Files
llama.cpp/tools/server/server-slot-sched.cpp
T
Xuan-Son NguyenandPascal 52975342ff server: slot_sched: fix lora and recurrent (#32434)
* server: fix lora and recurrent

* tests: run the hybrid SWA checkpoint test with one token per ubatch

With and without the cache, the prompt was not split into the same
ubatches, and the generated tokens are decoded one at a time while the
uncached run processes them in a batch. On CUDA this moves the top
logprobs of the tiny model by up to 0.03 and swaps near ties, so the
test failed while the cache behaved as expected.

With one token per ubatch both runs go through the same kernels and
the output is compared exactly.

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-11-18 00:12:20 +01:00

433 lines
15 KiB
C++

#include "server-slot-sched.h"
#include "common.h"
#include "llama.h"
#include "log.h"
#include <algorithm>
//
// server_prompt
//
server_prompt_reuse server_prompt::get_reuse(int32_t n_prefix, bool has_new_tokens, int32_t n_swa, llama_pos pos_min) const {
server_prompt_reuse res;
res.n_past = n_prefix;
res.pos_next = tokens.pos_next(n_prefix);
// the first evaluated token attends to the positions > pos_eval - n_swa, so a SWA state
// starting at pos_eval - n_swa + 1 is enough, while a state without SWA must end before pos_eval
const llama_pos pos_eval = res.pos_next - (has_new_tokens ? 0 : 1);
// a state or a checkpoint is useful when its pos_min is below this threshold
const auto pos_min_thold = std::max(0, n_swa > 0 ? pos_eval - n_swa + 2 : pos_eval);
if (pos_min < pos_min_thold) {
return res;
}
// search for a context checkpoint
const auto it = std::find_if(
checkpoints.rbegin(),
checkpoints.rend(),
[&](const auto & ckpt) {
const auto & cur = *ckpt;
// guarantee that a checkpoint will result in at least one token being processed [TAG_PROMPT_LOGITS]
SRV_TRC("checking checkpoint with [%d, %d] against %d...\n", cur.pos_min, cur.pos_max, pos_min_thold);
// workaround for [TAG_CHECKPOINTS_FIX_POS_MIN]
if (cur.pos_max > res.pos_next) {
return false;
}
return cur.pos_min < pos_min_thold || cur.pos_min == 0;
}
);
if (it == checkpoints.rend()) {
res.n_past = 0;
res.pos_next = 0;
return res;
}
const auto & cur = **it;
res.pos_next = std::min(res.pos_next, std::max(cur.pos_min + 1, cur.pos_max));
res.n_past = std::min(tokens.size_up_to_pos(res.pos_next), (size_t) cur.n_tokens);
res.ckpt = *it;
return res;
}
//
// server_prompt_cache
//
size_t server_prompt_cache::size() const {
size_t res = 0;
for (const auto & state : states) {
res += state.size();
}
return res;
}
size_t server_prompt_cache::n_tokens() const {
size_t res = 0;
for (const auto & state : states) {
res += state.prompt.n_tokens();
}
return res;
}
server_prompt_cache_state * server_prompt_cache::alloc(const server_prompt & prompt, size_t state_size_tgt, size_t state_size_dft) {
// first check if the current state is contained fully in the cache
for (auto it = states.begin(); it != states.end(); ++it) {
const int cur_lcp_len = it->prompt.tokens.get_common_prefix(prompt.tokens);
if (cur_lcp_len == (int) prompt.tokens.size() && are_lora_equal(it->prompt.lora, prompt.lora)) {
SRV_TRC("%s", " - prompt is already in the cache, skipping\n");
return nullptr;
}
}
// calculate checkpoints size to see if it will fit with the prompt
size_t checkpoints_size = 0;
for (const auto & ckpt : prompt.checkpoints) {
checkpoints_size += ckpt->size();
}
const size_t state_size_new = state_size_tgt + state_size_dft + checkpoints_size;
// skip over-limit entries to avoid disturbing the cache
if (limit_size > 0 && state_size_new > limit_size) {
SRV_WRN(" - prompt state size %.3f MiB exceeds cache size limit %.3f MiB, skipping\n",
state_size_new / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0));
return nullptr;
}
// remove any cached prompts that are fully contained in the current prompt
for (auto it = states.begin(); it != states.end();) {
const int len = it->prompt.tokens.get_common_prefix(prompt.tokens);
if (len == (int) it->prompt.tokens.size() && are_lora_equal(it->prompt.lora, prompt.lora)) {
SRV_TRC(" - removing obsolete cached prompt with length %d\n", len);
it = states.erase(it);
} else {
++it;
}
}
if (limit_size > 0) {
// make room before allocating the new vectors to avoid breaching the limit
while (!states.empty() && size() + state_size_new > limit_size) {
SRV_WRN(" - making room for prompt cache entry, removing oldest entry (size = %.3f MiB)\n",
