cui-llama.rn 1.7.3 → 1.7.6

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/common.h

@@ -126,7 +126,7 @@ enum common_grammar_trigger_type {
     COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN,
     COMMON_GRAMMAR_TRIGGER_TYPE_WORD,
     COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
-    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+    COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
 };
 
 struct common_grammar_trigger {
@@ -210,6 +210,9 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
+    lm_ggml_type cache_type_k = LM_GGML_TYPE_F16; // KV cache data type for the K
+    lm_ggml_type cache_type_v = LM_GGML_TYPE_F16; // KV cache data type for the V
+
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -226,7 +229,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {
@@ -306,6 +310,7 @@ struct common_params {
     int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
+    bool    offline                    = false;
 
     int32_t ppl_stride      = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int32_t ppl_output_type = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
@@ -368,7 +373,7 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    bool reranking         = false; // enable reranking support on server
+    std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
@@ -383,6 +388,7 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/ggml-common.h

@@ -1074,6 +1074,10 @@ LM_GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 LM_GGML_TABLE_END()
 
+LM_GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+LM_GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/ggml-cpu.h

@@ -101,6 +101,7 @@ extern "C" {
     LM_GGML_BACKEND_API int lm_ggml_cpu_has_riscv_v    (void);
     LM_GGML_BACKEND_API int lm_ggml_cpu_has_vsx        (void);
     LM_GGML_BACKEND_API int lm_ggml_cpu_has_vxe        (void);
+    LM_GGML_BACKEND_API int lm_ggml_cpu_has_nnpa       (void);
     LM_GGML_BACKEND_API int lm_ggml_cpu_has_wasm_simd  (void);
     LM_GGML_BACKEND_API int lm_ggml_cpu_has_llamafile  (void);
 

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

@@ -32,6 +32,8 @@
 extern "C" {
 #endif
 
+void lm_ggml_print_backtrace(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif
@@ -315,203 +317,81 @@ struct lm_ggml_cgraph lm_ggml_graph_view(struct lm_ggml_cgraph * cgraph, int i0,
 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)
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
 
-    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 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 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;
-    }
+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;
+}
 
-    #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)
+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;
 
-    // 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);
-    }
+    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;
 
-    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);
-        }
+    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);
+}
 
-        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);
+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);
     }
 
-    #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];
+    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_FP16_TO_FP32(x) lm_ggml_lookup_fp16_to_fp32(x)
-#endif
+#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)
 
-#if !defined(LM_GGML_FP32_TO_FP16)
+#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)
-#endif
 
 /**
  * Converts brain16 to float32.

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/ggml.h

@@ -490,6 +490,7 @@ extern "C" {
         LM_GGML_OP_UPSCALE, // nearest interpolate
         LM_GGML_OP_PAD,
         LM_GGML_OP_PAD_REFLECT_1D,
+        LM_GGML_OP_ROLL,
         LM_GGML_OP_ARANGE,
         LM_GGML_OP_TIMESTEP_EMBEDDING,
         LM_GGML_OP_ARGSORT,
@@ -936,6 +937,15 @@ extern "C" {
             struct lm_ggml_tensor  * a,
             struct lm_ggml_tensor  * b);
 
+    // repeat a to the specified shape
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_repeat_4d(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+                       int64_t    ne0,
+                       int64_t    ne1,
+                       int64_t    ne2,
+                       int64_t    ne3);
+
     // sums repetitions in a into shape of b
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_repeat_back(
             struct lm_ggml_context * ctx,
@@ -1793,6 +1803,17 @@ extern "C" {
             int                   p0,
             int                   p1);
 
+    // Move tensor elements by an offset given for each dimension. Elements that
+    // are shifted beyond the last position are wrapped around to the beginning.
+    LM_GGML_API struct lm_ggml_tensor * lm_ggml_roll(
+            struct lm_ggml_context * ctx,
+            struct lm_ggml_tensor  * a,
+            int                   shift0,
+            int                   shift1,
+            int                   shift2,
+            int                   shift3);
+
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]
@@ -2087,9 +2108,6 @@ extern "C" {
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_graph_get_grad    (const struct lm_ggml_cgraph * cgraph, const struct lm_ggml_tensor * node);
     LM_GGML_API struct lm_ggml_tensor * lm_ggml_graph_get_grad_acc(const struct lm_ggml_cgraph * cgraph, const struct lm_ggml_tensor * node);
 
-    LM_GGML_API void                 lm_ggml_graph_export(const struct lm_ggml_cgraph * cgraph, const char * fname);
-    LM_GGML_API struct lm_ggml_cgraph * lm_ggml_graph_import(const char * fname, struct lm_ggml_context ** ctx_data, struct lm_ggml_context ** ctx_eval);
-
     // print info and performance information for the graph
     LM_GGML_API void lm_ggml_graph_print(const struct lm_ggml_cgraph * cgraph);
 
