Enable bitsandbytes unit tests on Radeon

Author: sstamenk

tests/models/quantization/test_bitsandbytes.py

@@ -10,13 +10,14 @@
 
 from tests.quantization.utils import is_quant_method_supported
 from vllm.platforms import current_platform
+from vllm.platforms.rocm import on_gfx9
 
 from ...utils import compare_two_settings, multi_gpu_test
 from ..utils import check_embeddings_close, check_logprobs_close
 
 pytestmark = pytest.mark.skipif(
-    current_platform.is_rocm(),
-    reason="bitsandbytes quantization not supported on ROCm (CUDA-only kernels)",
+    current_platform.is_rocm() and on_gfx9(),
+    reason="bitsandbytes quantization not supported on Instinct (warp size 64 limitation)",
 )
 
 models_4bit_to_test = [