test_cpu_offload.py

# SPDX-License-Identifier: Apache-2.0

# Expanded quantized model tests for CPU offloading
# Base tests: tests/basic_correctness/test_cpu_offload.py

import pytest
import os

from tests.quantization.utils import is_quant_method_supported

from ..utils import compare_two_settings, models_path_prefix
from vllm.platforms import current_platform


@pytest.fixture(scope="function", autouse=True)
def use_v0_only(monkeypatch):
    # Fall back to V0 if cpu offloading is enabled.
    # Fixture is required to that baseline uses V0.
    monkeypatch.setenv('VLLM_USE_V1', '0')


@pytest.mark.skipif(not is_quant_method_supported("fp8") or current_platform.is_rocm(),
                    reason="fp8 is not supported on this GPU type.")
def test_cpu_offload_fp8():
    # Test quantization of an unquantized checkpoint
    compare_two_settings(os.path.join(models_path_prefix, "meta-llama/Llama-3.2-1B-Instruct"),
                         ["--quantization", "fp8"],
                         ["--quantization", "fp8", "--cpu-offload-gb", "1"],
                         max_wait_seconds=480)
    # Test loading a quantized checkpoint
    # compare_two_settings(os.path.join(models_path_prefix, "neuralmagic/Qwen2-1.5B-Instruct-FP8"), [],
    #                      ["--cpu-offload-gb", "1"],
    #                      max_wait_seconds=480)


@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin") or current_platform.is_rocm(),
                    reason="gptq_marlin is not supported on this GPU type.")
def test_cpu_offload_gptq():
    # Test GPTQ Marlin
    compare_two_settings(os.path.join(models_path_prefix, "Qwen/Qwen2-1.5B-Instruct-GPTQ-Int4"), [],
                         ["--cpu-offload-gb", "1"],
                         max_wait_seconds=480)
    # Test GPTQ
    compare_two_settings(os.path.join(models_path_prefix, "Qwen/Qwen2-1.5B-Instruct-GPTQ-Int4"),
                         ["--quantization", "gptq"],
                         ["--quantization", "gptq", "--cpu-offload-gb", "1"],
                         max_wait_seconds=480)


@pytest.mark.skipif(not is_quant_method_supported("awq_marlin") or current_platform.is_rocm(),
                    reason="awq_marlin is not supported on this GPU type.")
def test_cpu_offload_awq():
    # Test AWQ Marlin
    compare_two_settings(os.path.join(models_path_prefix, "Qwen/Qwen2-1.5B-Instruct-AWQ"), [],
                         ["--cpu-offload-gb", "1"],
                         max_wait_seconds=480)
    # Test AWQ
    compare_two_settings(os.path.join(models_path_prefix, "Qwen/Qwen2-1.5B-Instruct-AWQ"),
                         ["--quantization", "awq"],
                         ["--quantization", "awq", "--cpu-offload-gb", "1"],
                         max_wait_seconds=480)


@pytest.mark.skipif(not is_quant_method_supported("gptq_marlin") or current_platform.is_rocm(),
                    reason="gptq_marlin is not supported on this GPU type.")
def test_cpu_offload_compressed_tensors():
    # Test wNa16
    compare_two_settings(os.path.join(models_path_prefix, "nm-testing/tinyllama-oneshot-w4a16-channel-v2"), [],
                         ["--cpu-offload-gb", "1"],
                         max_wait_seconds=480)
    # Test w4a16_marlin24
    compare_two_settings(os.path.join(models_path_prefix, "nm-testing/llama7b-one-shot-2_4-w4a16-marlin24-t"),
                         [], ["--cpu-offload-gb", "1"],
                         max_wait_seconds=480)
    # Test w8a8
    compare_two_settings(
        os.path.join(models_path_prefix, "nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change"), [],
        ["--cpu-offload-gb", "1"],
        max_wait_seconds=480)