Require datasets 4.1.0

[dbutenhof]

Summary

Update Huggingface datasets dependency: unversioned dependency can break with an old previously installed datasets package.

Details

We previously had an unversioned dependency on datasets, and a recent report shows that GuideLLM is not compatible with versions of datasets prior to 3.1.0.

Since the "audio" extra already depends on datasets 4.1.0 we know that works and it seems a reasonable target.

Test Plan

Related Issues

---

• "I certify that all code in this PR is my own, except as noted below."

Use of AI

• Includes AI-assisted code completion
• Includes code generated by an AI application
• Includes AI-generated tests (NOTE: AI written tests should have a docstring that includes ## WRITTEN BY AI ##)

[dbutenhof]

Starting with #647, I'm seeing every PR unit-tests run failing, but a re-run has (so far) in each case succeeded. I'm not quite sure how this could have been affected by #647 and it's possibly a total coincidence ... but flaky tests make me nervous.

    async def test_requeue_with_positive_delay(self, worker_instance):
        """Test requeueing with positive delay sleeps then appends to turns_queue.
    
        ### WRITTEN BY AI ###
        """
        history = [("req1", "resp1")]
        conversation = [("req2", RequestInfo(request_id="req2"))]
        delay = 0.1
    
        start = time.time()
        await worker_instance._wait_then_requeue(history, conversation, delay)
        elapsed = time.time() - start
    
        # Should have slept for approximately the delay time
>       assert elapsed >= delay
E       assert 0.09976863861083984 >= 0.1

[dbutenhof]

Starting with #647, I've been seeing the first unit-tests run failing on each PR, but a re-run has (so far) in each case succeeded. I'm not quite sure how this could have been affected by #647 and it's possibly a total coincidence ... but flaky tests make me nervous.

And ... that's what I get for mentioning this. This PR's unit tests have failed 4 retries in a row... just my luck. 🪦

[sjmonson]

Starting with #647, I'm seeing every PR unit-tests run failing, but a re-run has (so far) in each case succeeded.Starting with #647, I'm seeing every PR unit-tests run failing, but a re-run has (so far) in each case succeeded.

I just filed an upstream issue MagicStack/uvloop#739 the reason we didn't see the CI issue before is that uvloop was only enabled in the benchmark entrypoint.

[jaredoconnell]

Starting with #647, I'm seeing every PR unit-tests run failing, but a re-run has (so far) in each case succeeded. I'm not quite sure how this could have been affected by #647 and it's possibly a total coincidence ... but flaky tests make me nervous.

    async def test_requeue_with_positive_delay(self, worker_instance):
        """Test requeueing with positive delay sleeps then appends to turns_queue.
    
        ### WRITTEN BY AI ###
        """
        history = [("req1", "resp1")]
        conversation = [("req2", RequestInfo(request_id="req2"))]
        delay = 0.1
    
        start = time.time()
        await worker_instance._wait_then_requeue(history, conversation, delay)
        elapsed = time.time() - start
    
        # Should have slept for approximately the delay time
>       assert elapsed >= delay
E       assert 0.09976863861083984 >= 0.1

Interesting. It's requeuing too fast. I would expect a minimum to not be flaky because you'd only expect it to exceed the expected value.

[dbutenhof]

I just filed an upstream issue MagicStack/uvloop#739 the reason we didn't see the CI issue before is that uvloop was only enabled in the benchmark entrypoint.

Ouch ... well, get used to occasional random CI failures until they (or we) do something about this. Your 2 PRs passed on the second try ... I had to retrigger this one 4 or 5 times before I got lucky.