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 10, 05:15 UTC
[ROCm][CI] Disable Async Scheduling For Qwen3-Next-80B-A3B-Instruct MTP Async EPLB Accuracy Test (#32275)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Micah Williamson committed
Jan 13, 2026 at 23:29 UTC
6fa6e7ef0c9b485e8a684211e96691731aad6faa
1 file changed
+5
-1
.buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh
+5
-1
@@ -18,15 +18,18 @@ wait_for_server() {
18
19
MODEL="Qwen/Qwen3-Next-80B-A3B-Instruct"
20
21
-# Set BACKENDS based on platform
21
+# Set BACKENDS and platform-specific args based on platform
22
if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
23
# ROCm platform
24
BACKENDS=("allgather_reducescatter")
25
# Disable MOE padding for ROCm since it is causing eplb to fail
26
export VLLM_ROCM_MOE_PADDING=0
27
+ PLATFORM_ARGS=("--no-async-scheduling")
28
+ echo "Disabled async scheduling for ROCm platform due to issues with spec decode."
29
else
30
# Non-ROCm platform (CUDA/other)
31
BACKENDS=("deepep_high_throughput" "deepep_low_latency")
32
+ PLATFORM_ARGS=()
33
fi
34
35
cleanup() {
@@ -54,6 +57,7 @@ for BACK in "${BACKENDS[@]}"; do
57
--trust-remote-code \
58
--max-model-len 2048 \
59
--gpu-memory-utilization 0.9 \
60
+ "${PLATFORM_ARGS[@]}" \
61
--port $PORT &
62
SERVER_PID=$!
63
wait_for_server $PORT