[ROCm][CI] Fix Kimi K3 KDA on ROCm (#50262)
Signed-off-by: Stefan Koncarevic <stefan.koncarevic@amd.com>
stefankoncarevic committed
Jul 29, 2026 at 17:47 UTC
381b6916202035f997467324a6b1495ab84d98a6
2 files changed
+8
-2
.buildkite/test-amd.yaml
+6
-2
@@ -1576,10 +1576,14 @@ steps:
1576
- vllm/third_party/flash_linear_attention/ops/kda.py
1577
- vllm/third_party/flash_linear_attention/ops/chunk_delta_h.py
1578
- vllm/third_party/flash_linear_attention/ops/l2norm.py
1579
- - tests/kernels/test_kda.py
1579
+ - vllm/models/kimi_k3/nvidia/kda.py
1580
+ - vllm/models/kimi_k3/nvidia/kda_metadata.py
1581
+ - vllm/models/kimi_k3/nvidia/ops/third_party/kda/
1582
+ - tests/models/kimi_k3/test_kda.py
1583
+ - tests/models/kimi_k3/test_kda_metadata.py
1584
- vllm/platforms/rocm.py
1585
commands:
1582
- - pytest -v -s kernels/test_kda.py
1586
+ - pytest -v -s models/kimi_k3/test_kda.py models/kimi_k3/test_kda_metadata.py
1587
1588
- label: Kernels Mamba Test # TBD
1589
timeout_in_minutes: 180
vllm/models/kimi_k3/nvidia/kda.py
+2
@@ -164,6 +164,8 @@ def is_flashkda_supported(
164
dtype: torch.dtype,
165
lower_bound: float | None,
166
) -> bool:
167
+ if not current_platform.is_cuda():
168
+ return False
169
capability = current_platform.get_device_capability()
170
return (
171
capability is not None