nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

common : fix state save in common_prompt_batch_decode (#23468)

Browse Source 
* common : fix state save in common_prompt_batch_decode

This commit addresses a bug in common_prompt_batch_decode that affects
the session state store/restore in completion.cpp and
save-load-state.cpp.

The motivation for this is that currently the code is saving n-1 tokens
in both the session_tokens and in the KV cache. Then when loading the
session tokens, and if the prompt matches, it would replay the last
saved token (n-1) into the next position, effectively replaying the
same token in the wrong position.

The fix is to store all n tokens in session_tokens, while the memory
state only reflects n-1 processed tokens as the saving happens before
the last token is decoded in common_prompt_batch_decode.

I ran both completion.cpp and save-load-state.cpp with a transformer, a
recurrent, and a hybrid model.

Resolves: https://github.com/ggml-org/llama.cpp/issues/23400

Co-authored-by: fairydreaming <166155368+fairydreaming@users.noreply.github.com>
This commit is contained in:
Daniel Bevenius
2026-06-02 15:44:15 +02:00
committed by GitHub
parent 60130d18f9
commit 0b7154066e
4 changed files with 31 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
common/common.cpp
View File

@@ -1982,36 +1982,37 @@ bool common_replay_last_token(struct llama_context * ctx, llama_token last_token
bool common_prompt_batch_decode( bool common_prompt_batch_decode(
struct llama_context * ctx, struct llama_context * ctx,
const std::vector<llama_token> & tokens, const std::vector<llama_token> & all_tokens,
int n_new,
int & n_past, int & n_past,
int n_batch, int n_batch,
std::string_view state_path, std::string_view state_path,
bool save_state) { bool save_state) {
const int n_eval = tokens.size(); if (n_new == 0) {
if (n_eval == 0) {
return true; return true;
} }
const int offset = all_tokens.size() - n_new;
if (save_state && n_eval > 1) { if (save_state && n_new > 1) {
const int n_tokens_before_last = n_eval - 1; const int n_tokens_before_last = n_new - 1;
GGML_ASSERT(n_eval <= n_batch); GGML_ASSERT(n_new <= n_batch);
// Decode all but the last token so we can save the memory state before decoding the last token. // Decode all but the last token so we can save the memory state before decoding the last token.
// This is done so we can restore the session state later and replay the last token. // This is done so we can restore the session state later and replay the last token.
// Memory implementations in recurrent/hybrid models don't support removing tokens from their // Memory implementations in recurrent/hybrid models don't support removing tokens from their
// memory, so we can't just remove the last token from the memory and replay the last token which // memory, so we can't just remove the last token from the memory and replay the last token which
// is the reason for this logic. // is the reason for this logic.
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_tokens_before_last))) { if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_tokens_before_last))) {
LOG_ERR("%s : failed to eval\n", __func__); LOG_ERR("%s : failed to eval\n", __func__);
return false; return false;
} }
n_past += n_tokens_before_last; n_past += n_tokens_before_last;
llama_state_save_file(ctx, state_path.data(), tokens.data(), n_tokens_before_last); llama_state_save_file(ctx, state_path.data(), all_tokens.data(), all_tokens.size());
LOG_INF("saved session before last token to %s, n_tokens = %d\n", state_path.data(), n_tokens_before_last); LOG_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size());
llama_token last_token = tokens.back(); llama_token last_token = all_tokens.back();
llama_batch batch = llama_batch_get_one(&last_token, 1); llama_batch batch = llama_batch_get_one(&last_token, 1);
int32_t pos = n_past; int32_t pos = n_past;
batch.pos = &pos; batch.pos = &pos;
@@ -2022,11 +2023,11 @@ bool common_prompt_batch_decode(
} }
n_past++; n_past++;
} else { } else {
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_eval))) { if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_new))) {
LOG_ERR("%s : failed to eval\n", __func__); LOG_ERR("%s : failed to eval\n", __func__);
return false; return false;
} }
n_past += n_eval; n_past += n_new;
} }
return true; return true;

3
common/common.h
View File

@@ -930,7 +930,8 @@ void common_batch_add(
// tokens from memory, so this approach works across all model architectures. // tokens from memory, so this approach works across all model architectures.
bool common_prompt_batch_decode( bool common_prompt_batch_decode(
struct llama_context * ctx, struct llama_context * ctx,
const std::vector<llama_token> & embd, const std::vector<llama_token> & all_tokens,
int n_new,
int & n_past, int & n_past,
int n_batch, int n_batch,
std::string_view state_path, std::string_view state_path,

