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 14, 04:15 UTC
[CI/Build][Bugfix]Fix Quantized Models Test on AMD (#27712)
Signed-off-by: zhewenli <zhewenli@meta.com>
Zhewen Li committed
Oct 28, 2025 at 23:27 UTC
83fd49b1fc6d3af6a4c77ef8bfa6451c39402239
3 files changed
+8
-2
.buildkite/test-amd.yaml
+1
-1
@@ -908,7 +908,7 @@ steps:
908
909
- label: Quantized Models Test # 45 min
910
timeout_in_minutes: 60
911
- mirror_hardwares: [amdexperimental]
911
+ mirror_hardwares: [amdexperimental, amdproduction]
912
agent_pool: mi325_1
913
# grade: Blocking
914
source_file_dependencies:
tests/models/quantization/test_bitsandbytes.py
+6
@@ -9,10 +9,16 @@ import pytest
9
from transformers import BitsAndBytesConfig
10
11
from tests.quantization.utils import is_quant_method_supported
12
+from vllm.platforms import current_platform
13
14
from ...utils import compare_two_settings, multi_gpu_test
15
from ..utils import check_embeddings_close, check_logprobs_close
16
17
+pytestmark = pytest.mark.skipif(
18
+ current_platform.is_rocm(),
19
+ reason="bitsandbytes quantization not supported on ROCm (CUDA-only kernels)",
20
+)
21
+
22
models_4bit_to_test = [
23
("facebook/opt-125m", "quantize opt model inflight"),
24
(
vllm/platforms/rocm.py
+1
-1
@@ -413,7 +413,7 @@ class RocmPlatform(Platform):
413
"Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ"
414
" is not set, enabling VLLM_USE_TRITON_AWQ."
415
)
416
- envs.VLLM_USE_TRITON_AWQ = True
416
+ os.environ["VLLM_USE_TRITON_AWQ"] = "1"
417
418
@classmethod
419
def get_punica_wrapper(cls) -> str: