cui-llama.rn 1.4.4 → 1.5.0

package/cpp/llama-context.h

@@ -1,128 +1,265 @@
-#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_adapter_cvec cvec;
-
-    std::unordered_map<struct llama_adapter_lora *, float> lora;
-
-    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);
+#pragma once
+
+#include "llama.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-graph.h"
+#include "llama-adapter.h"
+
+#include "ggml-cpp.h"
+
+#include <map>
+#include <vector>
+
+struct llama_model;
+struct llama_kv_cache;
+
+class llama_io_read_i;
+class llama_io_write_i;
+
+struct llama_context {
+    // init scheduler and compute buffers, reserve worst-case graphs
+    llama_context(
+            const llama_model & model,
+                  llama_context_params params);
+
+    ~llama_context();
+
+    void synchronize();
+
+    const llama_model & get_model() const;
+
+    uint32_t n_ctx()         const;
+    uint32_t n_ctx_per_seq() const;
+    uint32_t n_batch()       const;
+    uint32_t n_ubatch()      const;
+    uint32_t n_seq_max()     const;
+
+    uint32_t n_threads()       const;
+    uint32_t n_threads_batch() const;
+
+          llama_kv_cache * get_kv_self();
+    const llama_kv_cache * get_kv_self() const;
+
+    void kv_self_update();
+
+    enum llama_pooling_type pooling_type() const;
+
+    float * get_logits();
+    float * get_logits_ith(int32_t i);
+
+    float * get_embeddings();
+    float * get_embeddings_ith(int32_t i);
+    float * get_embeddings_seq(llama_seq_id seq_id);
+
+    void attach_threadpool(
+            lm_ggml_threadpool_t threadpool,
+            lm_ggml_threadpool_t threadpool_batch);
+
+    void detach_threadpool();
+
+    void set_n_threads(int32_t n_threads, int32_t n_threads_batch);
+
+    void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data);
+
+    void set_embeddings (bool value);
+    void set_causal_attn(bool value);
+    void set_warmup(bool value);
+
+    void set_adapter_lora(
+            llama_adapter_lora * adapter,
+            float scale);
+
+    bool rm_adapter_lora(
+            llama_adapter_lora * adapter);
+
+    void clear_adapter_lora();
+
+    bool apply_adapter_cvec(
+            const float * data,
+                 size_t   len,
+                int32_t   n_embd,
+                int32_t   il_start,
+                int32_t   il_end);
+
+    int encode(llama_batch & inp_batch);
+    int decode(llama_batch & inp_batch);
+
+    //
+    // state save/load
+    //
+
+    size_t state_get_size();
+    size_t state_get_data(      uint8_t * dst, size_t size);
+    size_t state_set_data(const uint8_t * src, size_t size);
+
+    size_t state_seq_get_size(llama_seq_id seq_id);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+
+    bool state_load_file(
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    bool state_save_file(
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    size_t state_seq_load_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+           llama_token * tokens_out,
+                size_t   n_token_capacity,
+                size_t * n_token_count_out);
+
+    size_t state_seq_save_file(
+          llama_seq_id   seq_id,
+            const char * filepath,
+     const llama_token * tokens,
+                size_t   n_token_count);
+
+    //
+    // perf
+    //
+
+    llama_perf_context_data perf_get_data() const;
+    void perf_reset();
+
+private:
+    //
+    // output
+    //
+
+    // Make sure enough space is available for outputs.
+    // Returns max number of outputs for which space was reserved.
+    int32_t output_reserve(int32_t n_outputs);
+
+    // make the outputs have the same order they had in the user-provided batch
+    // TODO: maybe remove this
+    void output_reorder();
+
+    //
+    // graph
+    //
+
+    int32_t graph_max_nodes() const;
+
+    // zero-out inputs and create the ctx_compute for the compute graph
+    lm_ggml_cgraph * graph_init();
+
+    llm_graph_result_ptr graph_build(
+            lm_ggml_context * ctx,
