1#include "llama-memory-hybrid-iswa.h"
  2
  3#include "llama-impl.h"
  4#include "llama-model.h"
  5#include "llama-context.h"
  6
  7//
  8// llama_memory_hybrid_iswa
  9//
 10
 11llama_memory_hybrid_iswa::llama_memory_hybrid_iswa(
 12        const llama_model & model,
 13                            /* attn */
 14                ggml_type   type_k,
 15                ggml_type   type_v,
 16                     bool   v_trans,
 17                     bool   swa_full,
 18                 uint32_t   kv_size,
 19                 uint32_t   n_ubatch,
 20                 uint32_t   n_pad,
 21                            /* recurrent */
 22                ggml_type   type_r,
 23                ggml_type   type_s,
 24                 uint32_t   rs_size,
 25                            /* common */
 26                 uint32_t   n_seq_max,
 27                     bool   offload,
 28                     bool   unified,
 29                            /* layer filters */
 30    const layer_filter_cb & filter_attn,
 31    const layer_filter_cb & filter_recr) :
 32    hparams(model.hparams),
 33    mem_attn(new llama_kv_cache_iswa(
 34        model,
 35        type_k,
 36        type_v,
 37        v_trans,
 38        offload,
 39        swa_full,
 40        unified,
 41        kv_size,
 42        n_seq_max,
 43        n_ubatch,
 44        n_pad,
 45        filter_attn == nullptr ?
 46            [&](int32_t il) { return !hparams.is_recurrent(il); }
 47            : filter_attn,
 48        nullptr
 49    )),
 50    mem_recr(new llama_memory_recurrent(
 51        model,
 52        type_r,
 53        type_s,
 54        offload,
 55        rs_size,
 56        n_seq_max,
 57        filter_recr == nullptr ?
 58            [&](int32_t il) { return hparams.is_recurrent(il); }
 59            : filter_recr
 60    )) {}
 61
 62llama_memory_context_ptr llama_memory_hybrid_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
 63    do {
 64        balloc.split_reset();
 65
 66        // follow the recurrent pattern for creating the ubatch splits
 67        std::vector<llama_ubatch> ubatches;
 68
 69        while (true) {
 70            llama_ubatch ubatch;
 71
 72            if (embd_all) {
 73                // if all tokens are output, split by sequence
 74                ubatch = balloc.split_seq(n_ubatch);
 75            } else {
 76                // TODO: non-sequential equal split can be done if using unified KV cache
 77                //       for simplicity, we always use sequential equal split for now
 78                ubatch = balloc.split_equal(n_ubatch, true);
 79            }
 80
 81            if (ubatch.n_tokens == 0) {
 82                break;
 83            }
 84
 85            ubatches.push_back(std::move(ubatch)); // NOLINT
 86        }
 87
 88        if (balloc.get_n_used() < balloc.get_n_tokens()) {
 89            // failed to find a suitable split
 90            break;
 91        }
 92
 93        // prepare the recurrent batches first
 94        if (!mem_recr->prepare(ubatches)) {
 95            // TODO: will the recurrent cache be in an undefined context at this point?
 96            LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
 97            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 98        }
 99
100        // prepare the attention cache (iswa version returns both base and swa slot infos)
101        auto sinfos_base = mem_attn->get_base()->prepare(ubatches);
102        if (sinfos_base.empty()) {
103            LLAMA_LOG_ERROR("%s: failed to prepare attention base ubatches\n", __func__);
104            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
105        }
106
107        auto sinfos_swa = mem_attn->get_swa()->prepare(ubatches);
108        if (sinfos_swa.empty()) {
109            LLAMA_LOG_ERROR("%s: failed to prepare attention swa ubatches\n", __func__);
110            return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
111        }
112
113        return std::make_unique<llama_memory_hybrid_iswa_context>(
114                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
115    } while(false);
116
117    return std::make_unique<llama_memory_hybrid_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
118}
119
120llama_memory_context_ptr llama_memory_hybrid_iswa::init_full() {
121    return std::make_unique<llama_memory_hybrid_iswa_context>(this);
122}
123
124llama_memory_context_ptr llama_memory_hybrid_iswa::init_update(llama_context * lctx, bool optimize) {
125    return std::make_unique<llama_memory_hybrid_iswa_context>(this, lctx, optimize);
126}
127
128bool llama_memory_hybrid_iswa::get_can_shift() const {
129    // Shifting is trivially supported for recurrent
130    return mem_attn->get_can_shift();
131}
132
133void llama_memory_hybrid_iswa::clear(bool data) {
134    mem_attn->clear(data);
135    mem_recr->clear(data);
136}
137
138bool llama_memory_hybrid_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
139    // Try removing from the recurrent cache first since it may fail. If it does
140    // fail, the cache will not have been mutated.
