cui-llama.rn 1.3.5 → 1.3.6

package/cpp/llama-context.h

@@ -0,0 +1,128 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-model.h"
+#include "llama-kv-cache.h"
+#include "llama-adapter.h"
+
+#include "ggml-cpp.h"
+
+#include <map>
+#include <unordered_map>
+#include <vector>
+#include <set>
+
+struct llama_context {
+    llama_context(const llama_model & model)
+        : model(model)
+        , t_start_us(model.t_start_us)
+        , t_load_us(model.t_load_us) {}
+
+    const struct llama_model & model;
+
+    struct llama_cparams        cparams;
+    struct llama_sbatch         sbatch;  // TODO: revisit if needed
+    struct llama_kv_cache       kv_self;
+    struct llama_control_vector cvec;
+
+    std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
+
+    std::vector<lm_ggml_backend_ptr> backends;
+    std::vector<std::pair<lm_ggml_backend_t, lm_ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
+    lm_ggml_backend_t backend_cpu = nullptr;
+
+    lm_ggml_threadpool_t threadpool       = nullptr;
+    lm_ggml_threadpool_t threadpool_batch = nullptr;
+
+    bool has_evaluated_once = false;
+
+    mutable int64_t t_start_us;
+    mutable int64_t t_load_us;
+    mutable int64_t t_p_eval_us = 0;
+    mutable int64_t t_eval_us   = 0;
+
+    mutable int64_t t_compute_start_us = 0;
+    mutable int64_t n_queued_tokens = 0;
+
+    mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
+    mutable int32_t n_eval   = 0; // number of eval calls
+
+    // host buffer for the model output (logits and embeddings)
+    lm_ggml_backend_buffer_ptr buf_output;
+
+    // decode output (2-dimensional array: [n_outputs][n_vocab])
+    size_t  logits_size = 0; // capacity (of floats) for logits
+    float * logits      = nullptr;
+
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
+    size_t  output_size = 0; // capacity (of tokens positions) for the output buffers
+    int32_t n_outputs   = 0; // number of actually-used outputs in the current ubatch or last logical batch
+
+    bool logits_all = false;
+
+    // embeddings output (2-dimensional array: [n_outputs][n_embd])
+    // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE
+    size_t  embd_size = 0; // capacity (of floats) for embeddings
+    float * embd      = nullptr;
+
+    // sequence embeddings output (map of [n_embd] vectors)
+    // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
+    std::map<llama_seq_id, std::vector<float>> embd_seq;
+
+    // whether we are computing encoder output or decoder output
+    bool is_encoding = false;
+
+    // TODO: find a better way to accommodate mutli-dimension position encoding methods
+    // number of position id each token get, 1 for each token in most cases.
+    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
+    int n_pos_per_token = 1;
+
+    // output of the encoder part of the encoder-decoder models
+    std::vector<float> embd_enc;
+    std::vector<std::set<llama_seq_id>> seq_ids_enc;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+    lm_ggml_backend_sched_ptr sched;
+
+    lm_ggml_abort_callback abort_callback      = nullptr;
+    void *              abort_callback_data = nullptr;
+
+    // input tensors
+    struct lm_ggml_tensor * inp_tokens;        // I32 [n_batch]
+    struct lm_ggml_tensor * inp_embd;          // F32 [n_embd, n_batch]
+    struct lm_ggml_tensor * inp_pos;           // I32 [n_batch]
+    struct lm_ggml_tensor * inp_out_ids;       // I32 [n_outputs]
+    struct lm_ggml_tensor * inp_KQ_mask;       // F32 [kv_size, n_batch]
+    struct lm_ggml_tensor * inp_KQ_mask_swa;   // F32 [kv_size, n_batch]
+    struct lm_ggml_tensor * inp_K_shift;       // I32 [kv_size]
+    struct lm_ggml_tensor * inp_mean;          // F32 [n_batch, n_batch]
+    struct lm_ggml_tensor * inp_cls;           // I32 [n_batch]
+    struct lm_ggml_tensor * inp_s_copy;        // I32 [kv_size]
+    struct lm_ggml_tensor * inp_s_mask;        // F32 [1, n_kv]
+    struct lm_ggml_tensor * inp_s_seq;         // I32 [n_kv, n_batch]
+    struct lm_ggml_tensor * inp_pos_bucket;    // I32 [n_batch|n_kv, n_batch]
+    struct lm_ggml_tensor * inp_embd_enc;      // F32 [n_embd, n_outputs_enc]
+    struct lm_ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
+};
+
+// TODO: make these methods of llama_context
+void llama_set_k_shift(struct llama_context & lctx);
+
+void llama_set_s_copy(struct llama_context & lctx);
+
+void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch);
+
+// Make sure enough space is available for outputs.
+// Returns max number of outputs for which space was reserved.
+size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs);
+
+// make the outputs have the same order they had in the user-provided batch
+void llama_output_reorder(struct llama_context & ctx);
+
+// For internal test use
+// TODO: remove
+const std::vector<std::pair<std::string, struct lm_ggml_tensor *>> & llama_internal_get_tensor_map(struct llama_context * ctx);

