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 14, 05:30 UTC
[ROCm] [Doc] Update ROCm installation docs (#27327)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Oct 28, 2025 at 13:00 UTC
5b3c35a68e049a93bd312c0e2012514b51875ed8
2 files changed
+75
-50
docs/getting_started/installation/gpu.rocm.inc.md
+74
-49
@@ -1,6 +1,6 @@
1
# --8<-- [start:installation]
2
3
-vLLM supports AMD GPUs with ROCm 6.3 or above.
3
+vLLM supports AMD GPUs with ROCm 6.3 or above, and torch 2.8.0 and above.
4
5
!!! tip
6
[Docker](#set-up-using-docker) is the recommended way to use vLLM on ROCm.
@@ -28,57 +28,63 @@ Currently, there are no pre-built ROCm wheels.
28
# --8<-- [end:pre-built-wheels]
29
# --8<-- [start:build-wheel-from-source]
30
31
+!!! tip
32
+ - If you found that the following installation step does not work for you, please refer to [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base). Dockerfile is a form of installation steps.
33
+
34
0. Install prerequisites (skip if you are already in an environment/docker with the following installed):
35
36
- [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/index.html)
37
- [PyTorch](https://pytorch.org/)
38
36
- For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm6.4.3_ubuntu24.04_py3.12_pytorch_release_2.6.0`, `rocm/pytorch-nightly`. If you are using docker image, you can skip to Step 3.
39
+ For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm7.0_ubuntu22.04_py3.10_pytorch_release_2.8.0`, `rocm/pytorch-nightly`. If you are using docker image, you can skip to Step 3.
40
41
Alternatively, you can install PyTorch using PyTorch wheels. You can check PyTorch installation guide in PyTorch [Getting Started](https://pytorch.org/get-started/locally/). Example:
42
43
```bash
44
# Install PyTorch
45
pip uninstall torch -y
43
- pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/rocm6.4
46
+ pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm7.0
47
```
48
46
-1. Install [Triton for ROCm](https://github.com/triton-lang/triton)
49
+1. Install [Triton for ROCm](https://github.com/ROCm/triton.git)
50
48
- Install ROCm's Triton (the default triton-mlir branch) following the instructions from [ROCm/triton](https://github.com/ROCm/triton/blob/triton-mlir/README.md)
51
+ Install ROCm's Triton following the instructions from [ROCm/triton](https://github.com/ROCm/triton.git)
52
53
```bash
54
python3 -m pip install ninja cmake wheel pybind11
55
pip uninstall -y triton
53
- git clone https://github.com/triton-lang/triton.git
56
+ git clone https://github.com/ROCm/triton.git
57
cd triton
55
- git checkout e5be006
58
+ # git checkout $TRITON_BRANCH
59
+ git checkout f9e5bf54
60
if [ ! -f setup.py ]; then cd python; fi
61
python3 setup.py install
62
cd ../..
63
```
64
65
!!! note
62
- If you see HTTP issue related to downloading packages during building triton, please try again as the HTTP error is intermittent.
66
+ - The validated `$TRITON_BRANCH` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
67
+ - If you see HTTP issue related to downloading packages during building triton, please try again as the HTTP error is intermittent.
68
64
-2. Optionally, if you choose to use CK flash attention, you can install [flash attention for ROCm](https://github.com/Dao-AILab/flash-attention)
69
+2. Optionally, if you choose to use CK flash attention, you can install [flash attention for ROCm](https://github.com/Dao-AILab/flash-attention.git)
70
66
- Install ROCm's flash attention (v2.7.2) following the instructions from [ROCm/flash-attention](https://github.com/ROCm/flash-attention#amd-rocm-support)
67
- Alternatively, wheels intended for vLLM use can be accessed under the releases.
71
+ Install ROCm's flash attention (v2.8.0) following the instructions from [ROCm/flash-attention](https://github.com/Dao-AILab/flash-attention#amd-rocm-support)
72
69
- For example, for ROCm 6.3, suppose your gfx arch is `gfx90a`. To get your gfx architecture, run `rocminfo |grep gfx`.
73
+ For example, for ROCm 7.0, suppose your gfx arch is `gfx942`. To get your gfx architecture, run `rocminfo |grep gfx`.
