hf_text-generation-inference/integration-tests/models/test_flash_neox_sharded.py

import pytest


@pytest.fixture(scope="module")
def flash_neox_sharded_handle(launcher):
    with launcher("OpenAssistant/oasst-sft-1-pythia-12b", num_shard=2) as handle:
        yield handle


@pytest.fixture(scope="module")
async def flash_neox_sharded(flash_neox_sharded_handle):
    await flash_neox_sharded_handle.health()
    return flash_neox_sharded_handle.client


@pytest.mark.release
@pytest.mark.asyncio
async def test_flash_neox(flash_neox_sharded, response_snapshot):
    response = await flash_neox_sharded.generate(
        "<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>",
        max_new_tokens=10,
        decoder_input_details=True,
    )

    assert response.details.generated_tokens == 10
    assert response == response_snapshot


@pytest.mark.release
@pytest.mark.asyncio
async def test_flash_neox_load(flash_neox_sharded, generate_load, response_snapshot):
    responses = await generate_load(
        flash_neox_sharded,
        "<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>",
        max_new_tokens=10,
        n=4,
    )

    assert len(responses) == 4
    assert all([r.generated_text == responses[0].generated_text for r in responses])

    assert responses == response_snapshot