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 7, 02:45 UTC
[CI] Enable Crosslayer KV layout tests for ROCm platforms (#35416)
qli88 committed
Mar 7, 2026 at 15:49 UTC
eebd14651f7618eddda5e79eab2d4ea0cdcc1770
2 files changed
+44
-13
.buildkite/test-amd.yaml
+28
@@ -1486,6 +1486,20 @@ steps:
1486
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1487
- DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1488
1489
+- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
1490
+ mirror_hardwares: [amdexperimental, amdproduction]
1491
+ agent_pool: mi325_4
1492
+ # grade: Blocking
1493
+ timeout_in_minutes: 30
1494
+ working_dir: "/vllm-workspace/tests"
1495
+ num_devices: 4
1496
+ source_file_dependencies:
1497
+ - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
1498
+ - tests/v1/kv_connector/nixl_integration/
1499
+ commands:
1500
+ - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
1501
+ - CROSS_LAYERS_BLOCKS=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
1502
+
1503
##### multi gpus test #####
1504
##### A100 test #####
1505
@@ -3136,6 +3150,20 @@ steps:
3150
- uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3151
- DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3152
3153
+- label: CrossLayer KV layout Distributed NixlConnector PD accuracy tests (4 GPUs)
3154
+ mirror_hardwares: [amdexperimental, amdproduction]
3155
+ agent_pool: mi355_4
3156
+ # grade: Blocking
3157
+ timeout_in_minutes: 30
3158
+ working_dir: "/vllm-workspace/tests"
3159
+ num_devices: 4
3160
+ source_file_dependencies:
3161
+ - vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
3162
+ - tests/v1/kv_connector/nixl_integration/
3163
+ commands:
3164
+ - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
3165
+ - CROSS_LAYERS_BLOCKS=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
3166
+
3167
##### multi gpus test #####
3168
##### A100 test #####
3169
tests/v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+16
-13
@@ -56,24 +56,27 @@ run_tests() {
56
echo "✅ All ${label} tests passed!"
57
}
58
59
-# Run tests
59
+# Set backend
60
+label="default backend"
61
+cmdline_args=""
62
if [[ -n "${ROCM_ATTN:-}" ]]; then
63
echo "ROCM_ATTN is set, running with --attention-backend ROCM_ATTN"
62
- run_tests "ROCM_ATTN backend" "--attention-backend ROCM_ATTN"
63
-else
64
- run_tests "default backend" ""
65
-fi
66
-
67
-# Check if FLASHINFER is set (non-empty)
68
-if [[ -n "${FLASHINFER:-}" ]]; then
69
- echo "FLASHINFER is set, rerunning with --attention-backend FLASHINFER"
70
- run_tests "FLASHINFER backend" "--attention-backend FLASHINFER"
64
+ label="ROCM_ATTN backend"
65
+ cmdline_args=" --attention-backend ROCM_ATTN "
66
+elif [[ -n "${FLASHINFER:-}" ]]; then
67
+ echo "FLASHINFER is set, running with --attention-backend FLASHINFER"
68
+ label="FLASHINFER backend"
69
+ cmdline_args=" --attention-backend FLASHINFER "
70
else
72
- echo "FLASHINFER not set, skipping FLASHINFER runs."
71
+ echo "running with default attention backend"
72
fi
73
74
# Check if cross-layers is enabled (non-empty)
75
if [[ -n "${CROSS_LAYERS_BLOCKS:-}" ]]; then
77
- echo "CROSS_LAYERS_BLOCKS is set, rerunning with --enable-cross-layers"
78
- run_tests "default backend" "--enable-cross-layers"
76
+ echo "CROSS_LAYERS_BLOCKS is set, running with --enable-cross-layers"
77
+ label+=" - CROSS_LAYERS_BLOCKS enabled"
78
+ cmdline_args+=" --enable-cross-layers "
79
fi
80
+
81
+# Run tests
82
+run_tests "${label}" "${cmdline_args}"