text-generation-inference/integration-tests/models/test_mllama.py
2025-02-12 12:26:52 +01:00

96 lines
2.8 KiB
Python

import pytest
import asyncio
@pytest.fixture(scope="module")
def mllama_handle(launcher):
with launcher(
"meta-llama/Llama-3.2-11B-Vision-Instruct",
num_shard=2,
) as handle:
yield handle
@pytest.fixture(scope="module")
async def mllama(mllama_handle):
await mllama_handle.health(300)
return mllama_handle.client
@pytest.mark.asyncio
async def test_mllama_simpl(mllama, response_snapshot):
response = await mllama.chat(
max_tokens=10,
temperature=0.0,
messages=[
{
"role": "user",
"content": [
{
"type": "text",
"text": "Describe the image in 10 words.",
},
{
"type": "image_url",
"image_url": {
"url": "https://raw.githubusercontent.com/huggingface/text-generation-inference/main/integration-tests/images/chicken_on_money.png"
},
},
],
},
],
)
assert response.usage == {
"completion_tokens": 10,
"prompt_tokens": 45,
"total_tokens": 55,
}
assert (
response.choices[0].message.content
== "A chicken sits on a pile of money, looking"
)
assert response == response_snapshot
@pytest.mark.release
@pytest.mark.asyncio
async def test_mllama_load(mllama, generate_load, response_snapshot):
futures = [
mllama.chat(
max_tokens=10,
temperature=0.0,
messages=[
{
"role": "user",
"content": [
{
"type": "text",
"text": "Describe the image in 10 words.",
},
{
"type": "image_url",
"image_url": {
"url": "https://raw.githubusercontent.com/huggingface/text-generation-inference/main/integration-tests/images/chicken_on_money.png"
},
},
],
},
],
)
# TODO with v3, 4 breaks here. Nothing accounts of the image VRAM
# because mllama is the only one doing its thing.
for i in range(2)
]
responses = await asyncio.gather(*futures)
generated_texts = [response.choices[0].message.content for response in responses]
# XXX: TODO: Fix this test.
assert generated_texts[0] == "A chicken sits on a pile of money, looking"
assert len(generated_texts) == 2
assert generated_texts, all(
[text == generated_texts[0] for text in generated_texts]
)
assert responses == response_snapshot