From 9a3101b2ba6821488f4b7a9b93124e479edc4d3e Mon Sep 17 00:00:00 2001 From: Charlie Fu Date: Fri, 21 Nov 2025 19:11:02 -0600 Subject: [PATCH] [Rocm][CI] Fix DeekSeek V2-Lite Accuracy CI (#29135) Signed-off-by: charlifu --- .../deepseek_v2_lite_ep_eplb.sh | 12 +++++++++++- .../qwen30b_a3b_fp8_block_ep.sh | 11 ++++++++++- 2 files changed, 21 insertions(+), 2 deletions(-) diff --git a/.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh b/.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh index 5302f524a0ae..8106f50f18f6 100644 --- a/.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh +++ b/.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh @@ -17,7 +17,17 @@ wait_for_server() { } MODEL="deepseek-ai/DeepSeek-V2-lite" -BACKENDS=("deepep_high_throughput" "deepep_low_latency") + +# Set BACKENDS based on platform +if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then + # ROCm platform + BACKENDS=("allgather_reducescatter") + # Disable MOE padding for ROCm since it is causing eplb to fail + export VLLM_ROCM_MOE_PADDING=0 +else + # Non-ROCm platform (CUDA/other) + BACKENDS=("deepep_high_throughput" "deepep_low_latency") +fi cleanup() { if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then diff --git a/.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh b/.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh index a5135299297e..0d06f53a183d 100644 --- a/.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh +++ b/.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh @@ -17,7 +17,16 @@ wait_for_server() { } MODEL="QWen/Qwen3-30B-A3B-FP8" -BACKENDS=("deepep_high_throughput" "deepep_low_latency") +# Set BACKENDS based on platform +if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then + # ROCm platform + BACKENDS=("allgather_reducescatter") + # Disable MOE padding for ROCm since it is causing eplb to fail + export VLLM_ROCM_MOE_PADDING=0 +else + # Non-ROCm platform (CUDA/other) + BACKENDS=("deepep_high_throughput" "deepep_low_latency") +fi cleanup() { if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then