states.front().size() / (1024.0 * 1024.0));
states.pop_front();
}
}
common_state_data state_data_tgt;
common_state_data state_data_dft;
// check if we can allocate enough memory for the new state
try {
state_data_tgt.resize(state_size_tgt);
state_data_dft.resize(state_size_dft);
} catch (const std::bad_alloc & e) {
SRV_ERR("failed to allocate memory for prompt cache state: %s\n", e.what());
limit_size = std::max<size_t>(1, 0.4*size());
SRV_WRN(" - cache size limit reduced to %.3f MiB\n", limit_size / (1024.0 * 1024.0));
update();
return nullptr;
}
states.push_back({
/*.prompt =*/ {
/*.tokens =*/ prompt.tokens.clone(),
/*.checkpoints =*/ prompt.checkpoints,
/*.lora =*/ prompt.lora,
},
/*.data =*/ {
/*.main =*/ std::move(state_data_tgt),
/*.drft =*/ std::move(state_data_dft),
},
});
return &states.back();
}
bool server_prompt_cache::load(server_prompt & prompt, const server_tokens & tokens_new, const std::vector<common_adapter_lora_info> & lora_new, llama_context * ctx_tgt, llama_context * ctx_dft, int32_t id_slot) {
// a prompt made with another lora config is cleared before the new one starts, so it counts as empty
const bool is_empty = prompt.tokens.empty() || lora_should_clear_cache(prompt.lora, lora_new);
int lcp_best = is_empty ? 0 : prompt.tokens.get_common_prefix(tokens_new);
float f_keep_best = is_empty ? -1.0f : float(lcp_best) / prompt.tokens.size(); // empty slot: any cache entry wins
float f_sim_best = float(lcp_best) / tokens_new.size();
SRV_TRC(" - looking for better prompt, base f_keep = %.3f, f_sim = %.3f\n", f_keep_best, f_sim_best);
auto it_best = states.end();
// find the most similar cached prompt, that would also preserve the most context
for (auto it = states.begin(); it != states.end(); ++it) {
if (lora_should_clear_cache(it->prompt.lora, lora_new)) {
SRV_TRC(" - prompt with length %7zu, skipping, other lora config\n", it->prompt.tokens.size());
continue;
}
const int lcp_cur = it->prompt.tokens.get_common_prefix(tokens_new);
const float f_keep_cur = float(lcp_cur) / it->prompt.tokens.size();
const float f_sim_cur = float(lcp_cur) / tokens_new.size();
SRV_TRC(" - prompt with length %7zu, lcp = %7d, f_keep = %.3f, f_sim = %.3f\n", it->prompt.tokens.size(), lcp_cur, f_keep_cur, f_sim_cur);
// don't trash large prompts
if (f_keep_cur < 0.25f) {
continue;
}
if (f_keep_best < f_keep_cur && f_sim_best < f_sim_cur) {
f_keep_best = f_keep_cur;
f_sim_best = f_sim_cur;
lcp_best = lcp_cur;
it_best = it;
}
}
if (it_best != states.end()) {
SRV_TRC(" - found better prompt with f_keep = %.3f, f_sim = %.3f\n", f_keep_best, f_sim_best);
{
auto & data = it_best->data.main;
const size_t size = data.size();
const size_t n = llama_state_seq_set_data_ext(ctx_tgt, data.data(), size, id_slot, 0);
if (n != size) {
SRV_ERR("failed to restore state with size %zu\n", size);
return false;
}
data.clear();
data.shrink_to_fit();
}
{
auto & data = it_best->data.drft;
if (!data.empty()) {
GGML_ASSERT(ctx_dft);
const size_t size = data.size();
const size_t n = llama_state_seq_set_data_ext(ctx_dft, data.data(), size, id_slot, 0);
if (n != size) {
SRV_WRN("failed to restore state with size %zu\n", size);
return false;
}
data.clear();
data.shrink_to_fit();
}
}
SRV_DBG("%s", "__TEST_TAG_PROMPT_CACHE_RESTORE__\n");
// the whole cached prompt is a prefix of the new one
if (lcp_best == (int) it_best->prompt.tokens.size()) {
SRV_DBG("%s", "__TEST_TAG_PROMPT_CACHE_FULL_RESTORE__\n");
}
prompt = std::move(it_best->prompt);
states.erase(it_best);
}
return true;
}
void server_prompt_cache::update() {
if (limit_size > 0) {
while (!states.empty() && size() > limit_size) {
SRV_WRN(" - cache size limit reached, removing oldest entry (size = %.3f MiB)\n", states.front().size() / (1024.0 * 1024.0));
states.pop_front();
}
}
// average size per token
const float size_per_token = std::max<float>(1.0f, float(size()) / (std::max<size_t>(1, n_tokens())));
// dynamically increase the token limit if it can fit in the memory limit
const size_t limit_tokens_cur = limit_size > 0 ? std::max<size_t>(limit_tokens, limit_size/size_per_token) : limit_tokens;
if (limit_tokens > 0) {
while (!states.empty() && n_tokens() > limit_tokens_cur) {
SRV_WRN(" - cache token limit (%zu, est: %zu) reached, removing oldest entry (size = %.3f MiB)\n",
limit_tokens, limit_tokens_cur, states.front().size() / (1024.0 * 1024.0));
states.pop_front();
