tests/test_api.py

# Copyright Lightning AI. Licensed under the Apache License 2.0, see LICENSE file.

import os
import re
import sys
from collections import OrderedDict
from pathlib import Path
from unittest.mock import MagicMock, patch

import pytest
import torch
from lightning.fabric.accelerators import CUDAAccelerator

from litgpt.api import LLM, benchmark_dict_to_markdown_table, calculate_number_of_devices
from litgpt.scripts.download import download_from_hub
from litgpt.utils import _RunIf

skip_in_ci_on_macos = pytest.mark.skipif(
    sys.platform == "darwin" and os.getenv("GITHUB_ACTIONS") == "true",
    reason="Skipped on macOS in CI environment because CI machine does not have enough memory to run this test.",
)


if sys.platform == "darwin" and os.getenv("GITHUB_ACTIONS") == "true":
    USE_MPS = False
elif torch.backends.mps.is_available():
    USE_MPS = True
else:
    USE_MPS = False


@pytest.fixture
def mock_llm():
    llm = MagicMock(spec=LLM)
    llm.model = MagicMock()
    llm.preprocessor = MagicMock()
    llm.prompt_style = MagicMock()
    llm.checkpoint_dir = MagicMock()
    llm.fabric = MagicMock()
    return llm


def test_load_model(mock_llm):
    assert isinstance(mock_llm, LLM)
    assert mock_llm.model is not None
    assert mock_llm.preprocessor is not None
    assert mock_llm.prompt_style is not None
    assert mock_llm.checkpoint_dir is not None
    assert mock_llm.fabric is not None


def test_generate(mock_llm):
    prompt = "What do Llamas eat?"
    mock_llm.generate.return_value = prompt + " Mock output"
    output = mock_llm.generate(prompt, max_new_tokens=10, temperature=0.8, top_k=5)
    assert isinstance(output, str)
    assert len(output) > len(prompt)


def test_stream_generate(mock_llm):
    prompt = "What do Llamas eat?"

    def iterator():
        outputs = (prompt + " Mock output").split()
        yield from outputs

    mock_llm.generate.return_value = iterator()
    output = mock_llm.generate(prompt, max_new_tokens=10, temperature=0.8, top_k=5, stream=True)
    result = "".join([out for out in output])
    assert len(result) > len(prompt)


def test_generate_token_ids(mock_llm):
    prompt = "What do Llamas eat?"
    mock_output_ids = MagicMock(spec=torch.Tensor)
    mock_output_ids.shape = [len(prompt) + 10]
    mock_llm.generate.return_value = mock_output_ids
    output_ids = mock_llm.generate(prompt, max_new_tokens=10, return_as_token_ids=True)
    assert isinstance(output_ids, torch.Tensor)
    assert output_ids.shape[0] > len(prompt)


def test_calculate_number_of_devices():
    assert calculate_number_of_devices(1) == 1
    assert calculate_number_of_devices([0, 1, 2]) == 3
    assert calculate_number_of_devices(None) == 0


def test_llm_load_random_init(tmp_path):
    download_from_hub(repo_id="EleutherAI/pythia-14m", tokenizer_only=True, checkpoint_dir=tmp_path)

    torch.manual_seed(123)
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(model="pythia-160m", init="random", tokenizer_dir=Path(tmp_path / "EleutherAI/pythia-14m"))

    input_text = "some text text"
    output_text = llm.generate(input_text, max_new_tokens=15)
    ln = len(llm.preprocessor.tokenizer.encode(output_text)) - len(llm.preprocessor.tokenizer.encode(input_text))
    assert ln <= 15

    # The following below tests that generate works with different prompt lengths
    # after the kv cache was set

    input_text = "some text"
    output_text = llm.generate(input_text, max_new_tokens=15)
    ln = len(llm.preprocessor.tokenizer.encode(output_text)) - len(llm.preprocessor.tokenizer.encode(input_text))
    assert ln <= 15

    input_text = "some text text text"
    output_text = llm.generate(input_text, max_new_tokens=15)
    ln = len(llm.preprocessor.tokenizer.encode(output_text)) - len(llm.preprocessor.tokenizer.encode(input_text))
    assert ln <= 15


def test_llm_load_hub_init(tmp_path):
    torch.manual_seed(123)
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(model="EleutherAI/pythia-14m", init="pretrained")

    text_1 = llm.generate("text", max_new_tokens=10, top_k=1)
    assert len(text_1) > 0

    text_2 = llm.generate("text", max_new_tokens=10, top_k=1, stream=True)
    text_2 = "".join(list(text_2))
    assert text_1 == text_2, (text_1, text_2)


def test_model_not_initialized(tmp_path):
    llm = LLM.load(model="EleutherAI/pythia-14m", init="pretrained", distribute=None)
    s = "The model is not initialized yet; use the .distribute() or .trainer_setup() method to initialize the model."
    with pytest.raises(AttributeError, match=re.escape(s)):
        llm.generate("text")

    llm = LLM.load(model="EleutherAI/pythia-14m", tokenizer_dir="EleutherAI/pythia-14m", init="random", distribute=None)
    s = "The model is not initialized yet; use the .distribute() or .trainer_setup() method to initialize the model."
    with pytest.raises(AttributeError, match=re.escape(s)):
        llm.generate("text")


