xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 06:34:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable bitsandbytes quantization on AMD GPUs that use warp size 32 (#27307)

Browse Source 
Signed-off-by: sstamenk <strahinja.stamenkovic@amd.com>
This commit is contained in:
Strahinja Stamenkovic 2025-11-19 04:12:31 +01:00 committed by GitHub
parent 20852c8f4c
commit 814843e021
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 10 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
tests/models/quantization/test_bitsandbytes.py
View File

@ -14,10 +14,13 @@ from vllm.platforms import current_platform
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)",
)
if current_platform.is_rocm():
from vllm.platforms.rocm import on_gfx9
pytestmark = pytest.mark.skipif(
on_gfx9(),
reason="bitsandbytes not supported on gfx9 (warp size 64 limitation)",
)
models_4bit_to_test = [
("facebook/opt-125m", "quantize opt model inflight"),

3
vllm/platforms/rocm.py
View File

@ -185,6 +185,9 @@ class RocmPlatform(Platform):
"petit_nvfp4",
"torchao",
]
# bitsandbytes not supported on gfx9 (warp size 64 limitation)
if not on_gfx9():
supported_quantization += ["bitsandbytes"]
@classmethod
def get_vit_attn_backend(