+             lm_ggml_cgraph * gf,
+      const llama_ubatch & ubatch,
+          llm_graph_type   gtype);
+
+    // returns the result of lm_ggml_backend_sched_graph_compute_async execution
+    lm_ggml_status graph_compute(
+            lm_ggml_cgraph * gf,
+                   bool   batched);
+
+    llm_graph_cb graph_get_cb() const;
+
+    // used by kv_self_update()
+    lm_ggml_tensor * build_rope_shift(
+        lm_ggml_context * ctx0,
+        lm_ggml_tensor * cur,
+        lm_ggml_tensor * shift,
+        lm_ggml_tensor * factors,
+              float   freq_base,
+              float   freq_scale,
+        lm_ggml_backend_buffer * bbuf) const;
+
+    llm_graph_result_ptr build_kv_self_shift(
+            lm_ggml_context * ctx0,
+            lm_ggml_cgraph * gf) const;
+
+    llm_graph_result_ptr build_kv_self_defrag(
+            lm_ggml_context * ctx0,
+            lm_ggml_cgraph * gf) const;
+
+    // TODO: read/write lora adapters and cvec
+    size_t state_write_data(llama_io_write_i & io);
+    size_t state_read_data (llama_io_read_i  & io);
+
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+
+    //
+    // members
+    //
+
+    const llama_model & model;
+
+    llama_cparams       cparams;
+    llama_adapter_cvec  cvec;
+    llama_adapter_loras loras;
+    llama_sbatch        sbatch;
+
+    llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably
+
+    std::unique_ptr<llama_kv_cache_unified> kv_self;
+
+    // TODO: remove
+    bool logits_all = false;
+
+    // decode output (2-dimensional array: [n_outputs][n_vocab])
+    size_t  logits_size = 0; // capacity (of floats) for logits
+    float * logits      = nullptr;
+
+    // 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;
+
+    int32_t n_outputs     = 0; // number of actually-used outputs in the current ubatch or last logical batch
+    int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers
+
+    std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
+
+    lm_ggml_backend_sched_ptr sched;
+
+    lm_ggml_backend_t backend_cpu = nullptr;
+    std::vector<lm_ggml_backend_ptr> backends;
+
+    lm_ggml_context_ptr ctx_compute;
+
+    lm_ggml_threadpool_t threadpool       = nullptr;
+    lm_ggml_threadpool_t threadpool_batch = nullptr;
+
+    lm_ggml_abort_callback abort_callback      = nullptr;
+    void *              abort_callback_data = nullptr;
+
+    std::vector<std::pair<lm_ggml_backend_t, lm_ggml_backend_set_n_threads_t>> set_n_threads_fns;
+
+    // buffer types used for the compute buffer of each backend
+    std::vector<lm_ggml_backend_t>             backend_ptrs;
+    std::vector<lm_ggml_backend_buffer_type_t> backend_buft;
+
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
+
+    // host buffer for the model output (logits and embeddings)
+    lm_ggml_backend_buffer_ptr buf_output;
+
+    bool has_evaluated_once = false;
+
+    // perf
+    mutable int64_t t_start_us  = 0;
+    mutable int64_t t_load_us   = 0;
+    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
+};

package/cpp/llama-cparams.cpp

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

package/cpp/llama-cparams.h

@@ -1,37 +1,38 @@
-#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;
-};
+#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;
+    bool warmup;
+
+    enum llama_pooling_type pooling_type;
+
+    lm_ggml_backend_sched_eval_callback cb_eval;
+    void * cb_eval_user_data;
+};

package/cpp/llama-cpp.h

@@ -1,30 +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_adapter_lora_deleter {
-    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(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_adapter_lora, llama_adapter_lora_deleter> llama_adapter_lora_ptr;
+#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_adapter_lora_deleter {
+    void operator()(llama_adapter_lora * adapter) { llama_adapter_lora_free(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_adapter_lora, llama_adapter_lora_deleter> llama_adapter_lora_ptr;