Skip to content

Commit ec7695f

Browse files
committed
feat: First pass at llama_kv_cache_hybrid_recurrent
This follows the pattern in iswa where the two child caches are held explicitly to support the case where a model requires a single attention cache and a single recurrent cache where each layer uses exactly one of the caches. This is a rewrite of the more generic approach in the original hybrid cache PR: #13276 Branch: HybridRecurrentCache Signed-off-by: Gabe Goodhart <[email protected]>
1 parent 4470221 commit ec7695f

File tree

2 files changed

+464
-2
lines changed

2 files changed

+464
-2
lines changed

src/llama-kv-cache.cpp

Lines changed: 238 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -1733,8 +1733,8 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
17331733
uint32_t n_seq_max,
17341734
uint32_t n_batch,
17351735
uint32_t n_pad) : hparams(model.hparams) {
1736-
llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
1737-
llama_kv_cache_unified::layer_filter_cb filter_swa = [&](int32_t il) { return model.hparams.is_swa(il); };
1736+
llama_kv_cache::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
1737+
llama_kv_cache::layer_filter_cb filter_swa = [&](int32_t il) { return model.hparams.is_swa(il); };
17381738

17391739
const uint32_t size_base = kv_size;
17401740

@@ -3082,3 +3082,239 @@ int32_t llama_kv_cache_recurrent_state::s_copy(int i) const {
30823082
float llama_kv_cache_recurrent_state::s_mask(int i) const {
30833083
return kv->s_mask(i);
30843084
}
3085+
3086+
//
3087+
// llama_kv_cache_hybrid_recurrent
3088+
//
3089+
3090+
llama_kv_cache_hybrid_recurrent::llama_kv_cache_hybrid_recurrent(
3091+
const llama_model & model,
3092+
/* attn */
3093+
ggml_type attn_type_k,
3094+
ggml_type attn_type_v,
3095+
bool attn_v_trans,
3096+
uint32_t attn_kv_size,
3097+
uint32_t attn_n_pad,
3098+
uint32_t attn_n_swa,
3099+
llama_swa_type attn_swa_type,
3100+
/* recurrent */
3101+
ggml_type recurrent_type_k,
3102+
ggml_type recurrent_type_v,
3103+
uint32_t recurrent_kv_size,
3104+
/* common */
3105+
uint32_t n_seq_max,
3106+
bool offload) :
3107+
hparams(model.hparams),
3108+
kv_attn(new llama_kv_cache_unified(
3109+
model,
3110+
[&](int32_t il) { return !model.hparams.recurrent_layer(il); },
3111+
attn_type_k,
3112+
attn_type_v,
3113+
attn_v_trans,
3114+
offload,
3115+
attn_kv_size,
3116+
n_seq_max,
3117+
attn_n_pad,
3118+
attn_n_swa,
3119+
attn_swa_type
3120+
)),
3121+
kv_recurrent(new llama_kv_cache_recurrent(
3122+
model,
3123+
[&](int32_t il) { return model.hparams.recurrent_layer(il); },
3124+
recurrent_type_k,
3125+
recurrent_type_v,
3126+
offload,
3127+
recurrent_kv_size,
3128+
n_seq_max
3129+
)) {}
3130+
3131+
void llama_kv_cache_hybrid_recurrent::clear() {
3132+
kv_attn ->clear();
3133+
kv_recurrent->clear();
3134+
}
3135+
3136+
bool llama_kv_cache_hybrid_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
3137+
// Try removing from the recurrent cache first since it may fail. If it does
3138+
// fail, the cache will not have been mutated.
3139+
if (!kv_recurrent->seq_rm(seq_id, p0, p1)) {
3140+
return false;
3141+
}
3142+
return kv_attn->seq_rm(seq_id, p0, p1);
3143+
}
3144+
3145+
void llama_kv_cache_hybrid_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
3146+
kv_attn ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
