    LM_GGML_ABORT("fatal error");
-}
-
-// computation graph
-
-enum lm_ggml_cgraph_eval_order {
-    LM_GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-    LM_GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-    LM_GGML_CGRAPH_EVAL_ORDER_COUNT
-};
-
-struct lm_ggml_cgraph {
-    int size;    // maximum number of nodes/leafs/grads/grad_accs
-    int n_nodes; // number of nodes currently in use
-    int n_leafs; // number of leafs currently in use
-
-    struct lm_ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
-    struct lm_ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
-    struct lm_ggml_tensor ** grad_accs; // accumulators for node gradients
-    struct lm_ggml_tensor ** leafs;     // tensors with constant data
-
-    struct lm_ggml_hash_set visited_hash_set;
-
-    enum lm_ggml_cgraph_eval_order order;
-};
-
-// returns a slice of cgraph with nodes [i0, i1)
-// the slice does not have leafs or gradients
-// if you need the gradients, get them from the original graph
-struct lm_ggml_cgraph lm_ggml_graph_view(struct lm_ggml_cgraph * cgraph, int i0, int i1);
-
-// Memory allocation
-
-LM_GGML_API void * lm_ggml_aligned_malloc(size_t size);
-LM_GGML_API void lm_ggml_aligned_free(void * ptr, size_t size);
-
-// FP16 to FP32 conversion
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-//
-// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
-// for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
-//
-#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-    #define LM_GGML_COMPUTE_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
-    #define LM_GGML_COMPUTE_FP32_TO_FP16(x) lm_ggml_compute_fp32_to_fp16(x)
-
-    #define LM_GGML_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
-
-    static inline float lm_ggml_compute_fp16_to_fp32(lm_ggml_fp16_t h) {
-        __fp16 tmp;
-        memcpy(&tmp, &h, sizeof(lm_ggml_fp16_t));
-        return (float)tmp;
-    }
-
-    static inline lm_ggml_fp16_t lm_ggml_compute_fp32_to_fp16(float f) {
-        lm_ggml_fp16_t res;
-        __fp16 tmp = f;
-        memcpy(&res, &tmp, sizeof(lm_ggml_fp16_t));
-        return res;
-    }
-
-#elif defined(__F16C__)
-
-    #ifdef _MSC_VER
-        #define LM_GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-        #define LM_GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-    #else
-        #define LM_GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-        #define LM_GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-    #endif
-
-#elif defined(__POWER9_VECTOR__)
-
-    #define LM_GGML_COMPUTE_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
-    #define LM_GGML_COMPUTE_FP32_TO_FP16(x) lm_ggml_compute_fp32_to_fp16(x)
-    /* the inline asm below is about 12% faster than the lookup method */
-    #define LM_GGML_FP16_TO_FP32(x) LM_GGML_COMPUTE_FP16_TO_FP32(x)
-    #define LM_GGML_FP32_TO_FP16(x) LM_GGML_COMPUTE_FP32_TO_FP16(x)
-
-    static inline float lm_ggml_compute_fp16_to_fp32(lm_ggml_fp16_t h) {
-        float f;
-        double d;
-        __asm__(
-            "mtfprd %0,%2\n"
-            "xscvhpdp %0,%0\n"
-            "frsp %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=f"(f):
-            /* in */   "r"(h));
-        return f;
-    }
-
-    static inline lm_ggml_fp16_t lm_ggml_compute_fp32_to_fp16(float f) {
-        double d;
-        lm_ggml_fp16_t r;
-        __asm__( /* xscvdphp can work on double or single precision */
-            "xscvdphp %0,%2\n"
-            "mffprd %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=r"(r):
-            /* in */   "f"(f));
-        return r;
-    }
-
-#elif defined(__riscv) && defined(LM_GGML_RV_ZFH)
-
-    static inline float lm_ggml_compute_fp16_to_fp32(lm_ggml_fp16_t h) {
-        float f;
-        __asm__(
-            "fmv.h.x %[f], %[h]\n\t"
-            "fcvt.s.h %[f], %[f]"
-            : [f] "=&f" (f)
-            : [h] "r" (h)
-        );
-        return f;
-    }
-
-    static inline lm_ggml_fp16_t lm_ggml_compute_fp32_to_fp16(float f) {
-        lm_ggml_fp16_t res;
-        __asm__(
-            "fcvt.h.s %[f], %[f]\n\t"
-            "fmv.x.h %[h], %[f]"
