[ROCm][CI] Adding qwen3 dp4 eplb (#47480)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Jul 2, 2026 at 19:58 UTC
6768fbc76f47bf8d9c50ea5e4f4b2cf633c3b4b3
2 files changed
+25
.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_dp4_async_eplb.sh
+4
@@ -18,6 +18,10 @@ wait_for_server() {
18
19
MODEL="Qwen/Qwen3-30B-A3B-FP8"
20
BACK="allgather_reducescatter"
21
+if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
22
+ # Disable MOE padding for ROCm since it is causing eplb to fail.
23
+ export VLLM_ROCM_MOE_PADDING=0
24
+fi
25
26
cleanup() {
27
if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
.buildkite/test-amd.yaml
+21
@@ -1159,6 +1159,27 @@ steps:
1159
commands:
1160
- bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
1161
1162
+- label: Qwen3-30B-A3B-FP8 DP4 Async EPLB Accuracy (4xH100-4xMI300) # TBD
1163
+ timeout_in_minutes: 180
1164
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
1165
+ agent_pool: mi300_4
1166
+ num_gpus: 4
1167
+ optional: true
1168
+ working_dir: "/vllm-workspace"
1169
+ source_file_dependencies:
1170
+ - vllm/model_executor/models/
1171
+ - vllm/model_executor/model_loader/
1172
+ - vllm/model_executor/layers/quantization/
1173
+ - vllm/distributed/eplb
1174
+ - vllm/model_executor/layers/fused_moe/
1175
+ - vllm/v1/attention/backends/
1176
+ - vllm/v1/attention/selector.py
1177
+ - .buildkite/scripts/scheduled_integration_test/
1178
+ - vllm/_aiter_ops.py
1179
+ - vllm/platforms/rocm.py
1180
+ commands:
1181
+ - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_dp4_async_eplb.sh 0.8 200 8050
1182
+
1183
- label: Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy # TBD
1184
timeout_in_minutes: 180
1185
mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]