A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 19:15 UTC
[bnb] Skip moe + bnb test (#36896)
Signed-off-by: Marc Sun <marc@huggingface.co>
Marc Sun committed
Mar 12, 2026 at 19:03 UTC
c973ecdeada2bccda0eb0d1ec73c30119fc8aa85
1 file changed
+8
tests/models/quantization/test_bitsandbytes.py
+8
@@ -6,7 +6,9 @@ Run `pytest tests/quantization/test_bitsandbytes.py`.
6
"""
7
8
import pytest
9
+from packaging.version import Version
10
from transformers import BitsAndBytesConfig
11
+from transformers import __version__ as TRANSFORMERS_VERSION
12
13
from tests.quantization.utils import is_quant_method_supported
14
from vllm.platforms import current_platform
@@ -138,6 +140,12 @@ def test_load_pp_4bit_bnb_model(model_name, description) -> None:
140
compare_two_settings(model_name, common_args, pp_args)
141
142
143
+@pytest.mark.skipif(
144
+ Version(TRANSFORMERS_VERSION) >= Version("5.0.0"),
145
+ reason="Need to add support for quantizing MoE experts with bnb"
146
+ " in transformers v5. See"
147
+ " https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1849",
148
+)
149
@pytest.mark.skipif(
150
not is_quant_method_supported("bitsandbytes"),
151
reason="bitsandbytes is not supported on this GPU type.",