cui-llama.rn 1.4.0 → 1.4.2

package/cpp/llama-hparams.h

@@ -1,140 +1,139 @@
-#pragma once
-
-#include "llama.h"
-
-#include <array>
-
-// bump if necessary
-#define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
-
-enum llama_expert_gating_func_type {
-    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,
-    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
-};
-
-struct llama_hparams_posnet {
-    uint32_t n_embd;
-    uint32_t n_layer;
-};
-
-struct llama_hparams_convnext {
-    uint32_t n_embd;
-    uint32_t n_layer;
-};
-
-struct llama_hparams {
-    bool vocab_only;
-    bool rope_finetuned;
-    bool use_par_res;
-    bool swin_norm;
-
-    uint32_t n_vocab = 0;
-    uint32_t n_ctx_train; // context size the model was trained on
-    uint32_t n_embd;
-    uint32_t n_embd_features = 0;
-    uint32_t n_layer;
-    uint32_t n_rot;
-    uint32_t n_swa = 0; // sliding window attention (SWA)
-    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
-    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
-    uint32_t n_expert = 0;
-    uint32_t n_expert_used = 0;
-    uint32_t n_vocab_type = 0; // for BERT-style token types
-    uint32_t n_rel_attn_bkts = 0;
-
-    // for WavTokenizer
-    struct llama_hparams_posnet   posnet;
-    struct llama_hparams_convnext convnext;
-
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
-    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
-
-    uint32_t n_layer_dense_lead = 0;
-    uint32_t n_lora_q           = 0;
-    uint32_t n_lora_kv          = 0;
-    uint32_t n_ff_exp           = 0;
-    uint32_t n_ff_shexp         = 0;
-    uint32_t n_expert_shared    = 0;
-    uint32_t n_norm_groups      = 0;
-
-    float    expert_weights_scale = 0.0;
-    bool     expert_weights_norm  = false;
-    uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
-
-    float f_norm_eps;
-    float f_norm_rms_eps;
-    float f_norm_group_eps;
-
-    float f_attn_logit_softcapping  = 50.0f;
-    float f_final_logit_softcapping = 30.0f;
-
-    // for RWKV
-    uint32_t rescale_every_n_layers = 0;
-    uint32_t time_mix_extra_dim     = 0;
-    uint32_t time_decay_extra_dim   = 0;
-    uint32_t wkv_head_size          = 0;
-
-    float    rope_attn_factor = 1.0f;
-    float    rope_freq_base_train;
-    float    rope_freq_scale_train;
-    uint32_t n_ctx_orig_yarn;
-    float    rope_yarn_log_mul;
-
-    std::array<int, 4> rope_sections;
-
-    // for State Space Models
-    uint32_t ssm_d_conv  = 0;
-    uint32_t ssm_d_inner = 0;
-    uint32_t ssm_d_state = 0;
-    uint32_t ssm_dt_rank = 0;
-
-    bool ssm_dt_b_c_rms = false;
-
-    float f_clamp_kqv      = 0.0f;
-    float f_max_alibi_bias = 0.0f;
-    float f_logit_scale    = 0.0f;
-
-    // Additional scale factors (Granite/Granite MoE)
-    float f_residual_scale  = 0.0f;
-    float f_embedding_scale = 0.0f;
-    float f_attention_scale = 0.0f;
-
-    bool causal_attn   = true;
-    bool use_alibi     = false;
-    bool attn_soft_cap = false;
-
-    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
-    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
-
-    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
-    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
-    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
-
-    uint32_t n_head(uint32_t il = 0) const;
-
-    uint32_t n_head_kv(uint32_t il = 0) const;
-
-    uint32_t n_ff(uint32_t il = 0) const;
-
-    uint32_t n_gqa(uint32_t il = 0) const;
-
-    // dimension of key embeddings across all k-v heads
-    uint32_t n_embd_k_gqa(uint32_t il = 0) const;
-
-    // dimension of value embeddings across all k-v heads
-    uint32_t n_embd_v_gqa(uint32_t il = 0) const;
-
-    // dimension of the rolling state embeddings
-    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
-
-    // dimension of the recurrent state embeddings
-    uint32_t n_embd_v_s() const;
-};
-
-static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
-
+#pragma once
+
+#include "llama.h"
+
+#include <array>
+
+// bump if necessary
+#define LLAMA_MAX_LAYERS  512
+#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+
+enum llama_expert_gating_func_type {
+    LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX = 1,
+    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID = 2,
+};
+
+struct llama_hparams_posnet {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams_convnext {
