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 15, 07:15 UTC
[ROCm][Build] Clean up the ROCm build (#19040)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Gregory Shtrasberg committed
Jun 2, 2025 at 23:47 UTC
9e6f61e8c3df833537e4bea6c33f85eca5d73b15
4 files changed
+2
-23
CMakeLists.txt
-4
@@ -182,9 +182,6 @@ include(FetchContent)
182
file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
183
message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
184
185
-#
186
-# Set rocm version dev int.
187
-#
185
if(VLLM_GPU_LANG STREQUAL "HIP")
186
#
187
# Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
@@ -192,7 +189,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
189
set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
190
set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
191
195
-
192
#
193
# Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
194
# a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
docker/Dockerfile.rocm
-17
@@ -1,7 +1,5 @@
1
# default base image
2
ARG REMOTE_VLLM="0"
3
-ARG USE_CYTHON="0"
4
-ARG BUILD_RPD="1"
3
ARG COMMON_WORKDIR=/app
4
ARG BASE_IMAGE=rocm/vllm-dev:base
5
@@ -36,12 +34,10 @@ FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm
34
# -----------------------
35
# vLLM build stages
36
FROM fetch_vllm AS build_vllm
39
-ARG USE_CYTHON
37
# Build vLLM
38
RUN cd vllm \
39
&& python3 -m pip install -r requirements/rocm.txt \
40
&& python3 setup.py clean --all \
44
- && if [ ${USE_CYTHON} -eq "1" ]; then python3 tests/build_cython.py build_ext --inplace; fi \
41
&& python3 setup.py bdist_wheel --dist-dir=dist
42
FROM scratch AS export_vllm
43
ARG COMMON_WORKDIR
@@ -90,13 +86,6 @@ RUN case "$(which python3)" in \
86
*) ;; esac
87
88
RUN python3 -m pip install --upgrade huggingface-hub[cli]
93
-ARG BUILD_RPD
94
-RUN if [ ${BUILD_RPD} -eq "1" ]; then \
95
- git clone -b nvtx_enabled https://github.com/ROCm/rocmProfileData.git \
96
- && cd rocmProfileData/rpd_tracer \
97
- && pip install -r requirements.txt && cd ../ \
98
- && make && make install \
99
- && cd hipMarker && python3 setup.py install ; fi
89
90
# Install vLLM
91
RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
@@ -117,12 +106,6 @@ ENV TOKENIZERS_PARALLELISM=false
106
# ENV that can improve safe tensor loading, and end-to-end time
107
ENV SAFETENSORS_FAST_GPU=1
108
120
-# User-friendly environment setting for multi-processing to avoid below RuntimeError.
121
-# RuntimeError: Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing,
122
-# you must use the 'spawn' start method
123
-# See https://pytorch.org/docs/stable/notes/multiprocessing.html#cuda-in-multiprocessing
124
-ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
125
-
109
# Performance environment variable.
110
ENV HIP_FORCE_DEV_KERNARG=1
111
docs/getting_started/installation/gpu/rocm.inc.md
-2
@@ -179,8 +179,6 @@ It is important that the user kicks off the docker build using buildkit. Either
179
It provides flexibility to customize the build of docker image using the following arguments:
180
181
- `BASE_IMAGE`: specifies the base image used when running `docker build`. The default value `rocm/vllm-dev:base` is an image published and maintained by AMD. It is being built using <gh-file:docker/Dockerfile.rocm_base>
182
-- `USE_CYTHON`: An option to run cython compilation on a subset of python files upon docker build
183
-- `BUILD_RPD`: Include RocmProfileData profiling tool in the image
182
- `ARG_PYTORCH_ROCM_ARCH`: Allows to override the gfx architecture values from the base docker image
183
184
Their values can be passed in when running `docker build` with `--build-arg` options.
requirements/rocm.txt
+2
@@ -12,5 +12,7 @@ ray>=2.10.0,<2.45.0
12
peft
13
pytest-asyncio
14
tensorizer>=2.9.0
15
+setuptools-scm>=8
16
+setuptools>=77.0.3,<80.0.0
17
runai-model-streamer==0.11.0
18
runai-model-streamer-s3==0.11.0