whispercpp 1.3.1 → 1.3.2

data/ext/sources/examples/talk-llama/llama-hparams.cpp

@@ -0,0 +1,95 @@
+#include "llama-hparams.h"
+
+#include "ggml.h"
+
+void llama_hparams::set_swa_pattern(uint32_t n_pattern) {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        swa_layers[il] = n_pattern == 0 || (il % n_pattern < (n_pattern - 1));
+    }
+}
+
+bool llama_hparams::is_swa_any() const {
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (swa_layers[il]) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+uint32_t llama_hparams::n_head(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_head_kv(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_kv_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_ff(uint32_t il) const {
+    if (il < n_layer) {
+        return n_ff_arr[il];
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_gqa(uint32_t il) const {
+    const uint32_t n_head    = this->n_head(il);
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    if (n_head_kv == 0) {
+        return 0;
+    }
+
+    return n_head/n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_k * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
+    const uint32_t n_head_kv = this->n_head_kv(il);
+
+    return n_embd_head_v * n_head_kv;
+}
+
+uint32_t llama_hparams::n_embd_k_s() const {
+    if (wkv_head_size != 0) {
+        // for RWKV models
+        return token_shift_count * n_embd;
+    }
+
+    // TODO: maybe support other convolution strides than 1
+    // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+}
+
+uint32_t llama_hparams::n_embd_v_s() const {
+    if (wkv_head_size != 0) {
+        // corresponds to RWKV's wkv_states size
+        return n_embd * wkv_head_size;
+    }
+
+    // corresponds to Mamba's ssm_states size
+    return ssm_d_state * ssm_d_inner;
+}
+
+bool llama_hparams::is_swa(uint32_t il) const {
+    if (il < n_layer) {
+        return swa_layers[il];
+    }
+
+    GGML_ABORT("fatal error");
+}

data/ext/sources/examples/talk-llama/llama-hparams.h

@@ -0,0 +1,190 @@
+#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,
+};
+
+enum llama_swa_type {
+    LLAMA_SWA_TYPE_NONE     = 0,
+    LLAMA_SWA_TYPE_STANDARD = 1,
+    LLAMA_SWA_TYPE_CHUNKED  = 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_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;
+
+    // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
+    uint32_t n_embd_head_k_mla = 0;
+    uint32_t n_embd_head_v_mla = 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;
+    uint32_t moe_every_n_layers   = 0;
+
+    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;
+    uint32_t n_lora_decay           = 0;
+    uint32_t n_lora_iclr            = 0;
+    uint32_t n_lora_value_res_mix   = 0;
+    uint32_t n_lora_gate            = 0;
+
+    float    rope_attn_factor = 1.0f;
+    float    rope_freq_base_train;
+    float    rope_freq_base_train_swa;
+    float    rope_freq_scale_train;
+    float    rope_freq_scale_train_swa;
+    uint32_t n_ctx_orig_yarn;
+    float    rope_yarn_log_mul;
+
+    std::array<int, 4> rope_sections;
+
+    // Sliding Window Attention (SWA)
+    llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+    // the size of the sliding window (0 - no SWA)
+    uint32_t n_swa = 0;
+    // if swa_layers[il] == true, then layer il is SWA
+    // if swa_layers[il] == false, then layer il is dense (i.e. non-SWA)
+    // by default, all layers are dense
+    std::array<bool, LLAMA_MAX_LAYERS> swa_layers;
+
+    // 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;
+    bool use_kq_norm   = true;
+
+    // llama4
+    uint32_t n_moe_layer_step        = 0;
+    uint32_t n_no_rope_layer_step    = 4;
+    uint32_t n_attn_temp_floor_scale = 8192;
+    float    f_attn_temp_scale       = 0.1;
+
+    // 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;
+
+    // this value n_pattern means that every nth layer is dense (i.e. non-SWA)
+    // note that if n_pattern == 0, all layers are SWA
+    //           if n_pattern == 1, all layers are dense
+    // example: n_pattern = 3
+    //   il == 0: swa
+    //   il == 1: swa
+    //   il == 2: dense
+    //   il == 3: swa
+    //   il == 4: swa
+    //   il == 5: dense
+    //   il == 6: swa
+    //   etc ...
+    void set_swa_pattern(uint32_t n_pattern);
+
+    // return true if one of the layers is SWA
+    bool is_swa_any() const;
+
+    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;
+
+    bool is_swa(uint32_t il) const;
+};
+
+static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
+

