[CI/Build][Bugfix]Fix Quantized Models Test on AMD (#27712)

Signed-off-by: zhewenli <zhewenli@meta.com>

.buildkite/test-amd.yaml

@@ -908,7 +908,7 @@ steps:
 
 - label: Quantized Models Test # 45 min
   timeout_in_minutes: 60
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:

tests/models/quantization/test_bitsandbytes.py

@@ -9,10 +9,16 @@ import pytest
 from transformers import BitsAndBytesConfig
 
 from tests.quantization.utils import is_quant_method_supported
+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)",
+)
+
 models_4bit_to_test = [
     ("facebook/opt-125m", "quantize opt model inflight"),
     (

vllm/platforms/rocm.py

@@ -413,7 +413,7 @@ class RocmPlatform(Platform):
                 "Using AWQ quantization with ROCm, but VLLM_USE_TRITON_AWQ"
                 " is not set, enabling VLLM_USE_TRITON_AWQ."
             )
-        envs.VLLM_USE_TRITON_AWQ = True
+        os.environ["VLLM_USE_TRITON_AWQ"] = "1"
 
     @classmethod
     def get_punica_wrapper(cls) -> str: