mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-14 10:59:41 +00:00
llama : add high-throughput mode (#14363)
* kv-cache : prepare K/V buffers for separation ggml-ci * batched-bench : fix oob write ggml-ci * llama : add "virtual sequences" ggml-ci * llama : use "stream" vs "virtual sequence" ggml-ci * graph : fix stream splitting when KV cache is not used ggml-ci * kv-cache : add multi-stream save/load support ggml-ci * llama : add "--attn-streams" flag ggml-ci * kv-cache : fix handling when find_slot fails ggml-ci * kv-cache : restore find_slot impl ggml-ci * kv-cache : add comments * kv-cache : add bounds checks for sequence id ggml-ci * cont : add n_seq_max to batch allocr ggml-ci * kv-cache : perform stream copies lazily after llama_synchronize ggml-ci * kv-cache : avoid throwing exceptions across the C boundary ggml-ci * CUDA: 4D FlashAttention support (#14628) * CUDA: 4D FlashAttention support * CUDA: fix WMMA FA kernel * llama : rename attn_streams -> kv_unified ggml-ci * common : rename kv_split -> kv_unified ggml-ci --------- Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Georgi Gerganov
GitHub
parent
ab14019821
commit
225e7a1438
30 changed files with 1080 additions and 460 deletions
|
|
@ -35,16 +35,50 @@ public:
|
|||
std::vector<uint32_t> ids;
|
||||
};
|
||||
|
||||
struct stream_copy_info {
|
||||
bool empty() const {
|
||||
assert(ssrc.size() == sdst.size());
|
||||
return ssrc.empty();
|
||||
}
|
||||
|
||||
std::vector<uint32_t> ssrc;
|
||||
std::vector<uint32_t> sdst;
|
||||
};
|
||||
|
||||
// for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
|
||||
// KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
|
||||
struct slot_info {
|
||||
// data for ggml_set_rows
|
||||
using idx_vec_t = std::vector<uint32_t>;
|
||||
|
||||
idx_vec_t idxs;
|
||||
// number of streams: ns = s1 - s0 + 1
|
||||
llama_seq_id s0;
|
||||
llama_seq_id s1;
|
||||
|
||||
std::vector<llama_seq_id> strm; // [ns]
|
||||
std::vector<idx_vec_t> idxs; // [ns]
|
||||
|
||||
uint32_t head() const {
|
||||
return idxs.at(0);
|
||||
GGML_ASSERT(idxs.size() == 1);
|
||||
GGML_ASSERT(!idxs[0].empty());
|
||||
|
||||
return idxs[0][0];
|
||||
}
|
||||
|
||||
void resize(size_t n) {
|
||||
strm.resize(n);
|
||||
idxs.resize(n);
|
||||
}
|
||||
|
||||
size_t size() const {
|
||||
GGML_ASSERT(idxs.size() == strm.size());
|
||||
GGML_ASSERT(!idxs.empty());
|
||||
|
||||
return idxs[0].size();
|
||||
}
|
||||
|
||||
size_t n_stream() const {
|
||||
return strm.size();
|
||||
}
|
||||
|
||||
bool empty() const {
|
||||
|
|
@ -54,9 +88,6 @@ public:
|
|||
void clear() {
|
||||
idxs.clear();
|
||||
}
|
||||
|
||||
// TODO: implement
|
||||
//std::vector<idx_vec_t> seq_idxs;
|
||||
};
|
||||
|
||||
using slot_info_vec_t = std::vector<slot_info>;
|
||||
|
|
@ -68,6 +99,7 @@ public:
|
|||
ggml_type type_v,
|
||||
bool v_trans,
|
||||
bool offload,
|
||||
bool unified,
|
||||
uint32_t kv_size,
|
||||
uint32_t n_seq_max,
|
||||
uint32_t n_pad,
|
||||
|
|
@ -111,7 +143,8 @@ public:
|
|||
// llama_kv_cache_unified specific API
|
||||
//
|
||||
|
||||
uint32_t get_size() const;
|
||||
uint32_t get_size() const;
|
||||
uint32_t get_n_stream() const;
|
||||
|
||||
bool get_has_shift() const;
|
||||
|
||||
|
|
@ -122,8 +155,8 @@ public:
|
|||
uint32_t get_n_kv() const;
|
||||
|
||||
// get views of the current state of the cache
|
||||
ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
|
||||
ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
|
||||
ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
|
||||
ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
|
||||
|
||||
// store k_cur and v_cur in the cache based on the provided head location
|
||||
ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
|
||||
|
|
@ -137,7 +170,7 @@ public:
|
|||
// return empty vector on failure
|
||||
slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
|
||||
|
||||
bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
|
||||
bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo, const stream_copy_info & sc_info);
|
||||
|
||||
// find a slot of kv cells that can hold the ubatch
|
||||
// if cont == true, then the slot must be continuous
|
||||
|
|
@ -157,8 +190,9 @@ public:
|
|||
void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
|
||||
void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
|
||||
|
||||
void set_input_k_shift(ggml_tensor * dst) const;
|
||||
|
||||
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:
|
||||
|
|
@ -172,15 +206,15 @@ private:
|
|||
|
||||
ggml_tensor * k;
|
||||
ggml_tensor * v;
|
||||
|
||||
std::vector<ggml_tensor *> k_stream;
|
||||
std::vector<ggml_tensor *> v_stream;
|
||||
};
|
||||
|
||||
bool v_trans = true; // the value tensor is transposed
|
||||
|
||||
// the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
|
||||
// note: this is not part of the KV state and it's only used to speed-up the find_slot() method
|
||||
uint32_t head = 0;
|
||||
|
||||
const uint32_t n_seq_max = 1;
|
||||
const uint32_t n_stream = 1;
|
||||
|
||||
// required padding
|
||||
const uint32_t n_pad = 1;
|
||||
|
|
@ -200,7 +234,17 @@ private:
|
|||
std::vector<ggml_context_ptr> ctxs;
|
||||
std::vector<ggml_backend_buffer_ptr> bufs;
|
||||
|
||||
llama_kv_cells_unified cells;
|
||||
// the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
|
||||
// note: this is not part of the KV state and it's only used to speed-up the find_slot() method
|
||||
std::vector<uint32_t> v_heads;
|
||||
|
||||
std::vector<llama_kv_cells_unified> v_cells;
|
||||
|
||||
// maps from a sequence id to a stream id
|
||||
std::vector<uint32_t> seq_to_stream;
|
||||
|
||||
// pending stream copies that will be applied during the next update
|
||||
stream_copy_info sc_info;
|
||||
|
||||
std::vector<kv_layer> layers;
|
||||
|
||||
|
|
@ -237,18 +281,25 @@ private:
|
|||
ggml_cgraph * gf,
|
||||
const defrag_info & dinfo) 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;
|
||||
struct cell_ranges_t {
|
||||
uint32_t strm;
|
||||
|
||||
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);
|
||||
std::vector<std::pair<uint32_t, uint32_t>> data; // ranges, from inclusive, to exclusive
|
||||
};
|
||||
|
||||
void state_write_meta(llama_io_write_i & io, const cell_ranges_t & cr, llama_seq_id seq_id = -1) const;
|
||||
void state_write_data(llama_io_write_i & io, const cell_ranges_t & cr) const;
|
||||
|
||||
bool state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id = -1);
|
||||
bool state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count);
|
||||
};
|
||||
|
||||
class llama_kv_cache_unified_context : public llama_memory_context_i {
|
||||
public:
|
||||
// some shorthands
|
||||
using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
|
||||
using defrag_info = llama_kv_cache_unified::defrag_info;
|
||||
using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
|
||||
using defrag_info = llama_kv_cache_unified::defrag_info;
|
||||
using stream_copy_info = llama_kv_cache_unified::stream_copy_info;
|
||||
|
||||
// used for errors
|
||||
llama_kv_cache_unified_context(llama_memory_status status);
|
||||
|
|
@ -262,7 +313,8 @@ public:
|
|||
llama_kv_cache_unified * kv,
|
||||
llama_context * lctx,
|
||||
bool do_shift,
|
||||
defrag_info dinfo);
|
||||
defrag_info dinfo,
|
||||
stream_copy_info sc_info);
|
||||
|
||||
// used to create a batch procesing context from a batch
|
||||
llama_kv_cache_unified_context(
|
||||
|
|
@ -320,6 +372,8 @@ private:
|
|||
|
||||
defrag_info dinfo;
|
||||
|
||||
stream_copy_info sc_info;
|
||||
|
||||
//
|
||||
// batch processing context
|
||||
//
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue