In our experience and testing, these types force us to represent data in inefficient ways because they are general purpose.
Ztunnel is not general purpose; it has an extremely tight goal.
We can exploit this to make a more efficient protocol, which is critical to achieve our resource footprint goals.

  pop_backward_tape.release()();

  std::vector<int64_t> targets;
  std::vector<Gradient*> used_in_grads;
  // We may end up with slightly fewer elements than we reserve, but grad.size()
  // should be a reasonably tight upper bound.
  targets.reserve(grad.size());
  used_in_grads.reserve(grad.size());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;

### Suggest solutions

* After being able to understand the question, you can give them a possible **answer**.

* In many cases, it's better to understand their **underlying problem or use case**, because there might be a better way to solve it than what they are trying to do.

And then, right after the `yield`, we unload the model. This code will be executed **after** the application **finishes handling requests**, right before the *shutdown*. This could, for example, release resources like memory or a GPU.

!!! tip

---

If you just don't know, use normal `def`.

---

**Note**: You can mix `def` and `async def` in your *path operation functions* as much as you need and define each one using the best option for you. FastAPI will do the right thing with them.

Anyway, in any of the cases above, FastAPI will still work asynchronously and be extremely fast.

But by following the steps above, it will be able to do some performance optimizations.

    @Override
    public int compare(@CheckForNull Object left, @CheckForNull Object right) {
      if (left == right) {
        return 0;
      } else if (left == null) {
        return -1;
      } else if (right == null) {
        return 1;
      }
      int leftCode = identityHashCode(left);
      int rightCode = identityHashCode(right);
      if (leftCode != rightCode) {
        return leftCode < rightCode ? -1 : 1;

    @ParametricNullness private final V right;

    static <V extends @Nullable Object> ValueDifference<V> create(
        @ParametricNullness V left, @ParametricNullness V right) {
      return new ValueDifferenceImpl<>(left, right);
    }

    private ValueDifferenceImpl(@ParametricNullness V left, @ParametricNullness V right) {
      this.left = left;
      this.right = right;
    }

    @Override

And it shows their true commitment to FastAPI and its **community** (you), as they not only want to provide you a **good service** but also want to make sure you have a **good and healthy framework**, FastAPI. 🙇

For example, you might want to try <a href="https://speakeasyapi.dev/?utm_source=fastapi+repo&utm_medium=github+sponsorship" class="external-link" target="_blank">Speakeasy</a>.

The client will get a **500 Internal Server Error** for that request, but the application will continue working for the next requests instead of just crashing completely.

### Bigger Errors - Crashes

Nevertheless, there might be cases where we write some code that **crashes the entire application** making Uvicorn and Python crash. 💥

				// Subsequent reads will reuse this buffer.
				p.buf[bufIdx] = make([]byte, p.shardSize)
			}
			// For the last shard, the shardsize might be less than previous shard sizes.
			// Hence the following statement ensures that the buffer size is reset to the right size.
			p.buf[bufIdx] = p.buf[bufIdx][:p.shardSize]
			n, err := rr.ReadAt(p.buf[bufIdx], p.offset)
			if err != nil {
				switch {