Skip to content

Improve turbomind's prefix cache #3835

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

Open
wants to merge 14 commits into
base: main
Choose a base branch
from

Conversation

lvhan028
Copy link
Collaborator

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily receiving feedbacks. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

Please briefly describe what modification is made in this PR.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

  1. Pre-commit or other linting tools are used to fix the potential lint issues.
  2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
  3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
  4. The documentation has been modified accordingly, like docstring or example tutorials.

@lvhan028 lvhan028 requested review from irexyc and lzhangzz and removed request for irexyc August 13, 2025 09:12
@lvhan028 lvhan028 mentioned this pull request Aug 13, 2025
8 tasks
@lvhan028 lvhan028 requested a review from irexyc August 13, 2025 09:13
@@ -80,6 +83,7 @@ def main(model_path, backend, **kwargs):
try:
for resp in resps:
print(resp.text, end='', flush=True)
pass
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

seems unnecessary

Comment on lines +1283 to +1307
const int output_len = state_->h_context_length[index];
auto& seq = *state_->sequences[index];

// Update token IDs
seq.tokens.resize(output_len);
// Update token IDs
seq.tokens.resize(output_len);

// output_ids is updated & synced in `Finish`
const auto output_ids = state_->requests[index]->output_ids.data();
std::copy_n(output_ids, output_len, seq.tokens.data());
// output_ids is updated & synced in `Finish`
const auto output_ids = state_->requests[index]->output_ids.data();
std::copy_n(output_ids, output_len, seq.tokens.data());
// Cache the generated tokens of the sequence
if (!force_stop) {
sequence_manager_->CacheGeneration(seq);
}

// Save random state in host memory
seq.random_state.resize(sizeof(curandState_t));
// This async copy must be synchronized by the caller
core::Copy((curandState_t*)state_->curand_state.data() + index, 1, (curandState_t*)seq.random_state.data());
// Save random state in host memory
seq.random_state.resize(sizeof(curandState_t));
// This async copy must be synchronized by the caller
core::Copy((curandState_t*)state_->curand_state.data() + index, 1, (curandState_t*)seq.random_state.data());

// Set unlock flag for corresponding blocks, will be unlocked in the next `Materialize()`
sequence_manager_->UpdateAndSetUnlock(seq);
// Set unlock flag for corresponding blocks, will be unlocked in the next `Materialize()`
sequence_manager_->UpdateAndSetUnlock(seq);

if (state_->requests[index]->session.end_flag) {
// Sequence is ending this round
FT_CHECK(sequence_manager_->Erase(state_->requests[index]->id));
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we should refer to previous code here. With state_->requests[index]->session.end_flag=True, there is no need to output token id and save random_state.

Comment on lines +120 to +124
if (param_.enable_prefix_caching && !r->session.start_flag) {
// Prefix caching is incompatible with interactive mode
TM_LOG_ERROR("Skip inconsistent %s request for ID %lu", type, r->id);
r->ec = Request::kInconsistency;
}
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think we could check session.step == 0 here, so that we don't need to modify line 233-236, 262 and maybe 1018.

const int history_len = state_->sequences[i]->tokens.size();
const int cache_len = state_->sequences[i]->cache_len;
const int history_len =
!param_.enable_prefix_caching ? state_->sequences[i]->tokens.size() : state_->requests[i]->session.step;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I am confused about the setting state_->requests[i]->session.step

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants