fix cache verification (#821)

src/turbomind/models/llama/BlockManager.cc

@@ -207,7 +207,7 @@ int BlockManager::Lock(const BlockIds& ids)
 
     for (const auto& i : ids) {
         auto& b = blocks_[i];
-        FT_CHECK(is_cached(b));
+        FT_CHECK_WITH_INFO(is_cached(b), to_string(b));
         if (++b.use_count == 1) {
             lock.push_back(i);
             FT_CHECK(is_active(b));

src/turbomind/models/llama/SequenceManager.cc

@@ -93,17 +93,16 @@ void SequenceManager::VerifyAndLockCached(const Sequences& sequences)
             continue;
         }
         FT_CHECK(seq.blocks.size() == seq.block_unique_ids.size());
-        if (need_verify_) {
+        // Verify cache blocks that may be invalidated
         const int count = block_manager_->Verify(seq.blocks, seq.block_unique_ids);
         seq.blocks.resize(count);
         seq.block_unique_ids.resize(count);
-        }
+
         blocks.insert(blocks.end(), seq.blocks.begin(), seq.blocks.end());
         seq.cache_len = std::min<int>(seq.cache_len, seq.blocks.size() * block_seq_len_);
         seq.status    = Sequence::kLocked;
     }
     block_manager_->Lock(blocks);
-    need_verify_ = false;
 }
 
 void SequenceManager::CommitUnlockAndFree()
@@ -435,7 +434,6 @@ auto SequenceManager::Materialize(Sequences                    sequences,
     // evict cached blocks -> free
     if (schedule.evict) {
         block_manager_->Evict(schedule.evict);
-        need_verify_ = true;
     }
 
     // allocate & assign blocks

src/turbomind/models/llama/SequenceManager.h

@@ -127,8 +127,6 @@ private:
     int    rank_;
     size_t val_offset_{};
 
-    bool need_verify_{};
-
     // Use `std::map` to avoid reference invalidation
     std::map<uint64_t, Sequence> sequences_;