mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-08-14 04:17:53 -04:00
kv-cache : separate recurrent vs non-recurrent impl (#12799)
* kv-cache : serparate recurrent vs non-recurrent impl (wip) ggml-ci * kv-cache : init -> contructor + add llama_memory_params ggml-ci * kv-cache : fix callback reference ggml-ci * context : llama_kv_cache -> llama_memory_i ggml-ci * context : move memory creation logic to model ggml-ci * llama : remove reference of memory during encode ggml-ci * kv-cache : hide padding details in the implementation ggml-ci * kv-cache : add ubatch_next() ggml-ci * context : simplify sbatch logic ggml-ci * kv-cache : hide defrag logic in the implementation ggml-ci * context : hide kv cache details in implementation ggml-ci * build : fix ggml-ci * cont : another fix ggml-ci * kv-cache : simplify interface (wip) ggml-ci * kv-cache : use separate KV cell structs for unified/recurrent ggml-ci * kv-cache : clean-up ggml-ci * model : better llama_model::create_model() signature ggml-ci * kv-cache : fix recurrent seq_rm() ggml-ci * kv-cache : replace `struct callbacks` with `llama_model &` ggml-ci * kv-cache : replace `struct graph_params` with `llama_context &` ggml-ci * kv-cache : fix offload check ggml-ci * context : avoid passing unique_ptr ggml-ci * kv-cache : avoid using the backends from the llama_context ref #13113 ggml-ci * kv-cache : more consistent debug logs [no ci] * kv-cache : do not pass the full llama_context for kv graphs ggml-ci * kv-cache : remove comment * kv-cache : ggml_rope_ext_inplace -> ggml_rope_ext ggml-ci * kv-cache : fix recurrent multi-user case ggml-ci * memory : remove comments [no ci]
This commit is contained in:
Georgi Gerganov
GitHub
@@ -2,32 +2,72 @@
|
||||
|
||||
#include "llama.h"
|
||||
#include "llama-io.h"
|
||||
#include "llama-graph.h"
|
||||
#include "llama-memory.h"
|
||||
|
||||
#include "ggml-cpp.h"
|
||||
|
||||
#include <functional>
|
||||
#include <set>
|
||||
#include <vector>
|
||||
|
||||
struct llama_cparams;
|
||||
struct llama_hparams;
|
||||
struct llama_ubatch;
|
||||
struct llama_sbatch;
|
||||
struct llama_model;
|
||||
struct llama_context;
|
||||
|
||||
struct llama_kv_cache : public llama_memory_i {
|
||||
using llama_memory_i::llama_memory_i;
|
||||
virtual ~llama_kv_cache() = default;
|
||||
|
||||
virtual void restore() = 0; // call if batch processing fails - restores the cache state
|
||||
virtual void commit() = 0; // call after successful batch processing - clears any pending state
|
||||
// call if batch processing fails - restores the cache state
|
||||
virtual void restore() = 0;
|
||||
|
||||
virtual int32_t get_n_tokens() const = 0;
|
||||
virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
|
||||
// call after successful batch processing - clears any pending state
|
||||
virtual void commit() = 0;
|
||||
|
||||
virtual bool get_can_shift() const = 0;
|
||||
// process any pending defrag/shift/etc. operations
|
||||
// optionally call once before processing a new batch
|
||||
virtual bool update(llama_context & lctx) = 0;
|
||||
|
||||
// schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
|
||||
virtual void defrag_sched(float thold) = 0;
|
||||
|
||||
// simulate full cache, used for allocating worst-case compute buffers
|
||||
virtual void set_full() = 0;
|
||||
|
||||
//
|
||||
// batch processing
|
||||
//
|
||||
|
||||
virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
|
||||
|
||||
// different KV caches require different batch splitting strategies
|
||||
virtual llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const = 0;
|
||||
|
||||
// find an empty slot of size "n_tokens" in the cache
|
||||
virtual bool find_slot(const llama_ubatch & batch) = 0;
|
||||
|
||||
// getters
|
||||
virtual int32_t get_n_tokens() const = 0;
|
||||
virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache
|
||||
virtual llama_pos get_pos_max() const = 0;
|
||||
virtual bool get_can_shift() const = 0;
|
||||
|
||||
bool get_can_edit() const override { return get_can_shift(); }
|
||||
|
||||
//
|
||||
// state write/read
|
||||
//
|
||||
|
||||
virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
|
||||
virtual void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) = 0;
|
||||
};
|
||||
|
||||
//
|
||||
// llama_kv_cache_guard
|
||||
//
|
||||
|
||||
struct llama_kv_cache_guard {
|
||||
llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {}
|
||||
|
||||
@@ -43,65 +83,50 @@ private:
|
||||
llama_kv_cache * kv;
|
||||
};
|
||||
|
||||
struct llama_kv_cell {
|
||||
llama_pos pos = -1;
|
||||
llama_pos delta = 0;
|
||||
int32_t src = -1; // used by recurrent state models to copy states
|
||||
int32_t tail = -1;
|
||||
//
|
||||
// llama_kv_cache_unified
|
||||
//
|
||||
|
||||
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
|
||||
// TODO: pimpl
|
||||
// TODO: add notion of max sequences
|
||||
class llama_kv_cache_unified : public llama_kv_cache {
|
||||
public:
|
||||
// can be used to query data from the model if needed
|
||||
struct callbacks {
|
||||
std::function<ggml_tensor * (uint32_t n_ctx_per_seq, int il)> get_rope_factors;
|
||||
struct kv_cell {
|
||||
llama_pos pos = -1;
|
||||
llama_pos delta = 0;
|
||||
|
||||
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 kv_cell & other) const {
|
||||
return seq_id == other.seq_id;
|
||||
}
|
||||
};
|
||||
|
||||
static uint32_t get_padding(const llama_cparams & cparams);
|
||||
|
||||
llama_kv_cache_unified(
|
||||
const llama_hparams & hparams,
|
||||
callbacks cbs);
|
||||
|
||||
virtual ~llama_kv_cache_unified() = default;
|
||||
|
||||
// TODO: become constructor
|
||||
bool init(
|
||||
const llama_model & model, // TODO: do not reference the model
|
||||
const llama_cparams & cparams,
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
uint32_t kv_size,
|
||||
bool offload);
|
||||
uint32_t padding);
|
||||
|
||||
int32_t get_n_tokens() const override;
|
||||
int32_t get_used_cells() const override;
|
||||
~llama_kv_cache_unified() = default;
|
||||
|
||||
size_t total_size() const;
|
||||
|
||||
// TODO: better data structures to reduce the cost of this operation
|
||||
llama_pos pos_max() const;
|
||||
//
|
||||
// llama_memory_i
|
||||
//
|
||||
|
||||
void clear() override;
|
||||
void defrag() override;
|
||||
|
||||
virtual void restore() override;
|
||||
virtual void commit() override;
|
||||
|
||||
bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
|
||||
void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
|
||||
@@ -111,63 +136,40 @@ public:
|
||||
|
||||
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
|
||||
|
||||
bool get_can_shift() const override;
|
||||
//
|
||||
// llama_kv_cache
|
||||
//
|
||||
|
||||
void restore() override;
|
||||
void commit() override;
|
||||
|
||||
bool update(llama_context & ctx) override;
|
||||
|
||||
void defrag_sched(float thold) override;
|
||||
|
||||
void set_full() override;
|
||||
|
||||
llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
|
||||
|
||||
llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
|
||||
|
||||
// find an empty slot of size "n_tokens" in the cache
|
||||
// updates the cache head
|
||||
// Note: On success, it's important that cache.head points
|
||||
// to the first cell of the slot.
