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 12, 18:30 UTC
[Rocm][CI] Fix DeekSeek V2-Lite Accuracy CI (#29135)
Signed-off-by: charlifu <charlifu@amd.com>
Charlie Fu committed
Nov 21, 2025 at 19:11 UTC
9a3101b2ba6821488f4b7a9b93124e479edc4d3e
2 files changed
+21
-2
.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh
+11
-1
@@ -17,7 +17,17 @@ wait_for_server() {
17
}
18
19
MODEL="deepseek-ai/DeepSeek-V2-lite"
20
-BACKENDS=("deepep_high_throughput" "deepep_low_latency")
20
+
21
+# Set BACKENDS based on platform
22
+if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
23
+ # ROCm platform
24
+ BACKENDS=("allgather_reducescatter")
25
+ # Disable MOE padding for ROCm since it is causing eplb to fail
26
+ export VLLM_ROCM_MOE_PADDING=0
27
+else
28
+ # Non-ROCm platform (CUDA/other)
29
+ BACKENDS=("deepep_high_throughput" "deepep_low_latency")
30
+fi
31
32
cleanup() {
33
if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh
+10
-1
@@ -17,7 +17,16 @@ wait_for_server() {
17
}
18
19
MODEL="QWen/Qwen3-30B-A3B-FP8"
20
-BACKENDS=("deepep_high_throughput" "deepep_low_latency")
20
+# Set BACKENDS based on platform
21
+if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
22
+ # ROCm platform
23
+ BACKENDS=("allgather_reducescatter")
24
+ # Disable MOE padding for ROCm since it is causing eplb to fail
25
+ export VLLM_ROCM_MOE_PADDING=0
26
+else
27
+ # Non-ROCm platform (CUDA/other)
28
+ BACKENDS=("deepep_high_throughput" "deepep_low_latency")
29
+fi
30
31
cleanup() {
32
if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then