-            : [h] "=&r" (res)
-            : [f] "f" (f)
-        );
-        return res;
-    }
-
-    #define LM_GGML_COMPUTE_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
-    #define LM_GGML_COMPUTE_FP32_TO_FP16(x) lm_ggml_compute_fp32_to_fp16(x)
-    #define LM_GGML_FP16_TO_FP32(x) LM_GGML_COMPUTE_FP16_TO_FP32(x)
-    #define LM_GGML_FP32_TO_FP16(x) LM_GGML_COMPUTE_FP32_TO_FP16(x)
-
-#else
-
-    // FP16 <-> FP32
-    // ref: https://github.com/Maratyszcza/FP16
-
-    static inline float fp32_from_bits(uint32_t w) {
-        union {
-            uint32_t as_bits;
-            float as_value;
-        } fp32;
-        fp32.as_bits = w;
-        return fp32.as_value;
-    }
-
-    static inline uint32_t fp32_to_bits(float f) {
-        union {
-            float as_value;
-            uint32_t as_bits;
-        } fp32;
-        fp32.as_value = f;
-        return fp32.as_bits;
-    }
-
-    static inline float lm_ggml_compute_fp16_to_fp32(lm_ggml_fp16_t h) {
-        const uint32_t w = (uint32_t) h << 16;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        const uint32_t two_w = w + w;
-
-        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float exp_scale = 0x1.0p-112f;
-    #else
-        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-    #endif
-        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-        const uint32_t magic_mask = UINT32_C(126) << 23;
-        const float magic_bias = 0.5f;
-        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-        const uint32_t result = sign |
-            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-        return fp32_from_bits(result);
-    }
-
-    static inline lm_ggml_fp16_t lm_ggml_compute_fp32_to_fp16(float f) {
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float scale_to_inf = 0x1.0p+112f;
-        const float scale_to_zero = 0x1.0p-110f;
-    #else
-        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-    #endif
-        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-        const uint32_t w = fp32_to_bits(f);
-        const uint32_t shl1_w = w + w;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-        if (bias < UINT32_C(0x71000000)) {
-            bias = UINT32_C(0x71000000);
-        }
-
-        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-        const uint32_t bits = fp32_to_bits(base);
-        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-        const uint32_t nonsign = exp_bits + mantissa_bits;
-        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-    }
-
-    #define LM_GGML_COMPUTE_FP16_TO_FP32(x) lm_ggml_compute_fp16_to_fp32(x)
-    #define LM_GGML_COMPUTE_FP32_TO_FP16(x) lm_ggml_compute_fp32_to_fp16(x)
-
-#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in lm_ggml_init()
-LM_GGML_API float lm_ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into lm_ggml_lookup_fp16_to_fp32,
-// so we define LM_GGML_FP16_TO_FP32 and LM_GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(LM_GGML_FP16_TO_FP32)
-inline static float lm_ggml_lookup_fp16_to_fp32(lm_ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return lm_ggml_table_f32_f16[s];
-}
-
-#define LM_GGML_FP16_TO_FP32(x) lm_ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(LM_GGML_FP32_TO_FP16)
-#define LM_GGML_FP32_TO_FP16(x) LM_GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float lm_ggml_compute_bf16_to_fp32(lm_ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline lm_ggml_bf16_t lm_ggml_compute_fp32_to_bf16(float s) {
-    lm_ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define LM_GGML_FP32_TO_BF16(x) lm_ggml_compute_fp32_to_bf16(x)
-#define LM_GGML_BF16_TO_FP32(x) lm_ggml_compute_bf16_to_fp32(x)
-
-#ifdef __cplusplus
-}
-#endif
-
-#ifdef __cplusplus
-#include <vector>
-
-// expose GGUF internals for test code
-LM_GGML_API size_t lm_gguf_type_size(enum lm_gguf_type type);
-LM_GGML_API struct lm_gguf_context * lm_gguf_init_from_file_impl(FILE * file, struct lm_gguf_init_params params);
-LM_GGML_API void lm_gguf_write_to_buf(const struct lm_gguf_context * ctx, std::vector<int8_t> & buf, bool only_meta);
-#endif // __cplusplus