mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-10-27 08:21:30 +00:00
kv-cache : add SWA support (#13194)
* kv-cache : prepare for SWA ggml-ci * kv-cache : initial iSWA implementation ggml-ci * kv-cache : rework error recovery logic ggml-ci * models : fix Phi-3 SWA parameters ggml-ci * model : adjust Granite to rope factor changes ggml-ci * server : check if context can do shifts ggml-ci * iswa : for now, always enable shifts (experiment) ggml-ci * kv-cache : simplify SWA logic ggml-ci * kv-cache : apply defrag when we fail to find slots for the batch ggml-ci * llama : update docs about llama_decode ggml-ci * kv-cache : update warning logs when no space for the batch is available ggml-ci * llama : add llama_kv_self_seq_pos_min() * kv-cache : keep track of partial SWA computes and print warnings * server : disallow use cases involving partial SWA context ggml-ci * llama : add param to control SWA cache size ggml-ci * minor : clean-up ggml-ci
This commit is contained in:
Georgi Gerganov
@@ -8,6 +8,7 @@
|
||||
#include "ggml-cpp.h"
|
||||
|
||||
#include <set>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
|
||||
struct llama_cparams;
|
||||
@@ -40,6 +41,9 @@ struct llama_kv_cache : public llama_memory_i {
|
||||
// batch processing
|
||||
//
|
||||
|
||||
// =============================================================================================================
|
||||
// TODO: refactor and simplify this
|
||||
|
||||
virtual llama_sbatch sbatch_init(const llama_batch & batch, bool logits_all) = 0;
|
||||
|
||||
// different KV caches require different batch splitting strategies
|
||||
@@ -48,6 +52,8 @@ struct llama_kv_cache : public llama_memory_i {
|
||||
// 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
|
||||
@@ -87,38 +93,24 @@ private:
|
||||
// llama_kv_cache_unified
|
||||
//
|
||||
|
||||
// TODO: add notion of max sequences
|
||||
class llama_kv_cache_unified : public llama_kv_cache {
|
||||
public:
|
||||
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);
|
||||
|
||||
// this callback is used to filter out layers that should not be included in the cache
|
||||
using layer_filter_cb = std::function<bool(int32_t il)>;
|
||||
|
||||
llama_kv_cache_unified(
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
uint32_t kv_size,
|
||||
uint32_t padding);
|
||||
const llama_model & model,
|
||||
layer_filter_cb && filter,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
uint32_t kv_size,
|
||||
uint32_t padding,
|
||||
uint32_t n_swa,
|
||||
llama_swa_type swa_type);
|
||||
|
||||
~llama_kv_cache_unified() = default;
|
||||
|
||||
@@ -130,10 +122,11 @@ public:
|
||||
|
||||
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_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_min(llama_seq_id seq_id) const override;
|
||||
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
|
||||
|
||||
//
|
||||
@@ -150,7 +143,6 @@ public:
|
||||
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;
|
||||
|
||||
// updates the cache head
|
||||
@@ -169,29 +161,72 @@ public:
|
||||
// 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;
|
||||
void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1) override;
|
||||
|
||||
uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
|
||||
uint32_t size = 0; // total number of cells, shared across all sequences
|
||||
uint32_t used = 0; // used cells (i.e. at least one seq_id)
|
||||
//
|
||||
// llama_kv_cache_unified specific API
|
||||
//
|
||||
|
||||
// computed before each graph build
|
||||
uint32_t n = 0;
|
||||
uint32_t get_n() const;
|
||||
uint32_t get_size() const;
|
||||
|
||||
std::vector<kv_cell> cells;
|
||||
// get views of the current state of the cache
|
||||
ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
|
||||
ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
|
||||
|
||||
std::vector<ggml_tensor *> k_l; // per layer
|
||||
std::vector<ggml_tensor *> v_l;
|
||||
// store k_cur and v_cur in the cache based on the current head location
|
||||
ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
|
||||
ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
|
||||
|
||||
void prune_swa(llama_seq_id seq_id, llama_pos pmin, llama_pos pmax);
|
||||
|
||||
void set_input_kq_mask (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
|
||||
void set_input_k_shift (ggml_tensor * dst) const;
|
||||
void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
|
||||
|
||||
private:
|
||||
const llama_model & model;
|
||||
const llama_hparams & hparams;
|
||||
|
||||
struct kv_cell {
|