8
tests/test-save-load-state.cpp
View File

@@ -63,7 +63,7 @@ static std::string test_baseline(struct llama_model * model, const struct common
auto tokens = common_tokenize(ctx.get(), params.prompt, true); auto tokens = common_tokenize(ctx.get(), params.prompt, true);
auto n_past = 0; auto n_past = 0;
if (!common_prompt_batch_decode(ctx.get(), tokens, n_past, params.n_batch, params.out_file, true)) { if (!common_prompt_batch_decode(ctx.get(), tokens, (int)tokens.size(), n_past, params.n_batch, params.out_file, true)) {
LOG_ERR("%s: failed to decode prompt\n", __func__); LOG_ERR("%s: failed to decode prompt\n", __func__);
return {}; return {};
} }
@@ -111,7 +111,7 @@ static bool test_state_load(struct llama_model * model, const struct common_para
LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out); LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out);
// Replay last token // Replay last token
int n_past = (int) n_token_count_out; int n_past = (int) n_token_count_out - 1;
if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) { if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) {
return false; return false;
} }
@@ -165,7 +165,7 @@ static bool test_seq_cp_host(struct llama_model * model, const struct common_par
LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out); LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out);
// Replay last token // Replay last token
int n_past = (int) n_token_count_out; int n_past = (int) n_token_count_out - 1;
if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) { if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) {
return false; return false;
} }
@@ -240,7 +240,7 @@ static bool test_seq_cp_device(struct llama_model * model, const struct common_p
LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out); LOG_TRC("%s: loaded state with %zu tokens\n", __func__, n_token_count_out);
// Replay last token // Replay last token
int n_past = (int) n_token_count_out; int n_past = (int) n_token_count_out - 1;
if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) { if (!common_replay_last_token(ctx.get(), tokens.back(), n_past)) {
return false; return false;
} }

25
tools/completion/completion.cpp
View File

@@ -373,16 +373,10 @@ int llama_completion(int argc, char ** argv) {
__func__, n_match, embd_inp.size()); __func__, n_match, embd_inp.size());
} }
if (session_tokens.size() == n_match) {
// [TAG_CONTEXT_STATE_LOGITS]
// in this case, we are going to reuse the logits from the session
// if we ever decide to remove the logits from the session, we need to handle this somehow
// ref: https://github.com/ggml-org/llama.cpp/pull/18862#issuecomment-3756330941
}
// remove any "future" tokens that we might have inherited from the previous session // remove any "future" tokens that we might have inherited from the previous session
if (session_tokens.size() > n_match) { if (session_tokens.size() > n_match) {
if (!llama_memory_seq_rm(mem, -1, n_match, -1)) { llama_pos pos = n_match > 0 ? (llama_pos)(n_match - 1) : 0;
if (!llama_memory_seq_rm(mem, -1, pos, -1)) {
LOG_WRN("%s: unable to reuse common prefix (for example, when the memory is recurrent)\n", __func__); LOG_WRN("%s: unable to reuse common prefix (for example, when the memory is recurrent)\n", __func__);
llama_memory_clear(mem, true); llama_memory_clear(mem, true);
session_tokens.clear(); session_tokens.clear();
@@ -398,7 +392,7 @@ int llama_completion(int argc, char ** argv) {
// Logits are not stored as part of the session state so we need to // Logits are not stored as part of the session state so we need to
// "replay" the last token to get logits for sampling. // "replay" the last token to get logits for sampling.
if (!session_tokens.empty() && n_match > 0 && n_match == session_tokens.size()) { if (!session_tokens.empty() && n_match > 0 && n_match == session_tokens.size()) {
if (!common_replay_last_token(ctx, session_tokens.back(), n_match)) { if (!common_replay_last_token(ctx, session_tokens.back(), n_match - 1)) {
return 1; return 1;
} }
@@ -695,12 +689,14 @@ int llama_completion(int argc, char ** argv) {
if (!embd.empty()) { if (!embd.empty()) {
const bool is_last_batch = (n_consumed >= (int) embd_inp.size()); const bool is_last_batch = (n_consumed >= (int) embd_inp.size());
const bool save_now = session_do_save && is_last_batch; const bool save_now = session_do_save && is_last_batch;
if (!common_prompt_batch_decode(ctx, embd, n_past, params.n_batch, path_session, save_now)) { session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
if (!common_prompt_batch_decode(ctx, session_tokens, embd.size(), n_past, params.n_batch, path_session, save_now)) {
return 1; return 1;
} }
session_tokens.insert(session_tokens.end(), embd.begin(), embd.end()); n_session_consumed += embd.size();
n_session_consumed = session_tokens.size(); if (save_now) {
session_do_save = false; session_do_save = false;
}
LOG_DBG("n_past = %d\n", n_past); LOG_DBG("n_past = %d\n", n_past);
@@ -991,7 +987,10 @@ int llama_completion(int argc, char ** argv) {
if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) { if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str()); LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size()); llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
LOG_INF("saved final session to %s, n_tokens = %ld\n", path_session.data(), session_tokens.size());
} }
LOG("\n\n"); LOG("\n\n");