|
||||
bool find_slot(const llama_ubatch & batch);
|
||||
bool find_slot(const llama_ubatch & batch) override;
|
||||
|
||||
// TODO: maybe not needed
|
||||
uint32_t get_padding(const llama_cparams & cparams) const;
|
||||
int32_t get_n_tokens() const override;
|
||||
int32_t get_used_cells() const override;
|
||||
|
||||
// find how many cells are currently in use
|
||||
uint32_t cell_max() const;
|
||||
// TODO: better data structures to reduce the cost of this operation
|
||||
llama_pos get_pos_max() const override;
|
||||
|
||||
size_t size_k_bytes() const;
|
||||
size_t size_v_bytes() const;
|
||||
|
||||
// defrag
|
||||
|
||||
struct {
|
||||
std::vector<uint32_t> ids;
|
||||
} defrag_info;
|
||||
|
||||
// return true if cells have been moved
|
||||
bool defrag_prepare(int32_t n_max_nodes);
|
||||
|
||||
// commit/restore cache
|
||||
|
||||
struct slot_range {
|
||||
uint32_t c0 = 0; // note: these are cell indices, not sequence positions
|
||||
uint32_t c1 = 0;
|
||||
};
|
||||
|
||||
// pending cell updates that are not yet committed
|
||||
struct {
|
||||
std::vector<slot_range> ranges;
|
||||
} pending;
|
||||
bool get_can_shift() const override;
|
||||
|
||||
// state write/load
|
||||
|
||||
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const;
|
||||
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1);
|
||||
|
||||
// members
|
||||
|
||||
const llama_hparams & hparams;
|
||||
|
||||
callbacks cbs;
|
||||
|
||||
bool has_shift = false;
|
||||
bool do_defrag = false;
|
||||
|
||||
// TODO: remove this and implement llama_kv_cache_recurrent instead
|
||||
bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token
|
||||
|
||||
bool v_trans = true; // the value tensor is transposed
|
||||
bool can_shift = false;
|
||||
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
|
||||
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) override;
|
||||
|
||||
// Note: The value of head isn't only used to optimize searching
|
||||
// for a free KV slot. llama_decode_impl also uses it, so it
|
||||
@@ -179,18 +181,213 @@ public:
|
||||
// computed before each graph build
|
||||
uint32_t n = 0;
|
||||
|
||||
std::vector<llama_kv_cell> cells;
|
||||
std::vector<kv_cell> cells;
|
||||
|
||||
std::vector<ggml_tensor *> k_l; // per layer
|
||||
std::vector<ggml_tensor *> v_l;
|
||||
|
||||
private:
|
||||
const llama_model & model;
|
||||
const llama_hparams & hparams;
|
||||
|
||||
bool has_shift = false;
|
||||
bool do_defrag = false;
|
||||
|
||||
bool v_trans = true; // the value tensor is transposed
|
||||
bool can_shift = false;
|
||||
|
||||
// required padding
|
||||
uint32_t padding = 1;
|
||||
|
||||
ggml_type type_k = GGML_TYPE_F16;
|
||||
ggml_type type_v = GGML_TYPE_F16;
|
||||
|
||||
std::vector<ggml_context_ptr> ctxs;
|
||||
std::vector<ggml_backend_buffer_ptr> bufs;
|
||||
|
||||
// defrag
|
||||
struct {
|
||||
std::vector<uint32_t> ids;
|
||||
} defrag_info;
|
||||
|
||||
// return true if cells have been moved
|
||||
bool defrag_prepare(int32_t n_max_nodes);
|
||||
|
||||
// commit/restore cache
|
||||
struct slot_range {
|
||||
uint32_t c0 = 0; // note: these are cell indices, not sequence positions
|
||||
uint32_t c1 = 0;
|
||||
};
|
||||
|
||||
// pending cell updates that are not yet committed
|
||||
struct {
|
||||
std::vector<slot_range> ranges;
|
||||
} pending;
|
||||
|
||||
// find how many cells are currently in use
|
||||
uint32_t cell_max() const;
|
||||
|
||||
size_t total_size() const;
|
||||
|
||||
size_t size_k_bytes() const;
|
||||
size_t size_v_bytes() const;
|
||||
|
||||
ggml_tensor * build_rope_shift(
|
||||
const llama_cparams & cparams,
|
||||
ggml_context * ctx,
|
||||
ggml_tensor * cur,
|
||||
ggml_tensor * shift,
|
||||
ggml_tensor * factors,
|
||||
float freq_base,
|
||||
float freq_scale) const;
|
||||
|
||||
llm_graph_result_ptr build_graph_shift(
|
||||
const llama_cparams & cparams,
|
||||
ggml_context * ctx,
|
||||
ggml_cgraph * gf) const;
|
||||
|
||||
llm_graph_result_ptr build_graph_defrag(
|
||||
const llama_cparams & cparams,
|
||||
ggml_context * ctx,
|
||||
ggml_cgraph * gf) const;
|
||||
|
||||
void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
|
||||
