Skip to content

Server.cpp loadPrompt(): Fix segfault when prompt length exceeds ctx size #3639

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
25 changes: 14 additions & 11 deletions examples/server/server.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -436,31 +436,34 @@ struct llama_server_context
}
params.n_keep = std::min(n_ctx - 4, params.n_keep);

// if input prompt is too big, truncate like normal
// if input prompt is too big, we will truncate in the same way when the embd becomes too big when generating tokens
if (num_prompt_tokens >= (size_t)n_ctx)
{
const int n_left = (n_ctx - params.n_keep) / 2;
const int n_left = n_ctx - params.n_keep;

// Keep n_keep tokens of start of prompt (at most n_ctx - 4)
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
std::copy(prompt_tokens.end() - n_ctx, prompt_tokens.end(), last_n_tokens.begin());

// Use half the left-over space in the context for the prompt
new_tokens.insert(new_tokens.end(), prompt_tokens.end() - n_left / 2, prompt_tokens.end());

LOG_VERBOSE("input truncated", {
{"n_ctx", n_ctx},
{"n_keep", params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
{"num_prompt_tokens", new_tokens.size()}
});

truncated = true;
prompt_tokens = new_tokens;
num_prompt_tokens = prompt_tokens.size();
}
else
{
const size_t ps = num_prompt_tokens;
std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
}

// Initialize last_n_tokens
const size_t ps = num_prompt_tokens;
std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);

// compare the evaluated prompt with the new prompt
n_past = common_part(embd, prompt_tokens);
Expand Down