package/cpp/llama-cparams.cpp

@@ -0,0 +1 @@
+#include "llama-cparams.h"

package/cpp/llama-cparams.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+
+struct llama_cparams {
+    uint32_t n_ctx;           // context size used during inference
+    uint32_t n_batch;
+    uint32_t n_ubatch;
+    uint32_t n_seq_max;
+    int      n_threads;       // number of threads to use for generation
+    int      n_threads_batch; // number of threads to use for batch processing
+
+    float rope_freq_base;
+    float rope_freq_scale;
+
+    uint32_t n_ctx_orig_yarn;
+    // These hyperparameters are not exposed in GGUF, because all
+    // existing YaRN models use the same values for them.
+    float yarn_ext_factor;
+    float yarn_attn_factor;
+    float yarn_beta_fast;
+    float yarn_beta_slow;
+    float defrag_thold;
+
+    bool embeddings;
+    bool causal_attn;
+    bool offload_kqv;
+    bool flash_attn;
+    bool no_perf;
+
+    enum llama_pooling_type pooling_type;
+
+    lm_ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};

package/cpp/llama-cpp.h

@@ -0,0 +1,30 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include <memory>
+
+#include "llama.h"
+
+struct llama_model_deleter {
+    void operator()(llama_model * model) { llama_model_free(model); }
+};
+
+struct llama_context_deleter {
+    void operator()(llama_context * context) { llama_free(context); }
+};
+
+struct llama_sampler_deleter {
+    void operator()(llama_sampler * sampler) { llama_sampler_free(sampler); }
+};
+
+struct llama_lora_adapter_deleter {
+    void operator()(llama_lora_adapter * lora_adapter) { llama_lora_adapter_free(lora_adapter); }
+};
+
+typedef std::unique_ptr<llama_model, llama_model_deleter> llama_model_ptr;
+typedef std::unique_ptr<llama_context, llama_context_deleter> llama_context_ptr;
+typedef std::unique_ptr<llama_sampler, llama_sampler_deleter> llama_sampler_ptr;
+typedef std::unique_ptr<llama_lora_adapter, llama_lora_adapter_deleter> llama_lora_adapter_ptr;

package/cpp/llama-grammar.cpp

@@ -1,5 +1,6 @@
 #include "llama-grammar.h"
 
+#include "llama-impl.h"
 #include "llama-vocab.h"
 #include "llama-sampling.h"
 

package/cpp/llama-grammar.h

@@ -1,8 +1,10 @@
 #pragma once
 
-#include "llama-impl.h"
+#include "llama.h"
 
 #include <map>
+#include <string>
+#include <vector>
 
 struct llama_vocab;
 

package/cpp/llama-hparams.cpp

@@ -0,0 +1,71 @@
+#include "llama-hparams.h"
+
+#include "ggml.h"
+
+uint32_t llama_hparams::n_head(uint32_t il) const {
+    if (il < n_layer) {
+        return n_head_arr[il];
+    }
+
+    LM_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];
+    }
+
+    LM_GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_ff(uint32_t il) const {
+    if (il < n_layer) {
+        return n_ff_arr[il];
+    }
+
+    LM_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 2 * 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;
+}

package/cpp/llama-hparams.h

@@ -0,0 +1,140 @@
+#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");
+

package/cpp/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 {
+    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);
+    }
+}