ik_llama.cpp/src/llama-context.h
2026-06-04 20:45:12 -03:00

489 lines
20 KiB
C++

#pragma once
#include "llama-impl.h"
#include "llama-cparams.h"
#include "llama-sampling.h"
#include "llama-spec-features.h"
struct llama_model;
#include <vector>
#include <map>
#include <set>
#include <memory>
struct llama_kv_cell {
llama_pos pos = -1;
llama_pos delta = 0;
int32_t src = 0; // used by recurrent state models to copy states
std::set<llama_seq_id> seq_id;
bool has_seq_id(const llama_seq_id & id) const {
return seq_id.find(id) != seq_id.end();
}
bool is_empty() const {
return seq_id.empty();
}
bool is_same_seq(const llama_kv_cell & other) const {
return seq_id == other.seq_id;
}
};
// ring-buffer of cached KV data
struct llama_kv_cache {
bool has_shift = false;
bool do_defrag = false;
bool do_copy = false;
bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
bool hybrid = false;
bool v_trans = true; // the value tensor is transposed
// Note: The value of head isn't only used to optimize searching
// for a free KV slot. llama_decode_internal also uses it, so it
// cannot be freely changed after a slot has been allocated.
uint32_t head = 0;
uint32_t size = 0;
uint32_t used = 0; // used cells (i.e. at least one seq_id)
// computed before each graph build
uint32_t n = 0;
ggml_type type_k = GGML_TYPE_F16;
ggml_type type_v = GGML_TYPE_F16;
std::vector<llama_kv_cell> cells;
std::vector<struct ggml_tensor *> k_l; // per layer
std::vector<struct ggml_tensor *> v_l;
std::vector<struct ggml_tensor *> s_l; // per layer recurrent state storage (Qwen3Next)
// When true, the delta_net graph builder will enable per-step SSM state saves
bool save_per_step_ssm = false;
std::vector<llama_split_tensor> split_k_l;
std::vector<llama_split_tensor> split_v_l;
std::vector<llama_split_tensor> split_s_l;
// Per-device replicas of the MLA compressed-latent KV cache (-sm graph for DEEPSEEK2/GLM_DSA/MISTRAL4).
std::vector<llama_split_tensor> replicated_k_l;
std::vector<struct ggml_context *> ctxs;
std::vector<ggml_backend_buffer_t> bufs;
size_t total_size() const {
size_t size = 0;
for (ggml_backend_buffer_t buf : bufs) {
size += ggml_backend_buffer_get_size(buf);
}
return size;
}
// GPU-resident checkpoint for recurrent/hybrid speculative decoding
struct gpu_checkpoint {
std::vector<llama_kv_cell> cells_snapshot;
uint32_t head_snapshot = 0;
uint32_t used_snapshot = 0;
std::vector<ggml_tensor *> s_l_shadow;
std::vector<std::vector<ggml_tensor *>> split_s_l_shadow;
// Per-step SSM state checkpoints for speculative decoding.
std::vector<std::vector<ggml_tensor *>> per_step_ssm;
// Per-step conv feature buffer: stores qkv_mixed features from the
// verification forward pass so conv state can be reconstructed at any step.
// One tensor per recurrent layer, each sized [conv_dim * max_tokens].
//std::vector<std::vector<ggml_tensor *>> per_step_qkv;
std::vector<std::vector<ggml_tensor *>> per_step_conv;
int32_t per_step_n_tokens = 0;
int32_t per_step_max_allocated = 0;
int64_t per_step_ssm_state_size = 0;
int64_t per_step_conv_state_dim = 0;
int64_t per_step_conv_dim = 0;
int32_t per_step_d_conv = 0;
int selected_spec_mode = -1;
int fixed_spec_mode = LLAMA_SPEC_CKPT_NONE;
int32_t fixed_max_tokens = 0;
// Serialised sequence state for CPU mode
std::vector<uint8_t> cpu_state_data;
// Separate storage for per-step allocations
std::vector<struct ggml_context *> per_step_ctxs;
std::vector<ggml_backend_buffer_t> per_step_bufs;
std::vector<struct ggml_context *> shadow_ctxs;
std::vector<ggml_backend_buffer_t> shadow_bufs;
bool allocated = false;
bool shadow_conv_only = false;
bool saved = false;
~gpu_checkpoint() {
for (struct ggml_context * ctx : shadow_ctxs) {
ggml_free(ctx);
}
for (ggml_backend_buffer_t buf : shadow_bufs) {
ggml_backend_buffer_free(buf);
}
for (struct ggml_context * ctx : per_step_ctxs) {
ggml_free(ctx);
}
for (ggml_backend_buffer_t buf : per_step_bufs) {
ggml_backend_buffer_free(buf);
}
}
};
gpu_checkpoint ckpt;
bool checkpoint_alloc_shadows(bool conv_only_shadow = false);
bool checkpoint_supported() const;
bool checkpoint_save(ggml_backend_sched_t sched);
bool checkpoint_restore(ggml_backend_sched_t sched);
void checkpoint_delete();
// Per-step checkpoint: allocate, restore step k's full state (SSM + conv) to cache
bool per_step_alloc(const llama_model & model, int max_tokens);
bool per_step_restore(const llama_model & model, ggml_backend_sched_t sched, int step);
~llama_kv_cache() {
for (struct ggml_context * ctx : ctxs) {
ggml_free(ctx);
}
for (ggml_backend_buffer_t buf : bufs) {
ggml_backend_buffer_free(buf);
}
}
};
struct llama_control_vector {
std::vector<struct ggml_tensor *> tensors; // per layer
std::vector<struct ggml_context *> ctxs;
std::vector<ggml_backend_buffer_t> bufs;
int32_t layer_start = -1;
int32_t layer_end = -1;
struct ggml_tensor * tensor_for(int il) const {
if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) {
return nullptr;
}
return tensors[il];
}
struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const {
ggml_tensor * layer_dir = tensor_for(il);
if (layer_dir != nullptr) {
cur = ggml_add(ctx, cur, layer_dir);
}
return cur;
}
~llama_control_vector() {
for (struct ggml_context * ctx : ctxs) {
ggml_free(ctx);
}
for (ggml_backend_buffer_t buf : bufs) {
ggml_backend_buffer_free(buf);
}
}
};
struct llama_context {
llama_context(const llama_model & model);
~llama_context();
const struct llama_model & model;
struct llama_cparams cparams;
struct llama_sampling sampling;
struct llama_kv_cache kv_self;
struct llama_context * mtp_target_ctx = nullptr;
struct llama_control_vector cvec;
std::vector<float> scale_data;
std::unordered_map<struct llama_lora_adapter *, float> lora_adapters;
std::vector<ggml_backend_t> backends;
#ifdef GGML_USE_METAL
ggml_backend_t backend_metal = nullptr;
#endif
#ifdef GGML_USE_BLAS
ggml_backend_t backend_blas = nullptr;
#endif
ggml_backend_t backend_cpu = nullptr;
bool has_evaluated_once = false;
int64_t t_start_us;
int64_t t_load_us;
int64_t t_p_eval_us = 0;
int64_t t_eval_us = 0;
int64_t t_compute_start_us = 0;
int64_t n_queued_tokens = 0;
int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1)
int32_t n_eval = 0; // number of eval calls
// host buffer for the model output (logits and embeddings)
ggml_backend_buffer_t buf_output = nullptr;
// 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
int32_t n_outputs_embd = 0; // number of embedding rows produced for the current 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;
// 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;
ggml_backend_sched_t sched = nullptr;
ggml_abort_callback abort_callback = nullptr;
void * abort_callback_data = nullptr;
const float * draft_input_hidden_state = nullptr;
size_t draft_input_hidden_state_n_floats = 0;
std::vector<float> draft_input_hidden_state_owned;
struct dflash_runtime {
struct target_window_state {
const float * features = nullptr;
size_t features_n_floats = 0;
int32_t features_n_rows = 0;
const float * append_features = nullptr;
size_t append_features_n_floats = 0;
int32_t append_features_n_rows = 0;
const llama_pos * positions = nullptr;
size_t positions_n = 0;
uint64_t version = 0;
int32_t keep_rows = 0;
int32_t append_rows = 0;
bool replace = false;
std::vector<float> features_owned;
std::vector<float> append_features_owned;
std::vector<llama_pos> positions_owned;
std::vector<float> features_padded;
std::vector<llama_pos> pos_ctx_data;
std::vector<float> kq_mask_data;
std::vector<float> kq_mask_swa_data;
};
struct kv_runtime_state {
std::vector<struct ggml_tensor *> k_ctx_cache;
std::vector<struct ggml_tensor *> v_ctx_cache;
std::vector<struct ggml_tensor *> k_ctx_workspace;
std::vector<struct ggml_tensor *> v_ctx_workspace;
struct ggml_context * cache_ctx = nullptr;
std::vector<ggml_backend_buffer_t> cache_bufs;
int32_t cache_write_pos = 0;
int32_t cache_n_filled = 0;
int32_t cache_update_rows = 0;
int32_t cache_reserved_rows = 0;
int32_t cache_view_write_pos = 0;
int32_t cache_view_n_filled = 0;
uint64_t cache_applied_window_version = 0;
bool cache_valid = false;
bool cache_view_valid = false;
int32_t workspace_write_pos = 0;
int32_t workspace_n_filled = 0;
int32_t workspace_reserved_rows = 0;
int32_t workspace_token_capacity = 0;
int32_t workspace_n_kv_total = 0;
uint64_t workspace_applied_window_version = 0;
bool workspace_valid = false;
bool workspace_sync_pending = false;
std::vector<uint8_t> cache_compute_meta;
std::vector<uint8_t> workspace_compute_meta;
ggml_backend_sched_t cache_sched = nullptr;
ggml_backend_sched_t workspace_sched = nullptr;
ggml_cgraph * cache_graph = nullptr;
ggml_cgraph * workspace_graph = nullptr;
int32_t cache_graph_rows = 0;
int32_t cache_graph_write_pos = 0;
int32_t workspace_graph_rows = 0;
int32_t workspace_graph_write_pos = 0;
struct ggml_tensor * cache_input_target_features = nullptr;
struct ggml_tensor * cache_input_pos_ctx = nullptr;
struct ggml_tensor * kq_mask_tensor = nullptr;
struct ggml_tensor * kq_mask_swa_tensor = nullptr;
};
struct capture_state {
std::vector<int32_t> layer_ids;
std::vector<std::vector<float>> layer_rows;
int32_t row_count = 0;
int32_t row_width = 0;
uint64_t capture_batch_id = 0;
std::vector<uint64_t> layer_seen_batch_id;
ggml_backend_sched_eval_callback prev_cb_eval = nullptr;
void * prev_cb_eval_user_data = nullptr;
};
struct input_state {
struct ggml_tensor * target_features = nullptr; // F32 [n_target_features, cross_ctx]
struct ggml_tensor * pos_ctx = nullptr; // I32 [cross_ctx]
struct ggml_tensor * kq_mask = nullptr; // F32 [cross_ctx + n_batch, GGML_PAD(n_batch)]
struct ggml_tensor * kq_mask_swa = nullptr; // F32 [cross_ctx + n_batch, GGML_PAD(n_batch)]
};
target_window_state target;
kv_runtime_state kv;
std::unique_ptr<capture_state> capture;
std::vector<float> feature_view_buffer;
input_state inputs;
int32_t visible_cross_ctx = 0;
llama_dflash_profile_stats profile;
};
dflash_runtime dflash;
using dflash_capture_state = dflash_runtime::capture_state;
const float * & dflash_target_features = dflash.target.features;
size_t & dflash_target_features_n_floats = dflash.target.features_n_floats;
int32_t & dflash_target_features_n_rows = dflash.target.features_n_rows;
const float * & dflash_target_append_features = dflash.target.append_features;
size_t & dflash_target_append_features_n_floats = dflash.target.append_features_n_floats;
int32_t & dflash_target_append_features_n_rows = dflash.target.append_features_n_rows;
const llama_pos * & dflash_target_positions = dflash.target.positions;
size_t & dflash_target_positions_n = dflash.target.positions_n;
uint64_t & dflash_target_window_version = dflash.target.version;
int32_t & dflash_target_window_keep_rows = dflash.target.keep_rows;
int32_t & dflash_target_window_append_rows = dflash.target.append_rows;
bool & dflash_target_window_replace = dflash.target.replace;
std::vector<float> & dflash_target_features_owned = dflash.target.features_owned;
std::vector<float> & dflash_target_append_features_owned = dflash.target.append_features_owned;
std::vector<llama_pos> & dflash_target_positions_owned = dflash.target.positions_owned;
std::vector<float> & dflash_target_features_padded = dflash.target.features_padded;
std::vector<float> & dflash_feature_view_buffer = dflash.feature_view_buffer;
std::vector<llama_pos> & dflash_pos_ctx_data = dflash.target.pos_ctx_data;
std::vector<float> & dflash_kq_mask_data = dflash.target.kq_mask_data;
std::vector<float> & dflash_kq_mask_swa_data = dflash.target.kq_mask_swa_data;
int32_t & dflash_visible_cross_ctx = dflash.visible_cross_ctx;
std::vector<struct ggml_tensor *> & dflash_k_ctx_cache = dflash.kv.k_ctx_cache;
std::vector<struct ggml_tensor *> & dflash_v_ctx_cache = dflash.kv.v_ctx_cache;
// Argmax token IDs from the DFlash draft graph, computed via GPU argmax.
// Populated in llama_decode_internal after graph compute.
std::vector<llama_token> dflash_draft_tokens;
struct ggml_tensor * dflash_draft_tokens_tensor = nullptr;
std::vector<struct ggml_tensor *> & dflash_k_ctx_workspace = dflash.kv.k_ctx_workspace;
std::vector<struct ggml_tensor *> & dflash_v_ctx_workspace = dflash.kv.v_ctx_workspace;
struct ggml_context * & dflash_cache_ctx = dflash.kv.cache_ctx;
std::vector<ggml_backend_buffer_t> & dflash_cache_bufs = dflash.kv.cache_bufs;
int32_t & dflash_kv_cache_write_pos = dflash.kv.cache_write_pos;
int32_t & dflash_kv_cache_n_filled = dflash.kv.cache_n_filled;
int32_t & dflash_kv_cache_update_rows = dflash.kv.cache_update_rows;
int32_t & dflash_kv_cache_reserved_rows = dflash.kv.cache_reserved_rows;
int32_t & dflash_kv_cache_view_write_pos = dflash.kv.cache_view_write_pos;
int32_t & dflash_kv_cache_view_n_filled = dflash.kv.cache_view_n_filled;
uint64_t & dflash_kv_cache_applied_window_version = dflash.kv.cache_applied_window_version;
bool & dflash_kv_cache_valid = dflash.kv.cache_valid;
bool & dflash_kv_cache_view_valid = dflash.kv.cache_view_valid;
int32_t & dflash_kv_workspace_write_pos = dflash.kv.workspace_write_pos;
int32_t & dflash_kv_workspace_n_filled = dflash.kv.workspace_n_filled;
int32_t & dflash_kv_workspace_reserved_rows = dflash.kv.workspace_reserved_rows;