@_RunIf(min_cuda_gpus=2)
def test_more_than_1_device_for_sequential_gpu(tmp_path):
    device_count = CUDAAccelerator.auto_device_count()

    if device_count <= 2:
        model_name = "EleutherAI/pythia-14m"
    else:
        model_name = "EleutherAI/pythia-160m"
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model=model_name,
        )

    with pytest.raises(
        NotImplementedError,
        match="Support for multiple devices is currently only implemented for generate_strategy='sequential'|'tensor_parallel'.",
    ):
        llm.distribute(devices=2)

    llm.distribute(devices=2, generate_strategy="sequential")
    assert isinstance(llm.generate("What do llamas eat?"), str)
    assert str(llm.model.transformer.h[0].mlp.fc.weight.device) == "cuda:0"
    last_layer_idx = len(llm.model.transformer.h) - 1
    assert str(llm.model.transformer.h[last_layer_idx].mlp.fc.weight.device) == "cuda:1"

    # Also check with default (devices="auto") setting
    llm.distribute(generate_strategy="sequential")
    assert isinstance(llm.generate("What do llamas eat?"), str)
    assert str(llm.model.transformer.h[0].mlp.fc.weight.device) == "cuda:0"
    assert str(llm.model.transformer.h[last_layer_idx].mlp.fc.weight.device) == f"cuda:{device_count - 1}"


@_RunIf(min_cuda_gpus=2)
@pytest.mark.skipif(bool(os.getenv("SKIP_WITH_CI")), reason="Skip this test in CI due to ...")
def test_more_than_1_device_for_tensor_parallel_gpu(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(model="EleutherAI/pythia-14m")

    # this crashes the CI, maybe because of process forking; works fine locally though
    llm.distribute(devices=2, generate_strategy="tensor_parallel")
    assert isinstance(llm.generate("What do llamas eat?"), str)


@_RunIf(min_cuda_gpus=1)
@pytest.mark.parametrize("strategy", ("sequential", "tensor_parallel"))
@pytest.mark.xfail(
    NotADirectoryError, reason="This test is expected to fail due to a NotADirectoryError.", strict=False
)
def test_sequential_tp_incompatibility_with_random_weights(strategy, tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(model="EleutherAI/pythia-14m", tokenizer_dir="EleutherAI/pythia-14m", init="random")
    with pytest.raises(
        NotImplementedError,
        match=re.escape(
            "The LLM was initialized with init='random' but .distribute() currently only supports pretrained weights."
        ),
    ):
        llm.distribute(devices=1, generate_strategy=strategy)


@pytest.mark.parametrize("strategy", ("sequential", "tensor_parallel"))
def test_sequential_tp_cpu(strategy, tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
            distribute=None,
        )
    with pytest.raises(
        NotImplementedError, match=f"generate_strategy='{strategy}' is only supported for accelerator='cuda'|'gpu'."
    ):
        llm.distribute(devices=1, accelerator="cpu", generate_strategy=strategy)


def test_initialization_for_trainer(tmp_path):
    llm = LLM.load(model="EleutherAI/pythia-14m", distribute=None)
    s = "The model is not initialized yet; use the .distribute() or .trainer_setup() method to initialize the model."
    with pytest.raises(AttributeError, match=re.escape(s)):
        llm.generate("hello world")

    llm.trainer_setup()
    llm.model.to(llm.preprocessor.device)
    assert isinstance(llm.generate("hello world"), str)


@_RunIf(min_cuda_gpus=1)
def test_quantization_is_applied(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )
    llm.distribute(devices=1, quantize="bnb.nf4", precision="bf16-true")
    strtype = str(type(llm.model.lm_head))
    assert "NF4Linear" in strtype, strtype


@_RunIf(min_cuda_gpus=1)
def test_fixed_kv_cache(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )
    llm.distribute(devices=1, fixed_kv_cache_size=100)

    # Request too many tokens
    with pytest.raises(NotImplementedError, match="max_seq_length 512 needs to be >= 9223372036854775809"):
        _ = llm.generate("hello world", max_new_tokens=2**63)


def test_invalid_accelerator(tmp_path):
    llm = LLM.load(model="EleutherAI/pythia-14m", distribute=None)
    with pytest.raises(ValueError, match="Invalid accelerator"):
        llm.distribute(accelerator="invalid")


def test_returned_benchmark_dir(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )

    text, bench_d = llm.benchmark(prompt="hello world")
    assert isinstance(bench_d["Inference speed in tokens/sec"], list)
    assert len(bench_d["Inference speed in tokens/sec"]) == 1
    assert isinstance(bench_d["Inference speed in tokens/sec"][0], float)

