[ROCm][CI] Pin `test_rocm_compressed_tensors_w8a8` to TRITON_ATTN (#46180)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Jun 19, 2026 at 15:20 UTC
4a083cc858f075209dd964ade48c0f8ec87c3393
2 files changed
+5
-2
.buildkite/test_areas/kernels.yaml
+1
@@ -104,6 +104,7 @@ steps:
104
source_file_dependencies:
105
- csrc/quantization/
106
- vllm/model_executor/layers/quantization
107
+ - vllm/config/
108
- tests/kernels/quantization
109
- tests/kernels/quantization/test_rocm_skinny_gemms.py
110
- vllm/_aiter_ops.py
tests/kernels/quantization/test_triton_scaled_mm.py
+4
-2
@@ -60,8 +60,10 @@ def test_rocm_compressed_tensors_w8a8(
60
vllm_runner, example_prompts, model_path, max_tokens, num_logprobs
61
):
62
dtype = "bfloat16"
63
-
64
- with vllm_runner(model_path, dtype=dtype) as vllm_model:
63
+ # Pin to TRITON_ATTN, see https://github.com/vllm-project/vllm/issues/46179
64
+ with vllm_runner(
65
+ model_path, dtype=dtype, attention_backend="TRITON_ATTN"
66
+ ) as vllm_model:
67
vllm_model.generate_greedy_logprobs(example_prompts, max_tokens, num_logprobs)
68
69