biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI/Build] Add is_quant_method_supported to control quantization test configurations (#5253)

Browse Source
This commit is contained in:
Michael Goin
2024-06-12 12:58:02 -04:00
committed by GitHub
parent 8f89d72090
commit 3dd6853bc8
8 changed files with 30 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
tests/models/test_aqlm.py
View File

@@ -4,17 +4,8 @@ Run `pytest tests/models/test_aqlm.py`.
"""
import pytest
import torch
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
aqlm_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
aqlm_not_supported = (capability <
QUANTIZATION_METHODS["aqlm"].get_min_capability())
from tests.quantization.utils import is_quant_method_supported
# In this test we hardcode prompts and generations for the model so we don't
# need to require the AQLM package as a dependency
@@ -67,7 +58,7 @@ ground_truth_generations = [
]
@pytest.mark.skipif(aqlm_not_supported,
@pytest.mark.skipif(not is_quant_method_supported("aqlm"),
reason="AQLM is not supported on this GPU type.")
@pytest.mark.parametrize("model", ["ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf"])
@pytest.mark.parametrize("dtype", ["half"])