A high-throughput and memory-efficient inference and serving engine for LLMs
[ROCm][CI] Re-route NixlConnector jobs (#44809)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jun 8, 2026 at 18:57 UTC
05cb606cad303f1a4af122bdbd5906dc3c98464d
2 files changed
+15
-11
.buildkite/test-amd.yaml
+10
-10
@@ -804,7 +804,7 @@ steps:
804
- vllm/platforms/rocm.py
805
commands:
806
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
807
- - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
807
+ - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
808
809
- label: V1 e2e (2 GPUs) # TBD
810
timeout_in_minutes: 180
@@ -830,7 +830,7 @@ steps:
830
- vllm/platforms/rocm.py
831
commands:
832
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
833
- - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
833
+ - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
834
835
#------------------------------------------------------------- mi250 · misc ------------------------------------------------------------#
836
@@ -2405,7 +2405,7 @@ steps:
2405
- vllm/platforms/rocm.py
2406
commands:
2407
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2408
- - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2408
+ - CROSS_LAYERS_BLOCKS=True ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2409
2410
- label: Distributed DP Tests (4 GPUs) # TBD
2411
timeout_in_minutes: 180
@@ -2441,7 +2441,7 @@ steps:
2441
- vllm/platforms/rocm.py
2442
commands:
2443
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2444
- - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2444
+ - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2445
2446
- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
2447
timeout_in_minutes: 180
@@ -2455,7 +2455,7 @@ steps:
2455
- vllm/platforms/rocm.py
2456
commands:
2457
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2458
- - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2458
+ - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2459
2460
- label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
2461
timeout_in_minutes: 180
@@ -2469,7 +2469,7 @@ steps:
2469
- vllm/platforms/rocm.py
2470
commands:
2471
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
2472
- - HYBRID_SSM=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2472
+ - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
2473
2474
- label: V1 e2e (4 GPUs) # TBD
2475
timeout_in_minutes: 180
@@ -3070,7 +3070,7 @@ steps:
3070
- vllm/_aiter_ops.py
3071
commands:
3072
- rocm-smi
3073
- - python3 examples/basic/offline_inference/chat.py
3073
+ - python3 examples/basic/offline_inference/chat.py --attention-backend TRITON_ATTN
3074
- pytest -v -s tests/kernels/attention/test_attention_selector.py
3075
3076
- label: Kernels Attention Test %N # TBD
@@ -3434,7 +3434,7 @@ steps:
3434
- vllm/platforms/rocm.py
3435
commands:
3436
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3437
- - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
3437
+ - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
3438
3439
- label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
3440
timeout_in_minutes: 180
@@ -3449,7 +3449,7 @@ steps:
3449
- vllm/platforms/rocm.py
3450
commands:
3451
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3452
- - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3452
+ - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3453
3454
- label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
3455
timeout_in_minutes: 180
@@ -3464,7 +3464,7 @@ steps:
3464
- vllm/platforms/rocm.py
3465
commands:
3466
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3467
- - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3467
+ - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3468
3469
#------------------------------------------------------ mi355 · weight_loading -------------------------------------------------------#
3470
tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+5
-1
@@ -75,7 +75,11 @@ run_tests() {
75
# Set backend
76
label="default backend"
77
cmdline_args=""
78
-if [[ -n "${ROCM_ATTN:-}" ]]; then
78
+if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
79
+ echo "ATTENTION_BACKEND is set, running with --attention-backend ${ATTENTION_BACKEND}"
80
+ label="${ATTENTION_BACKEND} backend"
81
+ cmdline_args=" --attention-backend ${ATTENTION_BACKEND} "
82
+elif [[ -n "${ROCM_ATTN:-}" ]]; then
83
echo "ROCM_ATTN is set, running with --attention-backend ROCM_ATTN"
84
label="ROCM_ATTN backend"
85
cmdline_args=" --attention-backend ROCM_ATTN "