74
75
```bash
76
git clone https://github.com/Dao-AILab/flash-attention.git
77
cd flash-attention
74
- git checkout 1a7f4dfa
78
+ # git checkout $FA_BRANCH
79
+ git checkout 0e60e394
80
git submodule update --init
76
- GPU_ARCHS="gfx90a" python3 setup.py install
81
+ GPU_ARCHS="gfx942" python3 setup.py install
82
cd ..
83
```
84
85
!!! note
81
- You might need to downgrade the "ninja" version to 1.10 as it is not used when compiling flash-attention-2 (e.g. `pip install ninja==1.10.2.4`)
86
+ - The validated `$FA_BRANCH` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
87
+
88
89
3. If you choose to build AITER yourself to use a certain branch or commit, you can build AITER using the following steps:
90
@@ -92,11 +98,13 @@ Currently, there are no pre-built ROCm wheels.
98
```
99
100
!!! note
95
- You will need to config the `$AITER_BRANCH_OR_COMMIT` for your purpose.
101
+ - You will need to config the `$AITER_BRANCH_OR_COMMIT` for your purpose.
102
+ - The validated `$AITER_BRANCH_OR_COMMIT` can be found in the [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base).
103
+
104
97
-4. Build vLLM. For example, vLLM on ROCM 6.3 can be built with the following steps:
105
+4. Build vLLM. For example, vLLM on ROCM 7.0 can be built with the following steps:
106
99
- ??? console "Commands"
107
+ ???+ console "Commands"
108
109
```bash
110
pip install --upgrade pip
@@ -109,31 +117,48 @@ Currently, there are no pre-built ROCm wheels.
117
scipy \
118
huggingface-hub[cli,hf_transfer] \
119
setuptools_scm
112
- pip install "numpy<2"
120
pip install -r requirements/rocm.txt
121
115
- # Build vLLM for MI210/MI250/MI300.
116
- export PYTORCH_ROCM_ARCH="gfx90a;gfx942"
122
+ # To build for a single architecture (e.g., MI300) for faster installation (recommended):
123
+ export PYTORCH_ROCM_ARCH="gfx942"
124
+
125
+ # To build vLLM for multiple arch MI210/MI250/MI300, use this instead
126
+ # export PYTORCH_ROCM_ARCH="gfx90a;gfx942"
127
+
128
python3 setup.py develop
129
```
130
131
This may take 5-10 minutes. Currently, `pip install .` does not work for ROCm installation.
132
133
!!! tip
123
- - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm-up step before collecting perf numbers.
124
- - Triton flash attention does not currently support sliding window attention. If using half precision, please use CK flash-attention for sliding window support.
125
- - To use CK flash-attention or PyTorch naive attention, please use this flag `export VLLM_USE_TRITON_FLASH_ATTN=0` to turn off triton flash attention.
134
- The ROCm version of PyTorch, ideally, should match the ROCm driver version.
135
136
!!! tip
137
- For MI300x (gfx942) users, to achieve optimal performance, please refer to [MI300x tuning guide](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/index.html) for performance optimization and tuning tips on system and workflow level.
130
- For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/workload.html#vllm-performance-optimization).
138
+ For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/rocm-for-ai/inference-optimization/vllm-optimization.html).
139
140
# --8<-- [end:build-wheel-from-source]
141
# --8<-- [start:pre-built-images]
142
143
The [AMD Infinity hub for vLLM](https://hub.docker.com/r/rocm/vllm/tags) offers a prebuilt, optimized
144
docker image designed for validating inference performance on the AMD Instinct™ MI300X accelerator.
145
+AMD also offers nightly prebuilt docker image from [Docker Hub](https://hub.docker.com/r/rocm/vllm-dev), which has vLLM and all its dependencies installed.
146
+
147
+???+ console "Commands"
148
+ ```bash
149
+ docker pull rocm/vllm-dev:nightly # to get the latest image
150
+ docker run -it --rm \
151
+ --network=host \
152
+ --group-add=video \
153
+ --ipc=host \
154
+ --cap-add=SYS_PTRACE \
155
+ --security-opt seccomp=unconfined \
156
+ --device /dev/kfd \
157
+ --device /dev/dri \
158
+ -v <path/to/your/models>:/app/models \
159
+ -e HF_HOME="/app/models" \
160
+ rocm/vllm-dev:nightly
161
+ ```
162
163
!!! tip
164
Please check [LLM inference performance validation on AMD Instinct MI300X](https://rocm.docs.amd.com/en/latest/how-to/performance-validation/mi300x/vllm-benchmark.html)
@@ -144,29 +169,29 @@ docker image designed for validating inference performance on the AMD Instinct
169
170
Building the Docker image from source is the recommended way to use vLLM with ROCm.