||||
llama_pos pos = -1;
|
||||
llama_pos delta = 0;
|
||||
|
||||
// TODO: replace with bitset uint64_t
|
||||
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;
|
||||
}
|
||||
};
|
||||
|
||||
struct kv_layer {
|
||||
// layer index in the model
|
||||
// note: can be different from the layer index in the KV cache
|
||||
uint32_t il;
|
||||
|
||||
ggml_tensor * k;
|
||||
ggml_tensor * v;
|
||||
};
|
||||
|
||||
bool has_shift = false;
|
||||
bool do_defrag = false;
|
||||
|
||||
bool v_trans = true; // the value tensor is transposed
|
||||
bool can_shift = false;
|
||||
|
||||
uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
|
||||
uint32_t size = 0; // total number of cells, shared across all sequences
|
||||
uint32_t used = 0; // used cells (i.e. at least one seq_id) (TODO: add `struct kv_cells` and keep track automaticallt)
|
||||
|
||||
// computed before each graph build
|
||||
uint32_t n = 0;
|
||||
|
||||
// required padding
|
||||
uint32_t padding = 1;
|
||||
@@ -199,9 +234,29 @@ private:
|
||||
ggml_type type_k = GGML_TYPE_F16;
|
||||
ggml_type type_v = GGML_TYPE_F16;
|
||||
|
||||
// SWA
|
||||
uint32_t n_swa = 0;
|
||||
|
||||
llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
|
||||
|
||||
std::vector<ggml_context_ptr> ctxs;
|
||||
std::vector<ggml_backend_buffer_ptr> bufs;
|
||||
|
||||
std::vector<kv_cell> cells; // TODO: replace with `struct kv_cells`
|
||||
std::vector<kv_layer> layers;
|
||||
|
||||
// model layer id -> KV cache layer id
|
||||
std::unordered_map<int32_t, int32_t> map_layer_ids;
|
||||
|
||||
// recovery information used to restore the KV cells to their original state in case of a failure
|
||||
struct {
|
||||
void clear() {
|
||||
cells.clear();
|
||||
}
|
||||
|
||||
std::unordered_map<uint32_t, kv_cell> cells;
|
||||
} recovery;
|
||||
|
||||
// defrag
|
||||
struct {
|
||||
std::vector<uint32_t> ids;
|
||||
@@ -210,17 +265,6 @@ private:
|
||||
// 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;
|
||||
|
||||
@@ -229,6 +273,8 @@ private:
|
||||
size_t size_k_bytes() const;
|
||||
size_t size_v_bytes() const;
|
||||
|
||||
bool is_masked_swa(llama_pos p0, llama_pos p1) const;
|
||||
|
||||
ggml_tensor * build_rope_shift(
|
||||
const llama_cparams & cparams,
|
||||
ggml_context * ctx,
|
||||
@@ -255,6 +301,106 @@ private:
|
||||
bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
|
||||
};
|
||||
|
||||
//
|
||||
// llama_kv_cache_unified_iswa
|
||||
//
|
||||
|
||||
// utilizes two instances of llama_kv_cache_unified
|
||||
// the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
|
||||
// upon successful commit, the SWA cache removes old tokens outside the n_swa window
|
||||
|
||||
class llama_kv_cache_unified_iswa : public llama_kv_cache {
|
||||
public:
|
||||
llama_kv_cache_unified_iswa(
|
||||
const llama_model & model,
|
||||
ggml_type type_k,
|
||||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
uint32_t kv_size,
|
||||
bool swa_full,
|
||||
uint32_t n_seq_max,
|
||||
uint32_t n_batch,
|
||||
uint32_t padding);
|
||||
|
||||
~llama_kv_cache_unified_iswa() = 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_min(llama_seq_id seq_id) const 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 & 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;
|
||||
|
||||
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;
|
||||
|
||||
// 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;
|
||||
|
||||
//
|
||||
// llama_kv_cache_unified_iswa specific API
|
||||
//
|
||||
|
||||
llama_kv_cache_unified * get_kv_base() const;
|
||||
llama_kv_cache_unified * get_kv_swa () const;
|
||||
|
||||
private:
|
||||
const llama_hparams & hparams;
|
||||
|
||||
bool do_prune = true;
|
||||
|
||||
struct {
|
||||
struct entry {
|
||||
llama_pos pmin;
|
||||
llama_pos pmax;
|
||||
};
|
||||
|
||||
void clear() {
|
||||
pos.clear();
|
||||
}
|
||||
|
||||
// used to perform SWA pruning of old tokens
|
||||
std::unordered_map<llama_seq_id, entry> pos;
|
||||
} pending;
|
||||
|
||||
std::unique_ptr<llama_kv_cache_unified> kv_base;
|
||||
std::unique_ptr<llama_kv_cache_unified> kv_swa;
|
||||
};
|
||||
|
||||
//
|
||||
// llama_kv_cache_recurrent
|
||||
//
|
||||
@@ -302,6 +448,7 @@ public:
|
||||
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_min(llama_seq_id seq_id) const override;
|
||||
llama_pos seq_pos_max(llama_seq_id seq_id) const override;
|
||||
|
||||
//
|
||||
@@ -318,7 +465,6 @@ public:
|
||||
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;
|
||||
|
||||
Reference in New Issue
Block a user