    text, bench_d = llm.benchmark(prompt="hello world", stream=True)
    assert isinstance(bench_d["Inference speed in tokens/sec"], list)
    assert len(bench_d["Inference speed in tokens/sec"]) == 1
    assert isinstance(bench_d["Inference speed in tokens/sec"][0], float)

    text, bench_d = llm.benchmark(num_iterations=10, prompt="hello world", stream=True)
    assert isinstance(bench_d["Inference speed in tokens/sec"], list)
    assert len(bench_d["Inference speed in tokens/sec"]) == 10
    assert isinstance(bench_d["Inference speed in tokens/sec"][0], float)


def test_benchmark_dict_to_markdown_table_single_values():
    bench_d = {
        "Inference speed in tokens/sec": [17.617540650112936],
        "Seconds to first token": [0.6533610639999097],
        "Seconds total": [1.4758019020000575],
        "Tokens generated": [26],
        "Total GPU memory allocated in GB": [5.923729408],
    }

    expected_output = (
        "| Metric                              | Mean                        | Std Dev                     |\n"
        "|-------------------------------------|-----------------------------|-----------------------------|\n"
        "| Inference speed in tokens/sec       | 17.62                       | nan                         |\n"
        "| Seconds to first token              | 0.65                        | nan                         |\n"
        "| Seconds total                       | 1.48                        | nan                         |\n"
        "| Tokens generated                    | 26.00                       | nan                         |\n"
        "| Total GPU memory allocated in GB    | 5.92                        | nan                         |\n"
    )

    assert benchmark_dict_to_markdown_table(bench_d) == expected_output


def test_benchmark_dict_to_markdown_table_multiple_values():
    bench_d_list = {
        "Inference speed in tokens/sec": [
            17.034547562152305,
            32.8974175404589,
            33.04784205046782,
            32.445697744648584,
            33.204480197756396,
            32.64187570945661,
            33.21232058140845,
            32.69377798373551,
            32.92351459309756,
            32.48909032591177,
        ],
        "Seconds to first token": [
            0.7403525039999295,
            0.022901020000063,
            0.02335712100011733,
            0.022969672000272112,
            0.022788318000039,
            0.02365505999978268,
            0.02320190000000366,
            0.022791139999753796,
            0.022871761999795126,
            0.023060415999680117,
        ],
        "Seconds total": [
            1.5263099829999192,
            0.7903355929997815,
            0.7867382069998712,
            0.8013389080001616,
            0.7830268640000213,
            0.7965228539997042,
            0.7828420160003589,
            0.7952583520000189,
            0.7897091279996857,
            0.8002686360000553,
        ],
        "Tokens generated": [26, 26, 26, 26, 26, 26, 26, 26, 26, 26],
        "Total GPU memory allocated in GB": [
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
            5.923729408,
        ],
    }

    expected_output = (
        "| Metric                              | Mean                        | Std Dev                     |\n"
        "|-------------------------------------|-----------------------------|-----------------------------|\n"
        "| Inference speed in tokens/sec       | 31.26                       | 5.01                        |\n"
        "| Seconds to first token              | 0.09                        | 0.23                        |\n"
        "| Seconds total                       | 0.87                        | 0.23                        |\n"
        "| Tokens generated                    | 26.00                       | 0.00                        |\n"
        "| Total GPU memory allocated in GB    | 5.92                        | 0.00                        |\n"
    )

    assert benchmark_dict_to_markdown_table(bench_d_list) == expected_output


def test_state_dict(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )
    assert isinstance(llm.state_dict(), OrderedDict)
    assert llm.state_dict()["lm_head.weight"].shape == torch.Size([50304, 128])


def test_save_method(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )

    target_dir = "saved_model"
    llm.save(target_dir)

    expected_files = [
        "config.json",
        "generation_config.json",
        "lit_model.pth",
        "model_config.yaml",
        "prompt_style.yaml",
        "tokenizer_config.json",
        "tokenizer.json",
    ]

    files_in_directory = os.listdir(target_dir)
    for file_name in expected_files:
        assert file_name in files_in_directory, f"{file_name} is missing from {target_dir}"


def test_forward_method(tmp_path):
    with patch("torch.backends.mps.is_available", return_value=USE_MPS):
        llm = LLM.load(
            model="EleutherAI/pythia-14m",
        )
    inputs = torch.ones(6, 128, dtype=torch.int64).to(next(llm.model.parameters()).device)

    assert llm(inputs).shape == torch.Size([6, 128, 50304])
    logits, loss = llm(inputs, target_ids=inputs)
    assert logits.shape == torch.Size([6, 128, 50304])
    assert isinstance(loss.item(), float)


@skip_in_ci_on_macos  # The macOS CI machine segfaults here (it works fine locally though)
def test_precision_selection(tmp_path):
    llm = LLM.load(model="EleutherAI/pythia-14m", init="pretrained")

    llm.distribute(precision="16-true")
    assert llm.model._forward_module.lm_head.weight.dtype == torch.float16, (
        f"Expected float16, but got {llm.model._forward_module.lm_head.weight.dtype}"
    )