vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama: reverting kv_cache in case of failed compute

Browse source
This commit is contained in:
Michael Podvitskiy 2024-09-24 21:12:47 +02:00
parent 5e354e3ca2
commit 4701893233
1 changed files with 49 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

43
src/llama.cpp
View file

@ -2811,6 +2811,42 @@ struct llama_kv_cache {
} }
}; };
class llama_kv_cache_state {
struct llama_kv_cache_state_short {
uint32_t head = 0;
uint32_t size = 0;
uint32_t used = 0;
uint32_t n = 0;
std::vector<llama_kv_cell> cells;
} old_state;
bool saved = false;
public:
void save_state(const llama_kv_cache& cache) {
old_state.head = cache.head;
old_state.size = cache.size;
old_state.used = cache.used;
old_state.n = cache.n;
old_state.cells = cache.cells;
saved = true;
}
void restore(llama_kv_cache& cache) {
if (saved) {
cache.head = old_state.head;
cache.size = old_state.size;
cache.used = old_state.used;
cache.n = old_state.n;
cache.cells = std::move(old_state.cells);
saved = false;
}
}
};
struct llama_control_vector { struct llama_control_vector {
std::vector<struct ggml_tensor *> tensors; // per layer std::vector<struct ggml_tensor *> tensors; // per layer
std::vector<ggml_context_ptr> ctxs; std::vector<ggml_context_ptr> ctxs;
@ -17256,6 +17292,7 @@ static int llama_decode_internal(
lctx.n_queued_tokens += n_tokens_all; lctx.n_queued_tokens += n_tokens_all;
auto & kv_self = lctx.kv_self; auto & kv_self = lctx.kv_self;
llama_kv_cache_state kv_cache_state_holder;
const int64_t n_embd = hparams.n_embd; const int64_t n_embd = hparams.n_embd;
const int64_t n_vocab = hparams.n_vocab; const int64_t n_vocab = hparams.n_vocab;
@ -17333,6 +17370,7 @@ static int llama_decode_internal(
// non-causal masks do not use the KV cache // non-causal masks do not use the KV cache
if (hparams.causal_attn) { if (hparams.causal_attn) {
llama_kv_cache_update(&lctx); llama_kv_cache_update(&lctx);
kv_cache_state_holder.save_state(kv_self);
// if we have enough unused cells before the current head -> // if we have enough unused cells before the current head ->
// better to start searching from the beginning of the cache, hoping to fill it // better to start searching from the beginning of the cache, hoping to fill it
@ -17390,9 +17428,9 @@ static int llama_decode_internal(
llama_set_inputs(lctx, ubatch); llama_set_inputs(lctx, ubatch);
const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool); const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
if (compute_status != GGML_STATUS_SUCCESS) {
kv_cache_state_holder.restore(kv_self);
switch (compute_status) { switch (compute_status) {
case GGML_STATUS_SUCCESS:
break;
case GGML_STATUS_ABORTED: case GGML_STATUS_ABORTED:
return 2; return 2;
case GGML_STATUS_ALLOC_FAILED: case GGML_STATUS_ALLOC_FAILED:
@ -17401,6 +17439,7 @@ static int llama_decode_internal(
default: default:
return -3; return -3;
} }
}
// update the kv ring buffer // update the kv ring buffer
{ {