3147+
kv_recurrent->seq_cp(seq_id_src, seq_id_dst, p0, p1);
3148+
}
3149+
3150+
void llama_kv_cache_hybrid_recurrent::seq_keep(llama_seq_id seq_id) {
3151+
kv_attn ->seq_keep(seq_id);
3152+
kv_recurrent->seq_keep(seq_id);
3153+
}
3154+
3155+
void llama_kv_cache_hybrid_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
3156+
kv_attn->seq_add(seq_id, p0, p1, shift);
3157+
kv_recurrent->seq_add(seq_id, p0, p1, shift);
3158+
}
3159+
3160+
void llama_kv_cache_hybrid_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
3161+
kv_attn ->seq_div(seq_id, p0, p1, d);
3162+
kv_recurrent->seq_div(seq_id, p0, p1, d);
3163+
}
3164+
3165+
llama_pos llama_kv_cache_hybrid_recurrent::seq_pos_min(llama_seq_id seq_id) const {
3166+
// the min of the total cache is the max of the two caches' min values
3167+
return std::max(kv_attn->seq_pos_min(seq_id), kv_recurrent->seq_pos_min(seq_id));
3168+
}
3169+
3170+
llama_pos llama_kv_cache_hybrid_recurrent::seq_pos_max(llama_seq_id seq_id) const {
3171+
// the max of the total cache is the min of the two caches' max values
3172+
return std::min(kv_attn->seq_pos_max(seq_id), kv_recurrent->seq_pos_max(seq_id));
3173+
}
3174+
3175+
llama_memory_state_ptr llama_kv_cache_hybrid_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
3176+
3177+
// since this includes a recurrent cache, we cannot use split_simple
3178+
auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
3179+
3180+
// follow the recurrent pattern for creating the ubatch splits
3181+
std::vector<llama_ubatch> ubatches;
3182+
while (sbatch.n_tokens > 0) {
3183+
llama_ubatch ubatch;
3184+
3185+
if (embd_pooled) {
3186+
// Pooled embeddings cannot be split across ubatches (yet)
3187+
ubatch = sbatch.split_seq(n_ubatch);
3188+
} else {
3189+
ubatch = sbatch.split_equal(n_ubatch);
3190+
}
3191+
3192+
ubatches.push_back(ubatch);
3193+
}
3194+
3195+
// prepare the recurrent batches first
3196+
if (!kv_recurrent->prepare(ubatches)) {
3197+
// TODO: will the recurrent cache be in an undefined state at this point?
3198+
LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
3199+
return std::make_unique<llama_kv_cache_hybrid_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
3200+
}
3201+
3202+
// prepare the attention cache
3203+
auto heads_attn = kv_attn->prepare(ubatches);
3204+
if (heads_attn.empty()) {
3205+
LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__);
3206+
return std::make_unique<llama_kv_cache_hybrid_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
3207+
}
3208+
3209+
return std::make_unique<llama_kv_cache_hybrid_recurrent_state>(
3210+
this, std::move(sbatch), std::move(heads_attn), std::move(ubatches));
3211+
}
3212+
3213+
llama_memory_state_ptr llama_kv_cache_hybrid_recurrent::init_full() {
3214+
return std::make_unique<llama_kv_cache_hybrid_recurrent_state>(this);
3215+
}
3216+
3217+
bool llama_kv_cache_hybrid_recurrent::update(llama_context & lctx) {
3218+
bool res = false;
3219+
3220+
res = res | kv_attn ->update(lctx);
3221+
res = res | kv_recurrent->update(lctx);
3222+
3223+
return res;
3224+
}
3225+
3226+
void llama_kv_cache_hybrid_recurrent::defrag_sched(float thold) {
3227+
kv_attn ->defrag_sched(thold);
3228+
kv_recurrent->defrag_sched(thold);
3229+
}
3230+
3231+
bool llama_kv_cache_hybrid_recurrent::get_can_shift() const {
3232+