data/ext/sources/examples/talk-llama/llama-impl.cpp

@@ -0,0 +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 {
+    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 : ggml_time_us()), t_acc(t_acc) {}
+
+time_meas::~time_meas() {
+        if (t_start_us >= 0) {
+            t_acc += ggml_time_us() - t_start_us;
+        }
+    }
+
+void llama_log_set(ggml_log_callback log_callback, void * user_data) {
+    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(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(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(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);
+    GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    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 ggml_tensor * t) {
+    char buf[256];
+    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+    }
+    return buf;
+}
+
+static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
+    switch (type) {
+        case GGUF_TYPE_UINT8:   return std::to_string(((const uint8_t  *)data)[i]);
+        case GGUF_TYPE_INT8:    return std::to_string(((const int8_t   *)data)[i]);
+        case GGUF_TYPE_UINT16:  return std::to_string(((const uint16_t *)data)[i]);
+        case GGUF_TYPE_INT16:   return std::to_string(((const int16_t  *)data)[i]);
+        case GGUF_TYPE_UINT32:  return std::to_string(((const uint32_t *)data)[i]);
+        case GGUF_TYPE_INT32:   return std::to_string(((const int32_t  *)data)[i]);
+        case GGUF_TYPE_UINT64:  return std::to_string(((const uint64_t *)data)[i]);
+        case GGUF_TYPE_INT64:   return std::to_string(((const int64_t  *)data)[i]);
+        case GGUF_TYPE_FLOAT32: return std::to_string(((const float    *)data)[i]);
+        case GGUF_TYPE_FLOAT64: return std::to_string(((const double   *)data)[i]);
+        case GGUF_TYPE_BOOL:    return ((const bool *)data)[i] ? "true" : "false";
+        default:                return format("unknown type %d", type);
+    }
+}
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
+    const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
+
+    switch (type) {
+        case GGUF_TYPE_STRING:
+            return gguf_get_val_str(ctx_gguf, i);
+        case GGUF_TYPE_ARRAY:
+            {
+                const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
+                int arr_n = gguf_get_arr_n(ctx_gguf, i);
+                const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
+                std::stringstream ss;
+                ss << "[";
+                for (int j = 0; j < arr_n; j++) {
+                    if (arr_type == GGUF_TYPE_STRING) {
+                        std::string val = gguf_get_arr_str(ctx_gguf, i, j);
+                        // escape quotes
+                        replace_all(val, "\\", "\\\\");
+                        replace_all(val, "\"", "\\\"");
+                        ss << '"' << val << '"';
+                    } else if (arr_type == GGUF_TYPE_ARRAY) {
+                        ss << "???";
+                    } else {
+                        ss << gguf_data_to_str(arr_type, data, j);
+                    }
+                    if (j < arr_n - 1) {
+                        ss << ", ";
+                    }
+                }
+                ss << "]";
+                return ss.str();
+            }
+        default:
+            return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
+    }
+}

data/ext/sources/examples/talk-llama/llama-impl.h

@@ -0,0 +1,61 @@
+#pragma once
+
+#include "ggml.h" // for ggml_log_level
+
+#include <string>
+#include <vector>
+
+#ifdef __GNUC__
+#    if defined(__MINGW32__) && !defined(__clang__)
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#    else
+#        define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#    endif
+#else
+#    define LLAMA_ATTRIBUTE_FORMAT(...)
+#endif
+
+//
+// logging
+//
+
+LLAMA_ATTRIBUTE_FORMAT(2, 3)
+void llama_log_internal        (ggml_log_level level, const char * format, ...);
+void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
+
+#define LLAMA_LOG(...)       llama_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
+#define LLAMA_LOG_INFO(...)  llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define LLAMA_LOG_WARN(...)  llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define LLAMA_LOG_DEBUG(...) llama_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LLAMA_LOG_CONT(...)  llama_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__)
+
+//
+// helpers
+//
+
+template <typename T>
+struct no_init {
+    T value;
+    no_init() { /* do nothing */ }
+};
+
+struct time_meas {
+    time_meas(int64_t & t_acc, bool disable = false);
+    ~time_meas();
+
+    const int64_t t_start_us;
+
+    int64_t & t_acc;
+};
+
+void replace_all(std::string & s, const std::string & search, const std::string & replace);
+
+// TODO: rename to llama_format ?
+LLAMA_ATTRIBUTE_FORMAT(1, 2)
+std::string format(const char * fmt, ...);
+
+std::string llama_format_tensor_shape(const std::vector<int64_t> & ne);
+std::string llama_format_tensor_shape(const struct ggml_tensor * t);
+
+std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);

data/ext/sources/examples/talk-llama/llama-io.cpp

@@ -0,0 +1,15 @@
+#include "llama-io.h"
+
+void llama_io_write_i::write_string(const std::string & str) {
+    uint32_t str_size = str.size();
+
+    write(&str_size,  sizeof(str_size));
+    write(str.data(), str_size);
+}
+
+void llama_io_read_i::read_string(std::string & str) {
+    uint32_t str_size;
+    read_to(&str_size, sizeof(str_size));
+
+    str.assign((const char *) read(str_size), str_size);
+}

data/ext/sources/examples/talk-llama/llama-io.h

@@ -0,0 +1,35 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+
+struct ggml_tensor;
+
+class llama_io_write_i {
+public:
+    llama_io_write_i() = default;
+    virtual ~llama_io_write_i() = default;
+
+    virtual void write(const void * src, size_t size) = 0;
+    virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0;
+
+    // bytes written so far
+    virtual size_t n_bytes() = 0;
+
+    void write_string(const std::string & str);
+};
+
+class llama_io_read_i {
+public:
+    llama_io_read_i() = default;
+    virtual ~llama_io_read_i() = default;
+
+    virtual const uint8_t * read(size_t size) = 0;
+    virtual void read_to(void * dst, size_t size) = 0;
+
+    // bytes read so far
+    virtual size_t n_bytes() = 0;
+
+    void read_string(std::string & str);
+};