Unit Tests for On Device Sampling

[quic-sanising]

This PR adds the following Unit Tests for On Device Sampling:

1. test_sampler_transform: Test if SamplerTransform adds nodes at the output of a QEffForCausalLM model to enable the sampling of next tokens at the device (instead of the host) and returns the next tokens and/or probability distributions.
2. test_greedy_sampling: Test greedy sampling with QPC compiled with and without On Device Sampling.
3. test_random_sampling: Test random sampling with QPC compiled with and without On Device Sampling.

[quic-rishinr]

@quic-sanising can you add a small feature description under /docs/source/quick_start.md supported feature section? also provide the example script link in the description

[quic-sanising]

@quic-sanising can you add a small feature description under /docs/source/quick_start.md supported feature section? also provide the example script link in the description

Done

[quic-amitraj]

Please fix lint error.

[quic-sanising]

Please fix lint error.

@quic-amitraj The lint failures were happening because the linter is installing ruff v0.12.2 whereas the .pre-commit-config.yaml file has an older version of v0.5.2.

To fix the errors, we need to either install ruff v0.5.2 in the linter or update the .pre-commit-config.yaml file to version v0.12.2.

[ochougul]

Everything else LGTM, just see if we can use single layer model in our tests

[quic-rishinr]

Can we include this check in line 489 itself?

[quic-sanising]

Could you clarify this?

[quic-rishinr]

Can we make this test more robust? The goal is to verify whether the random sampler is working as expected right? One idea is to check if the sampler produces the same output when given the same random seed or random values, and different outputs when the seed or values change. Secondly if we want to compare outputs with and without the sampler cam we use cosine similarity with a threshold instead of exact matching? im open for suggestions

[quic-sanising]

Thanks for the suggestion! The current test checks for exact output match, which allows us to reproduce it with a fixed seed value.

As for using cosine similarity or other metrics like perplexity or rogue score, they are useful for semantic comparison. But can be vague as there are no clear-defined or universally accepted thresholds.

So, for this test, exact matching might help keep things deterministic and easier to maintain in my opinion. Happy to explore alternatives if you want to...