171
147
-#### (Optional) Build an image with ROCm software stack
172
+??? info "(Optional) Build an image with ROCm software stack"
173
149
-Build a docker image from [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base) which setup ROCm software stack needed by the vLLM.
150
-**This step is optional as this rocm_base image is usually prebuilt and store at [Docker Hub](https://hub.docker.com/r/rocm/vllm-dev) under tag `rocm/vllm-dev:base` to speed up user experience.**
151
-If you choose to build this rocm_base image yourself, the steps are as follows.
174
+ Build a docker image from [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base) which setup ROCm software stack needed by the vLLM.
175
+ **This step is optional as this rocm_base image is usually prebuilt and store at [Docker Hub](https://hub.docker.com/r/rocm/vllm-dev) under tag `rocm/vllm-dev:base` to speed up user experience.**
176
+ If you choose to build this rocm_base image yourself, the steps are as follows.
177
153
-It is important that the user kicks off the docker build using buildkit. Either the user put DOCKER_BUILDKIT=1 as environment variable when calling docker build command, or the user needs to set up buildkit in the docker daemon configuration /etc/docker/daemon.json as follows and restart the daemon:
178
+ It is important that the user kicks off the docker build using buildkit. Either the user put DOCKER_BUILDKIT=1 as environment variable when calling docker build command, or the user needs to set up buildkit in the docker daemon configuration /etc/docker/daemon.json as follows and restart the daemon:
179
155
-```json
156
-{
157
- "features": {
158
- "buildkit": true
180
+ ```json
181
+ {
182
+ "features": {
183
+ "buildkit": true
184
+ }
185
}
160
-}
161
-```
186
+ ```
187
163
-To build vllm on ROCm 6.3 for MI200 and MI300 series, you can use the default:
188
+ To build vllm on ROCm 7.0 for MI200 and MI300 series, you can use the default:
189
165
-```bash
166
-DOCKER_BUILDKIT=1 docker build \
167
- -f docker/Dockerfile.rocm_base \
168
- -t rocm/vllm-dev:base .
169
-```
190
+ ```bash
191
+ DOCKER_BUILDKIT=1 docker build \
192
+ -f docker/Dockerfile.rocm_base \
193
+ -t rocm/vllm-dev:base .
194
+ ```
195
196
#### Build an image with vLLM
197
@@ -181,7 +206,7 @@ It is important that the user kicks off the docker build using buildkit. Either
206
}
207
```
208
184
-[docker/Dockerfile.rocm](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm) uses ROCm 6.3 by default, but also supports ROCm 5.7, 6.0, 6.1, and 6.2, in older vLLM branches.
209
+[docker/Dockerfile.rocm](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm) uses ROCm 7.0 by default, but also supports ROCm 5.7, 6.0, 6.1, 6.2, 6.3, and 6.4, in older vLLM branches.
210
It provides flexibility to customize the build of docker image using the following arguments:
211
212
- `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 [docker/Dockerfile.rocm_base](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.rocm_base)
@@ -189,16 +214,16 @@ It provides flexibility to customize the build of docker image using the followi
214
215
Their values can be passed in when running `docker build` with `--build-arg` options.
216
192
-To build vllm on ROCm 6.3 for MI200 and MI300 series, you can use the default:
217
+To build vllm on ROCm 7.0 for MI200 and MI300 series, you can use the default:
218
194
-```bash
195
-DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.rocm -t vllm-rocm .
196
-```
219
+???+ console "Commands"
220
+ ```bash
221
+ DOCKER_BUILDKIT=1 docker build -f docker/Dockerfile.rocm -t vllm-rocm .
222
+ ```
223
224
To run the above docker image `vllm-rocm`, use the below command:
225
200
-??? console "Command"
201
-
226
+???+ console "Commands"
227
```bash
228
docker run -it \
229
--network=host \
docs/getting_started/installation/python_env_setup.inc.md
+1
-1
@@ -1,4 +1,4 @@
1
-It's recommended to use [uv](https://docs.astral.sh/uv/), a very fast Python environment manager, to create and manage Python environments. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following commands:
1
+On NVIDIA CUDA only, it's recommended to use [uv](https://docs.astral.sh/uv/), a very fast Python environment manager, to create and manage Python environments. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following commands:
2
3
```bash
4
uv venv --python 3.12 --seed