Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

STRATA_QFUSE=1: a one-token window did not self-commit, but commit() skipped the commit graph as if it had (#3948)

Niko1221789d83903ba0
diff --git a/src/core/verify.cpp b/src/core/verify.cpp
index 193b8f5ba..391b95312 100644
--- a/src/core/verify.cpp
+++ b/src/core/verify.cpp
@@ -3427,8 +3427,9 @@ bool Verifier::commit(int n_keep, std::string& err) {
h_commit_[0] = n_keep;
h_commit_[1] = n_keep - 1;
for (int t = 0; t < max_t_; ++t) h_commit_[2 + t] = t < n_keep ? (int32_t) (last_pos0_ + t) : -1;
- if (last_t_ == 1 && !always_publish_ && one_token_self_commit()) {
- // a one-token window has advanced the state itself (record_window): no commit graph
+ if (last_t_ == 1 && !always_publish_ && one_token_self_commit() && !g_qfuse()) {
+ // a one-token window has advanced the state itself (record_window): no commit graph. (Not with STRATA_QFUSE: its windows
+ // do not self-commit, see `self_commit` there - skipping the commit here left the GDN state and the indexer tail one token behind.)
} else {
std::atomic_thread_fence(std::memory_order_seq_cst);
// STRATA_COMMIT_TIMING=1 (measurement only): the commit graph's GPU time, from an event pair around it (read at the