+    uint32_t n_embd;
+    uint32_t n_layer;
+};
+
+struct llama_hparams {
+    bool vocab_only;
+    bool rope_finetuned;
+    bool use_par_res;
+    bool swin_norm;
+
+    uint32_t n_ctx_train; // context size the model was trained on
+    uint32_t n_embd;
+    uint32_t n_embd_features = 0;
+    uint32_t n_layer;
+    uint32_t n_rot;
+    uint32_t n_swa = 0; // sliding window attention (SWA)
+    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_expert = 0;
+    uint32_t n_expert_used = 0;
+    uint32_t n_rel_attn_bkts = 0;
+
+    // for WavTokenizer
+    struct llama_hparams_posnet   posnet;
+    struct llama_hparams_convnext convnext;
+
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
+    std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
+
+    uint32_t n_layer_dense_lead = 0;
+    uint32_t n_lora_q           = 0;
+    uint32_t n_lora_kv          = 0;
+    uint32_t n_ff_exp           = 0;
+    uint32_t n_ff_shexp         = 0;
+    uint32_t n_expert_shared    = 0;
+    uint32_t n_norm_groups      = 0;
+
+    float    expert_weights_scale = 0.0;
+    bool     expert_weights_norm  = false;
+    uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
+
+    float f_norm_eps;
+    float f_norm_rms_eps;
+    float f_norm_group_eps;
+
+    float f_attn_logit_softcapping  = 50.0f;
+    float f_final_logit_softcapping = 30.0f;
+
+    // for RWKV
+    uint32_t rescale_every_n_layers = 0;
+    uint32_t time_mix_extra_dim     = 0;
+    uint32_t time_decay_extra_dim   = 0;
+    uint32_t wkv_head_size          = 0;
+    uint32_t token_shift_count      = 2;
+
+    float    rope_attn_factor = 1.0f;
+    float    rope_freq_base_train;
+    float    rope_freq_scale_train;
+    uint32_t n_ctx_orig_yarn;
+    float    rope_yarn_log_mul;
+
+    std::array<int, 4> rope_sections;
+
+    // for State Space Models
+    uint32_t ssm_d_conv  = 0;
+    uint32_t ssm_d_inner = 0;
+    uint32_t ssm_d_state = 0;
+    uint32_t ssm_dt_rank = 0;
+
+    bool ssm_dt_b_c_rms = false;
+
+    float f_clamp_kqv      = 0.0f;
+    float f_max_alibi_bias = 0.0f;
+    float f_logit_scale    = 0.0f;
+
+    // Additional scale factors (Granite/Granite MoE)
+    float f_residual_scale  = 0.0f;
+    float f_embedding_scale = 0.0f;
+    float f_attention_scale = 0.0f;
+
+    bool causal_attn   = true;
+    bool use_alibi     = false;
+    bool attn_soft_cap = false;
+
+    // needed by encoder-decoder models (e.g. T5, FLAN-T5)
+    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
+    llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
+
+    enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
+    enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
+    enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
+
+    uint32_t n_head(uint32_t il = 0) const;
+
+    uint32_t n_head_kv(uint32_t il = 0) const;
+
+    uint32_t n_ff(uint32_t il = 0) const;
+
+    uint32_t n_gqa(uint32_t il = 0) const;
+
+    // dimension of key embeddings across all k-v heads
+    uint32_t n_embd_k_gqa(uint32_t il = 0) const;
+
+    // dimension of value embeddings across all k-v heads
+    uint32_t n_embd_v_gqa(uint32_t il = 0) const;
+
+    // dimension of the rolling state embeddings
+    // corresponds to Mamba's conv_states size or RWKV's token_shift states size
+    uint32_t n_embd_k_s() const;
+
+    // dimension of the recurrent state embeddings
+    uint32_t n_embd_v_s() const;
+};
+
+static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+

package/cpp/llama-impl.cpp

@@ -1,167 +1,167 @@
-#include "llama-impl.h"
-
-#include "gguf.h"
-#include "llama.h"
-
-#include <cinttypes>
-#include <climits>
-#include <cstdarg>
-#include <cstring>
-#include <vector>
-#include <sstream>
-
-struct llama_logger_state {
-    lm_ggml_log_callback log_callback = llama_log_callback_default;
-    void * log_callback_user_data = nullptr;
-};
-
-static llama_logger_state g_logger_state;
-
-time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : lm_ggml_time_us()), t_acc(t_acc) {}
-
-time_meas::~time_meas() {
-        if (t_start_us >= 0) {
-            t_acc += lm_ggml_time_us() - t_start_us;
-        }
-    }
-
-void llama_log_set(lm_ggml_log_callback log_callback, void * user_data) {
-    lm_ggml_log_set(log_callback, user_data);
-    g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
-    g_logger_state.log_callback_user_data = user_data;
-}
-
-static void llama_log_internal_v(lm_ggml_log_level level, const char * format, va_list args) {
-    va_list args_copy;
-    va_copy(args_copy, args);
-    char buffer[128];
-    int len = vsnprintf(buffer, 128, format, args);
-    if (len < 128) {
-        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
-    } else {
-        char * buffer2 = new char[len + 1];
-        vsnprintf(buffer2, len + 1, format, args_copy);
-        buffer2[len] = 0;
-        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
-        delete[] buffer2;
-    }
-    va_end(args_copy);
-}
-
-void llama_log_internal(lm_ggml_log_level level, const char * format, ...) {
-    va_list args;
-    va_start(args, format);
-    llama_log_internal_v(level, format, args);
-    va_end(args);
-}
-
-void llama_log_callback_default(lm_ggml_log_level level, const char * text, void * user_data) {
-    (void) level;
-    (void) user_data;
-    fputs(text, stderr);
-    fflush(stderr);
-}
-
-void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    if (search.empty()) {
-        return;
-    }
-    std::string builder;
-    builder.reserve(s.length());
-    size_t pos = 0;
-    size_t last_pos = 0;
-    while ((pos = s.find(search, last_pos)) != std::string::npos) {
-        builder.append(s, last_pos, pos - last_pos);
-        builder.append(replace);
-        last_pos = pos + search.length();
-    }
-    builder.append(s, last_pos, std::string::npos);
-    s = std::move(builder);
-}
-
-std::string format(const char * fmt, ...) {
-    va_list ap;
-    va_list ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    LM_GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    LM_GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
-std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
-    char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
-    for (size_t i = 1; i < ne.size(); i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
-    }
-    return buf;
-}
-
-std::string llama_format_tensor_shape(const struct lm_ggml_tensor * t) {
-    char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
-    for (int i = 1; i < LM_GGML_MAX_DIMS; i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
-    }
-    return buf;
-}
-
-static std::string lm_gguf_data_to_str(enum lm_gguf_type type, const void * data, int i) {
-    switch (type) {
-        case LM_GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
-        case LM_GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
-        case LM_GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
-        case LM_GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
-        case LM_GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
-        case LM_GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
-        case LM_GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
-        case LM_GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
-        case LM_GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
-        case LM_GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
-        case LM_GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
-        default:                return format("unknown type %d", type);
-    }
-}
-
-std::string lm_gguf_kv_to_str(const struct lm_gguf_context * ctx_gguf, int i) {
-    const enum lm_gguf_type type = lm_gguf_get_kv_type(ctx_gguf, i);
-
-    switch (type) {
-        case LM_GGUF_TYPE_STRING:
-            return lm_gguf_get_val_str(ctx_gguf, i);
-        case LM_GGUF_TYPE_ARRAY:
-            {
-                const enum lm_gguf_type arr_type = lm_gguf_get_arr_type(ctx_gguf, i);
-                int arr_n = lm_gguf_get_arr_n(ctx_gguf, i);
-                const void * data = arr_type == LM_GGUF_TYPE_STRING ? nullptr : lm_gguf_get_arr_data(ctx_gguf, i);
-                std::stringstream ss;
-                ss << "[";
-                for (int j = 0; j < arr_n; j++) {
-                    if (arr_type == LM_GGUF_TYPE_STRING) {
-                        std::string val = lm_gguf_get_arr_str(ctx_gguf, i, j);
-                        // escape quotes
-                        replace_all(val, "\\", "\\\\");
-                        replace_all(val, "\"", "\\\"");
-                        ss << '"' << val << '"';
-                    } else if (arr_type == LM_GGUF_TYPE_ARRAY) {
-                        ss << "???";
-                    } else {
-                        ss << lm_gguf_data_to_str(arr_type, data, j);
-                    }
-                    if (j < arr_n - 1) {
-                        ss << ", ";
-                    }
-                }
-                ss << "]";
-                return ss.str();
-            }
-        default:
-            return lm_gguf_data_to_str(type, lm_gguf_get_val_data(ctx_gguf, i), 0);
-    }
-}
+#include "llama-impl.h"
+
+#include "gguf.h"
+#include "llama.h"
+
+#include <cinttypes>
+#include <climits>
+#include <cstdarg>
+#include <cstring>
+#include <vector>
+#include <sstream>
+
+struct llama_logger_state {
+    lm_ggml_log_callback log_callback = llama_log_callback_default;
+    void * log_callback_user_data = nullptr;
+};
+
+static llama_logger_state g_logger_state;
+
+time_meas::time_meas(int64_t & t_acc, bool disable) : t_start_us(disable ? -1 : lm_ggml_time_us()), t_acc(t_acc) {}
+
+time_meas::~time_meas() {
+        if (t_start_us >= 0) {
+            t_acc += lm_ggml_time_us() - t_start_us;
+        }
+    }
+
+void llama_log_set(lm_ggml_log_callback log_callback, void * user_data) {
+    lm_ggml_log_set(log_callback, user_data);
+    g_logger_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
+}
+
+static void llama_log_internal_v(lm_ggml_log_level level, const char * format, va_list args) {
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = new char[len + 1];
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        delete[] buffer2;
+    }
+    va_end(args_copy);
+}
+
+void llama_log_internal(lm_ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    llama_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+void llama_log_callback_default(lm_ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    if (search.empty()) {
+        return;
+    }
+    std::string builder;
+    builder.reserve(s.length());
+    size_t pos = 0;
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
+    }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
+}
+
+std::string format(const char * fmt, ...) {
+    va_list ap;
+    va_list ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LM_GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LM_GGML_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
+    for (size_t i = 1; i < ne.size(); i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
+    }
+    return buf;
+}
+
+std::string llama_format_tensor_shape(const struct lm_ggml_tensor * t) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    for (int i = 1; i < LM_GGML_MAX_DIMS; i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+    }
+    return buf;
+}
+
+static std::string lm_gguf_data_to_str(enum lm_gguf_type type, const void * data, int i) {
+    switch (type) {
+        case LM_GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case LM_GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case LM_GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case LM_GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case LM_GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case LM_GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case LM_GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case LM_GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case LM_GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case LM_GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case LM_GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+std::string lm_gguf_kv_to_str(const struct lm_gguf_context * ctx_gguf, int i) {
+    const enum lm_gguf_type type = lm_gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case LM_GGUF_TYPE_STRING:
+            return lm_gguf_get_val_str(ctx_gguf, i);
+        case LM_GGUF_TYPE_ARRAY:
+            {
+                const enum lm_gguf_type arr_type = lm_gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = lm_gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == LM_GGUF_TYPE_STRING ? nullptr : lm_gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == LM_GGUF_TYPE_STRING) {
+                        std::string val = lm_gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == LM_GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << lm_gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return lm_gguf_data_to_str(type, lm_gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}