* **FastAPI**:
    * The same way that Starlette uses Uvicorn and cannot be faster than it, **FastAPI** uses Starlette, so it cannot be faster than it.

   - Reject untested PRs unless testing is infeasible, then assist with test setup.

## Tips for Success

- **Small PRs**: Easier to review, faster to merge. Split large changes logically.
- **Clear Commits**: Use `git rebase -i` to refine history before submitting.
- **Engage Early**: Discuss complex changes in issues or Slack (https://slack.min.io) before coding.

# For wheel verification, and uploading
auditwheel ~= 6.1.0
twine ~= 6.1.0
id
urllib3
requests

# For XLA
pyyaml

# For JAX
build ~= 1.2.2
# uv is faster than pip for installing Python packages.
uv ~= 0.5.30

# For running wheel verification script

 *
 * <p>This service allows implementations to:</p>
 * <ul>
 *   <li>Pool identical objects to reduce memory footprint</li>
 *   <li>Intern objects for faster equality comparisons</li>
 *   <li>Apply custom optimization strategies</li>
 *   <li>Transform or modify objects during building</li>
 * </ul>
 *

## Other middlewares { #other-middlewares }

There are many other ASGI middlewares.

For example:

#   request - Cache only for current request + its child requests
#   none    - Disable caching entirely
#
# REFERENCE OPTIONS:
#   hard - Strong reference (stays in memory, faster access)
#   soft - Weak reference (can be garbage collected under memory pressure)
#
# CONFIGURATION BREAKDOWN:
#   SourceCacheKey                   - File and RAW model requests (session/hard for consistent file access)

In most setups this is sufficient to heal the content after drive replacements. Setting `max_sleep` to a *lower* value and setting `max_io` to a *higher* value would make heal go faster.

1. This is the generator function. It's a "generator function" because it contains `yield` statements inside.
2. By using a `with` block, we make sure that the file-like object is closed after the generator function is done. So, after it finishes sending the response.

        long time2 = System.nanoTime() - start2;

        assertEquals("testValue", bean1.getName());
        assertEquals("testValue", bean2.getName());

        // Second call should be faster (though this is not guaranteed in all environments)
        // We mainly verify that both calls work correctly
        assertTrue(time2 >= 0, "Expected " + time2 + " to be >= " + 0); // Just verify it completed
    }

EOF
    find . | tr '[A-Z]' '[a-z]' | sort | uniq -d | tee $BATS_FILE_TMPDIR/repeats
    [[ ! -s $BATS_FILE_TMPDIR/repeats ]]
}

# It's unclear why, but running this on //tensorflow/... is faster than running
# only on affected targets, usually. There are targets in //tensorflow/lite that
# don't pass --nobuild, so they're on their own.
#
# Although buildifier checks for formatting as well, "bazel build nobuild"