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 5, 01:00 UTC
[AMD][CI] Update DeepEP branch (#38396)
Signed-off-by: Ryan Rock <ryan.rock@amd.com>
Ryan Rock committed
Apr 17, 2026 at 14:30 UTC
58da4ee047a97b71776488301f66612838b79788
2 files changed
+8
-9
.buildkite/test-amd.yaml
+1
-1
@@ -2613,6 +2613,7 @@ steps:
2613
- vllm/platforms/rocm.py
2614
commands:
2615
- export TORCH_NCCL_BLOCKING_WAIT=1
2616
+ - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
2617
- pytest -v -s tests/v1/distributed/test_dbo.py
2618
2619
@@ -3601,7 +3602,6 @@ steps:
3602
commands:
3603
- export TORCH_NCCL_BLOCKING_WAIT=1
3604
- pytest -v -s tests/distributed/test_context_parallel.py
3604
- - VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
3605
- pytest -v -s tests/v1/distributed/test_dbo.py
3606
3607
docker/Dockerfile.rocm
+7
-8
@@ -192,9 +192,10 @@ RUN cd /opt/rixl && mkdir -p /app/install && \
192
FROM base AS build_deep
193
ARG ROCSHMEM_BRANCH="ba0bf0f3"
194
ARG ROCSHMEM_REPO="https://github.com/ROCm/rocm-systems.git"
195
-ARG DEEPEP_BRANCH="e84464ec"
195
+ARG DEEPEP_BRANCH="5d90af8b"
196
ARG DEEPEP_REPO="https://github.com/ROCm/DeepEP.git"
197
ARG DEEPEP_NIC="cx7"
198
+ARG DEEPEP_ROCM_ARCH="gfx942;gfx950"
199
ENV ROCSHMEM_DIR=/opt/rocshmem
200
201
RUN git clone ${ROCSHMEM_REPO} \
@@ -202,13 +203,11 @@ RUN git clone ${ROCSHMEM_REPO} \
203
&& git checkout ${ROCSHMEM_BRANCH} \
204
&& mkdir -p projects/rocshmem/build \
205
&& cd projects/rocshmem/build \
205
- && cmake .. \
206
- -DCMAKE_INSTALL_PREFIX="${ROCSHMEM_DIR}" \
207
- -DROCM_PATH=/opt/rocm \
208
- -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
209
- -DUSE_EXTERNAL_MPI=OFF \
210
- && make -j \
211
- && make install
206
+ && bash ../scripts/build_configs/all_backends \
207
+ -DCMAKE_INSTALL_PREFIX="${ROCSHMEM_DIR}" \
208
+ -DROCM_PATH=/opt/rocm \
209
+ -DGPU_TARGETS="${DEEPEP_ROCM_ARCH}" \
210
+ -DUSE_EXTERNAL_MPI=OFF
211
212
# Build DeepEP wheel.
213
# DeepEP looks for rocshmem at ROCSHMEM_DIR.