int32_t & dflash_kv_workspace_token_capacity = dflash.kv.workspace_token_capacity;
int32_t & dflash_kv_workspace_n_kv_total = dflash.kv.workspace_n_kv_total;
uint64_t & dflash_kv_workspace_applied_window_version = dflash.kv.workspace_applied_window_version;
bool & dflash_kv_workspace_valid = dflash.kv.workspace_valid;
bool & dflash_kv_workspace_sync_pending = dflash.kv.workspace_sync_pending;
std::vector<uint8_t> & dflash_buf_compute_meta = dflash.kv.cache_compute_meta;
std::vector<uint8_t> & dflash_workspace_buf_compute_meta = dflash.kv.workspace_compute_meta;
ggml_backend_sched_t & dflash_sched = dflash.kv.cache_sched;
ggml_backend_sched_t & dflash_workspace_sched = dflash.kv.workspace_sched;
ggml_cgraph * & dflash_kv_graph = dflash.kv.cache_graph;
ggml_cgraph * & dflash_kv_workspace_graph = dflash.kv.workspace_graph;
int32_t & dflash_kv_graph_rows = dflash.kv.cache_graph_rows;
int32_t & dflash_kv_graph_write_pos = dflash.kv.cache_graph_write_pos;
int32_t & dflash_kv_workspace_graph_rows = dflash.kv.workspace_graph_rows;
int32_t & dflash_kv_workspace_graph_write_pos = dflash.kv.workspace_graph_write_pos;
struct ggml_tensor * & dflash_kv_input_target_features = dflash.kv.cache_input_target_features;
struct ggml_tensor * & dflash_kv_input_pos_ctx = dflash.kv.cache_input_pos_ctx;
struct ggml_tensor * & dflash_kq_mask_tensor = dflash.kv.kq_mask_tensor;
struct ggml_tensor * & dflash_kq_mask_swa_tensor = dflash.kv.kq_mask_swa_tensor;
std::unique_ptr<dflash_capture_state> & dflash_capture = dflash.capture;
llama_dflash_profile_stats & dflash_profile = dflash.profile;
struct ggml_tensor * & inp_dflash_target_features = dflash.inputs.target_features;
struct ggml_tensor * & inp_dflash_pos_ctx = dflash.inputs.pos_ctx;
struct ggml_tensor * & inp_dflash_kq_mask = dflash.inputs.kq_mask;
struct ggml_tensor * & inp_dflash_kq_mask_swa = dflash.inputs.kq_mask_swa;
// input tensors
struct ggml_tensor * inp_tokens; // I32 [n_batch]
struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch]
struct ggml_tensor * inp_pos; // I32 [n_batch]
struct ggml_tensor * inp_out_ids; // I32 [n_outputs]
struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch]
struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch]
struct ggml_tensor * inp_K_shift; // I32 [kv_size]
struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch]
struct ggml_tensor * inp_cls; // I32 [n_batch]
struct ggml_tensor * inp_s_copy; // I32 [kv_size]
struct ggml_tensor * inp_s_mask; // F32 [1, n_kv]
struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch]
struct ggml_tensor * inp_s_seq_qnext; // I32 [1, n_batch]
struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch]
struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc]
struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch]
struct ggml_tensor * inp_scale = nullptr; // F32 [n_tokens]
struct ggml_tensor * inp_mtp_states = nullptr;
ggml_backend_t ggml_backend_by_name(const char * name);
struct Prev;
std::unique_ptr<Prev> prev;
std::unique_ptr<Prev> prev_mtp;
void reset_scheduler();
bool can_reuse_graph(const llama_batch & u_batch);
struct CacheCopy {
ggml_tensor * cpy = nullptr;
size_t step = 0;
};
std::vector<CacheCopy> cache_copies;
bool update_cache_copies();
bool ensure_dflash_kv_cache_tensors(int32_t cross_ctx);
void free_dflash_kv_cache_tensors();
bool prepare_mtp_graph_inputs(
struct llama_context & lctx);
void set_mtp_op_type(llama_mtp_op_type value);
};