llama : fix K/V and recurrent state cleanup after failed restores (#27530)

* llama : add discard for deferred state writes

* llama : add tensor zeroing helper for backends without tensor memset

* llama : clear K/V data after failed sequence restore

* llama : clear recurrent state data after failed sequence restore

* llama : simplify discard and restore cleanup

* llama : report error when abnormal cell count is found in state_read_meta

* llama : clear attention state on hybrid restore failure

* tests : cover failed state restore cleanup

* llama : clear MLA state on dsa restore failure

* tests : update test for rebased test suite

* llama : clarify comment in llama_memory_recurrent::state_read
This commit is contained in:
Chipmunk authored and GitHub committed 2026-09-26 10:23:03 +03:00
1 parent a1de614ba3
commit 08618ff8e7
11 files changed
+383 -17

No files matched your search

+9 -1
View File
@@ -168,7 +168,15 @@ void llama_kv_cache_dsa::state_write(llama_io_write_i & io, llama_seq_id seq_id,
void llama_kv_cache_dsa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
kv_mla->state_read(io, seq_id, flags);
kv_lid->state_read(io, seq_id, flags);
try {
kv_lid->state_read(io, seq_id, flags);
} catch (...) {
// the MLA part is already restored - undo it, so that a failed restore leaves nothing behind
kv_mla->state_clear(seq_id);
throw;
}
}
llama_kv_cache * llama_kv_cache_dsa::get_mla() const {