void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
|
||||
|
||||
bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
|
||||
bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
|
||||
};
|
||||
|
||||
//
|
||||
// llama_kv_cache_recurrent
|
||||
//
|
||||
|
||||
class llama_kv_cache_recurrent : public llama_kv_cache {
|
||||
public:
|
||||
struct kv_cell {
|
||||
llama_pos pos = -1;
|
||||
int32_t src = -1; // used to copy states
|
||||
int32_t tail = -1;
|
||||
|
||||
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 kv_cell & other) const {
|
||||
return seq_id == other.seq_id;
|
||||
}
|
||||
};
|
||||
|
||||
llama_kv_cache_recurrent(
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool offload,
|
||||
uint32_t kv_size);
|
||||
|
||||
~llama_kv_cache_recurrent() = default;
|
||||
|
||||
//
|
||||
// llama_memory_i
|
||||
//
|
||||
|
||||
void clear() override;
|
||||
|
||||
bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override;
|
||||
void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
|
||||
void seq_keep(llama_seq_id seq_id) override;
|
||||
void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) override;
|
||||
void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override;
|
||||
|
||||
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
|
||||
|
||||
//
|
||||
// llama_kv_cache
|
||||
//
|
||||
|
||||
void restore() override;
|
||||
void commit() override;
|
||||
|
||||
bool update(llama_context & lctx) override;
|
||||
|
||||
void defrag_sched(float thold) override;
|
||||
|
||||
void set_full() override;
|
||||
|
||||
llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) override;
|
||||
|
||||
llama_ubatch ubatch_next(llama_sbatch & sbatch, uint32_t n_ubatch, bool embd_pooled) const override;
|
||||
|
||||
bool find_slot(const llama_ubatch & batch) override;
|
||||
|
||||
int32_t get_n_tokens() const override;
|
||||
int32_t get_used_cells() const override;
|
||||
|
||||
// TODO: better data structures to reduce the cost of this operation
|
||||
llama_pos get_pos_max() const override;
|
||||
|
||||
bool get_can_shift() const override;
|
||||
|
||||
// TODO: temporary methods - they are not really const as they do const_cast<>, fix this
|
||||
int32_t s_copy(int i) const;
|
||||
float s_mask(int i) const;
|
||||
|
||||
// state write/load
|
||||
|
||||
void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
|
||||
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) override;
|
||||
|
||||
// Note: The value of head isn't only used to optimize searching
|
||||
// for a free KV slot. llama_decode_impl 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;
|
||||
|
||||
std::vector<kv_cell> cells;
|
||||
|
||||
std::vector<ggml_tensor *> k_l; // per layer
|
||||
std::vector<ggml_tensor *> v_l;
|
||||
|
||||
private:
|
||||
//const llama_model & model;
|
||||
const llama_hparams & hparams;
|
||||
|
||||
// commit/restore cache
|
||||
// TODO: rework for recurrent cache
|
||||
struct slot_range {
|
||||
uint32_t c0 = 0; // note: these are cell indices, not sequence positions
|
||||
uint32_t c1 = 0;
|
||||
};
|
||||
|
||||
// pending cell updates that are not yet committed
|
||||
struct {
|
||||
std::vector<slot_range> ranges;
|
||||
} pending;
|
||||
|
||||
ggml_type type_k = GGML_TYPE_F16;
|
||||
ggml_type type_v = GGML_TYPE_F16;
|
||||
|
||||
std::vector<ggml_context_ptr> ctxs;
|
||||
std::vector<ggml_backend_buffer_ptr> bufs;
|
||||
|
||||
// find how many cells are currently in use
|
||||
uint32_t cell_max() const;
|
||||
|
||||
size_t total_size() const;
|
||||
|
||||
size_t size_k_bytes() const;
|
||||
size_t size_v_bytes() const;
|
||||
|
||||
void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
|
||||
void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
|
||||
|
||||
@@ -198,11 +395,6 @@ private:
|
||||
bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
|
||||
};
|
||||
|
||||
// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified
|
||||
//class llama_kv_cache_recurrent : public llama_kv_cache_unified {
|
||||
//public:
|
||||
// using llama_kv_cache_unified::llama_kv_cache_unified;
|
||||
//};
|
||||
|
||||
//
|
||||
// kv cache view
|
||||
|
||||
Reference in New Issue
Block a user