141    if (!mem_recr->seq_rm(seq_id, p0, p1)) {
142        return false;
143    }
144    return mem_attn->seq_rm(seq_id, p0, p1);
145}
146
147void llama_memory_hybrid_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
148    mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
149    mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
150}
151
152void llama_memory_hybrid_iswa::seq_keep(llama_seq_id seq_id) {
153    mem_attn->seq_keep(seq_id);
154    mem_recr->seq_keep(seq_id);
155}
156
157void llama_memory_hybrid_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
158    mem_attn->seq_add(seq_id, p0, p1, shift);
159    mem_recr->seq_add(seq_id, p0, p1, shift);
160}
161
162void llama_memory_hybrid_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
163    mem_attn->seq_div(seq_id, p0, p1, d);
164    mem_recr->seq_div(seq_id, p0, p1, d);
165}
166
167llama_pos llama_memory_hybrid_iswa::seq_pos_min(llama_seq_id seq_id) const {
168    // the min of the total cache is the max of the two caches' min values
169    return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
170}
171
172llama_pos llama_memory_hybrid_iswa::seq_pos_max(llama_seq_id seq_id) const {
173    // the max of the total cache is the min of the two caches' max values
174    return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
175}
176
177std::map<ggml_backend_buffer_type_t, size_t> llama_memory_hybrid_iswa::memory_breakdown() const {
178    std::map<ggml_backend_buffer_type_t, size_t> mb = mem_attn->memory_breakdown();
179    for (const auto & buft_size : mem_recr->memory_breakdown()) {
180        mb[buft_size.first] += buft_size.second;
181    }
182    return mb;
183}
184
185void llama_memory_hybrid_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
186    mem_attn->state_write(io, seq_id, flags);
187    mem_recr->state_write(io, seq_id, flags);
188}
189
190void llama_memory_hybrid_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
191    mem_attn->state_read(io, seq_id, flags);
192    mem_recr->state_read(io, seq_id, flags);
193}
194
195llama_kv_cache_iswa * llama_memory_hybrid_iswa::get_mem_attn() const {
196    return mem_attn.get();
197}
198
199llama_memory_recurrent * llama_memory_hybrid_iswa::get_mem_recr() const {
200    return mem_recr.get();
201}
202
203//
204// llama_memory_hybrid_iswa_context
205//
206
207llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_status status) : status(status) {}
208
209llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(llama_memory_hybrid_iswa * mem) :
210    ctx_attn(mem->get_mem_attn()->init_full()),
211    ctx_recr(mem->get_mem_recr()->init_full()),
212    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
213}
214
215llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
216        llama_memory_hybrid_iswa * mem,
217                   llama_context * lctx,
218                            bool   optimize) :
219    ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
220    ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
221    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
222}
223
224llama_memory_hybrid_iswa_context::llama_memory_hybrid_iswa_context(
225           llama_memory_hybrid_iswa * mem,
226                    slot_info_vec_t   sinfos_base,
227                    slot_info_vec_t   sinfos_swa,
228          std::vector<llama_ubatch>   ubatches) :
229    ubatches(std::move(ubatches)),
230    // note: here we copy the ubatches. not sure if this is ideal
231    ctx_attn(new llama_kv_cache_iswa_context(mem->get_mem_attn(), std::move(sinfos_base), std::move(sinfos_swa), this->ubatches)),
232    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(), this->ubatches)),
233    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
234}
235
236bool llama_memory_hybrid_iswa_context::next() {
237    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
238
239    ctx_attn->next();
240    ctx_recr->next();
241
242    if (++i_next >= ubatches.size()) {
243        return false;
244    }
245
246    return true;
247}
248
249bool llama_memory_hybrid_iswa_context::apply() {
250    assert(!llama_memory_status_is_fail(status));
251
252    bool res = true;
253
254    res = res & ctx_attn->apply();
255    res = res & ctx_recr->apply();
256
257    return res;
258}
259
260llama_memory_status llama_memory_hybrid_iswa_context::get_status() const {
261    return status;
262}
263
264const llama_ubatch & llama_memory_hybrid_iswa_context::get_ubatch() const {
265    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
266    return ubatches[i_next];
267}
268
269const llama_kv_cache_iswa_context * llama_memory_hybrid_iswa_context::get_attn() const {
270    return static_cast<const llama_kv_cache_iswa_context *>(ctx_attn.get());
271}
272
273const llama_memory_recurrent_context * llama_memory_hybrid_iswa_context::get_recr() const {
274    return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
275}