// TODO: Should this return true if the attention cache can shift?
3233+
return false;
3234+
}
3235+
3236+
void llama_kv_cache_hybrid_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
3237+
kv_attn ->state_write(io, seq_id);
3238+
kv_recurrent->state_write(io, seq_id);
3239+
}
3240+
3241+
void llama_kv_cache_hybrid_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
3242+
kv_attn ->state_read(io, seq_id);
3243+
kv_recurrent->state_read(io, seq_id);
3244+
}
3245+
3246+
llama_kv_cache_unified * llama_kv_cache_hybrid_recurrent::get_kv_attn() const {
3247+
return kv_attn.get();
3248+
}
3249+
3250+
llama_kv_cache_recurrent * llama_kv_cache_hybrid_recurrent::get_kv_recurrent() const {
3251+
return kv_recurrent.get();
3252+
}
3253+
3254+
llama_kv_cache_hybrid_recurrent_state::llama_kv_cache_hybrid_recurrent_state(llama_memory_status status)
3255+
: status(status), state_attn(status), state_recurrent(status) {}
3256+
3257+
llama_kv_cache_hybrid_recurrent_state::llama_kv_cache_hybrid_recurrent_state(llama_kv_cache_hybrid_recurrent * kv)
3258+
: status(LLAMA_MEMORY_STATUS_SUCCESS),
3259+
kv(kv),
3260+
state_attn(status, kv->get_kv_attn()),
3261+
state_recurrent(status, kv->get_kv_recurrent()) {}
3262+
3263+
llama_kv_cache_hybrid_recurrent_state::llama_kv_cache_hybrid_recurrent_state(
3264+
llama_kv_cache_hybrid_recurrent * kv,
3265+
llama_sbatch sbatch,
3266+
std::vector<uint32_t> heads_attn,
3267+
std::vector<llama_ubatch> ubatches)
3268+
: status(LLAMA_MEMORY_STATUS_SUCCESS),
3269+
kv(kv),
3270+
sbatch(std::move(sbatch)),
3271+
heads_attn(std::move(heads_attn)),
3272+
ubatches(std::move(ubatches)),
3273+
// NOTE: these child states are only used as wrapper APIs for the
3274+
// const methods, so we use the "init full" signature since the
3275+
// actual state is not used.
3276+
state_attn(LLAMA_MEMORY_STATUS_SUCCESS, kv->get_kv_attn()),
3277+
state_recurrent(LLAMA_MEMORY_STATUS_SUCCESS, kv->get_kv_recurrent()) {}
3278+
3279+
3280+
bool llama_kv_cache_hybrid_recurrent_state::next() {
3281+
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
3282+
3283+
if (++i_next >= ubatches.size()) {
3284+
return false;
3285+
}
3286+
3287+
return true;
3288+
}
3289+
3290+
bool llama_kv_cache_hybrid_recurrent_state::apply() {
3291+
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
3292+
3293+
kv->get_kv_attn() ->apply_ubatch(heads_attn[i_next], ubatches[i_next]);
3294+
kv->get_kv_recurrent()->find_slot(ubatches[i_next]);
3295+
3296+
return true;
3297+
}
3298+
3299+
std::vector<int64_t> & llama_kv_cache_hybrid_recurrent_state::out_ids() {
3300+
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
3301+
3302+
return sbatch.out_ids;
3303+
}
3304+
3305+
llama_memory_status llama_kv_cache_hybrid_recurrent_state::get_status() const {
3306+
return status;
3307+
}
3308+
3309+
const llama_ubatch & llama_kv_cache_hybrid_recurrent_state::get_ubatch() const {
3310+
assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
3311+
return ubatches[i_next];
3312+
}
3313+
3314+
const llama_kv_cache_unified_state * llama_kv_cache_hybrid_recurrent_state::get_state_attn () const {
3315+
return &state_attn;
3316+
}
3317+
3318+
const llama_kv_cache_recurrent_state * llama_kv_cache_hybrid_recurrent_state::get_state_recurrent() const {
3319+
return &state_recurrent;
3320+
}

0 commit comments

Comments
 (0)