@@ -2173,6 +2191,7 @@ extern "C" {
 
     // scheduling priorities
     enum lm_ggml_sched_priority {
+        LM_GGML_SCHED_PRIO_LOW = -1,
         LM_GGML_SCHED_PRIO_NORMAL,
         LM_GGML_SCHED_PRIO_MEDIUM,
         LM_GGML_SCHED_PRIO_HIGH,

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/json-partial.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#include "nlohmann/json.hpp"
+
+// Healing marker (empty if the JSON was fully parsed / wasn't healed).
+struct common_healing_marker {
+    // Raw marker.
+    std::string marker;
+
+    // Cutting the `common_json.json.dump()` string at the (only) occurrence of this marker should yield the original partial JSON string (modulo spaces / if it had the same dump format).
+    std::string json_dump_marker;
+};
+
+// Represents a parsed JSON object, with its optional healing marker (a JSON dump fragment that can be used to find the position of healing in the JSON dump string)
+struct common_json {
+    nlohmann::ordered_json json;
+
+    common_healing_marker healing_marker;
+};
+
+// Parse the JSON string, healing (closing) any partial JSON if `healing_marker` is not empty.
+//
+// Healing completes partial JSON strings by adding a (possibly modified) healing marker, then whatever is needed to close the JSON.
+// This allows to parse the resulting healed JSON string, yet be able to cut it again if needed at the healing marker.
+// (this is used when parsing JSON outputs from the models, then crafting partial JSONs for the partial tool calls in OAI format).
+//
+// For instance, parsing `{` with a healing marker `foo` will produce a healed JSON `{"foo":1}`, w/ json_dump_marker = `"foo"` (which can be used to break the JSON again).
+bool common_json_parse(
+    const std::string & input,
+    const std::string & healing_marker,
+    common_json & out);
+
+// Parse the JSON string (see overload above), but advancing an iterator to the end of the input when the (potentially partial) parsing succeeds.
+bool common_json_parse(
+    std::string::const_iterator & it,
+    const std::string::const_iterator & end,
+    const std::string & healing_marker,
+    common_json & out);

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/json-schema-to-grammar.h

@@ -1,9 +1,9 @@
 #pragma once
 
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to LM_GGML_ASSERT:
-#define JSON_ASSERT LM_GGML_ASSERT
-#include "json.hpp"
+#include "nlohmann/json_fwd.hpp"
+
+#include <functional>
+#include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);

package/ios/rnllama.xcframework/ios-arm64_x86_64-simulator/rnllama.framework/Headers/llama-arch.h

@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
+    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
@@ -45,6 +46,7 @@ enum llm_arch {
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
+    LLM_ARCH_GEMMA3N,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
     LLM_ARCH_XVERSE,
@@ -76,6 +78,8 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -148,6 +152,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -190,13 +195,13 @@ enum llm_kv {
     LLM_KV_TOKENIZER_MASK_ID,
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_SEP,
     LLM_KV_TOKENIZER_ADD_PREFIX,
     LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
-    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -213,6 +218,8 @@ enum llm_kv {
     LLM_KV_CONVNEXT_EMBEDDING_LENGTH,
     LLM_KV_CONVNEXT_BLOCK_COUNT,
 
+    LLM_KV_CLASSIFIER_OUTPUT_LABELS,
+
     // deprecated:
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
@@ -263,6 +270,22 @@ enum llm_tensor {
     LLM_TENSOR_LAYER_OUT_NORM,
     LLM_TENSOR_POST_ATTN_NORM,
     LLM_TENSOR_POST_MLP_NORM,
+    LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
+    LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n
+    LLM_TENSOR_PER_LAYER_INP_GATE,   // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ,       // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ_NORM,  // gemma3n
+    LLM_TENSOR_PER_LAYER_POST_NORM,  // gemma3n
+    LLM_TENSOR_ALTUP_PROJ,           // gemma3n
+    LLM_TENSOR_ALTUP_UNEMBD_PROJ,    // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_SCALE,  // gemma3n
+    LLM_TENSOR_ALTUP_PREDICT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER,         // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER_NORM,    // gemma3n
+    LLM_TENSOR_LAUREL_L,             // gemma3n
+    LLM_TENSOR_LAUREL_R,             // gemma3n
+    LLM_TENSOR_LAUREL_POST_NORM,     // gemma3n
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
@@ -435,3 +458,6 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
+
+bool llm_arch_is_recurrent(const llm_arch & arch);
+bool llm_arch_is_hybrid   (const llm_arch & arch);