}
}
SRV_TRC(" - cache state: %zu prompts, %.3f MiB (limits: %.3f MiB, %zu tokens, %zu est)\n",
states.size(), size() / (1024.0 * 1024.0), limit_size / (1024.0 * 1024.0), limit_tokens, limit_tokens_cur);
for (const auto & state : states) {
SRV_TRC(" - prompt %p: %7d tokens, checkpoints: %2zu, %9.3f MiB\n",
(const void *)&state, state.prompt.n_tokens(), state.prompt.checkpoints.size(), state.size() / (1024.0 * 1024.0));
}
}
//
// server_slot_sched
//
static std::vector<const server_slot_info *> get_free_slot_infos(const std::vector<server_slot_info> & slots, size_t n_slots_needed, int exclude_id_slot) {
std::vector<const server_slot_info *> free_slots;
for (const auto & slot : slots) {
if (!slot.is_processing && slot.id != exclude_id_slot) {
free_slots.push_back(&slot);
}
}
std::stable_sort(free_slots.begin(), free_slots.end(), [](const server_slot_info * a, const server_slot_info * b) {
if ((a->prompt->n_tokens() == 0) != (b->prompt->n_tokens() == 0)) {
return a->prompt->n_tokens() == 0;
}
return a->t_last_used < b->t_last_used;
});
free_slots.resize(std::min(free_slots.size(), n_slots_needed));
return free_slots;
}
std::vector<int> server_slot_sched::get_free_slots(const std::vector<server_slot_info> & slots, size_t n_slots_needed, int exclude_id_slot) const {
std::vector<int> res;
for (const auto * slot : get_free_slot_infos(slots, n_slots_needed, exclude_id_slot)) {
res.push_back(slot->id);
}
return res;
}
server_slot_plan server_slot_sched::select(const std::vector<server_slot_info> & slots, const server_task & task, const std::vector<common_adapter_lora_info> & lora) const {
const server_slot_info * ret = nullptr;
bool update_cache = false;
// if a specific slot is requested, use it (still goes through cache update logic below)
if (task.id_slot != -1 && !slots.empty()) {
// note: allow id_slot to be out of bounds (wrap around)
const int id_slot = task.id_slot % slots.size();
for (const auto & slot : slots) {
if (slot.id == id_slot) {
ret = &slot;
break;
}
}
if (ret) {
// a busy slot is returned untouched, the caller defers the task
if (ret->is_processing) {
return { ret->id, false };
}
SLT_INF(*ret, "selected slot by id (%d)\n", task.id_slot);
}
}
// find the slot that has at least n% prompt similarity
if (slot_prompt_similarity != 0.0f) {
float f_sim_best = 0;
for (const auto & slot : slots) {
if (task.id_slot != -1 && slot.id != task.id_slot) {
continue;
}
// skip the slot if it is not available
if (slot.is_processing) {
SLT_TRC(slot, " - skipping, is_processing = %d\n", slot.is_processing);
continue;
}
const auto & tokens = slot.prompt->tokens;
// skip the slot if it does not contains cached tokens
if (tokens.empty()) {
SLT_TRC(slot, "%s", " - skipping, slot is empty\n");
continue;
}
// the task would clear this slot, its state cannot be reused with another lora config
if (lora_should_clear_cache(slot.prompt->lora, lora)) {
SLT_TRC(slot, "%s", " - skipping, other lora config\n");
continue;
}
// fraction of the Longest Common Prefix length with respect to the input prompt length
const size_t lcp_len = tokens.get_common_prefix(task.tokens);
const float f_sim_cur = float(lcp_len) / task.tokens.size();
SLT_TRC(slot, " - checking sim = %.3f (%zu/%zu) > %.3f\n", f_sim_cur, lcp_len, task.tokens.size(), slot_prompt_similarity);
// select the current slot if the criteria match
if (f_sim_cur > f_sim_best && f_sim_cur > slot_prompt_similarity) {
f_sim_best = f_sim_cur;
ret = &slot;
}
}
if (ret != nullptr) {
const float f_keep = (f_sim_best*task.tokens.size()) / ret->prompt->tokens.size();
if (task.id_slot == -1) {
SLT_INF(*ret, "selected slot by LCP similarity, f_sim_best = %.3f (> %.3f thold), f_keep = %.3f\n",
f_sim_best, slot_prompt_similarity, f_keep);
}
// if we are about to lose a large portion of the existing context - save it in the prompt cache
if (f_keep < 0.5f) {
update_cache = true;
}
}
}
// find an empty slot, or the one that has been least recently used
if (ret == nullptr) {
const auto free_slots = get_free_slot_infos(slots, 1, -1);
if (!free_slots.empty()) {
ret = free_slots[0];
SLT_INF(*ret, "selected slot by LRU, n_tokens = %d, t_last = %" PRId64 "\n", ret->prompt->n_tokens(), ret->t_last_used);
update_cache = true;
}
}
if (ret == nullptr) {
return {};
}
// cache prompts only for the tasks that reuse them
update_cache = update_cache && !task.is_stateless();
return { ret->id, update_cache };
}