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 16, 22:00 UTC
Build some nightly wheels by default (#6380)
Simon Mo committed
Jul 12, 2024 at 13:56 UTC
d719ba24c5b4e669bf51b49293cab09f2ce7361c
2 files changed
+46
-2
.buildkite/release-pipeline.yaml
+32
-2
@@ -1,5 +1,24 @@
1
steps:
2
- - block: "Build wheels"
2
+ - label: "Build wheel default - Python {{matrix.python_version}}, CUDA {{matrix.cuda_version}}"
3
+ agents:
4
+ queue: cpu_queue
5
+ commands:
6
+ - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --build-arg PYTHON_VERSION={{matrix.python_version}} --tag vllm-ci:build-image -e CMAKE_BUILD_TYPE=Release --target build --progress plain ."
7
+ - "mkdir artifacts"
8
+ - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image cp -r dist /artifacts_host"
9
+ # rename the files to change linux -> manylinux1
10
+ - "for f in artifacts/dist/*.whl; do mv -- \"$f\" \"${f/linux/manylinux1}\"; done"
11
+ - "aws s3 cp --recursive --acl public-read artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
12
+ - "aws s3 cp --recursive --acl public-read artifacts/dist s3://vllm-wheels/nightly/"
13
+ matrix:
14
+ setup:
15
+ cuda_version:
16
+ - "12.1.0"
17
+ python_version:
18
+ - "3.10"
19
+ - "3.11"
20
+
21
+ - block: "Build wheels full"
22
23
- label: "Build wheel - Python {{matrix.python_version}}, CUDA {{matrix.cuda_version}}"
24
agents:
@@ -8,7 +27,9 @@ steps:
27
- "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg CUDA_VERSION={{matrix.cuda_version}} --build-arg PYTHON_VERSION={{matrix.python_version}} --tag vllm-ci:build-image --target build --progress plain ."
28
- "mkdir artifacts"
29
- "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image cp -r dist /artifacts_host"
11
- - "aws s3 cp --recursive artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
30
+ - "for f in artifacts/dist/*.whl; do mv -- \"$f\" \"${f/linux/manylinux1}\"; done
31
+ - "aws s3 cp --recursive --acl public-read artifacts/dist s3://vllm-wheels/$BUILDKITE_COMMIT/"
32
+ - "aws s3 cp --recursive --acl public-read artifacts/dist s3://vllm-wheels/nightly/"
33
matrix:
34
setup:
35
cuda_version:
@@ -19,3 +40,12 @@ steps:
40
- "3.9"
41
- "3.10"
42
- "3.11"
43
+ adjustments:
44
+ - with:
45
+ cuda_version: "12.1.0"
46
+ python_version: "3.10"
47
+ skip: true
48
+ - with:
49
+ cuda_version: "12.1.0"
50
+ python_version: "3.11"
51
+ skip: true
docs/source/getting_started/installation.rst
+14
@@ -42,6 +42,20 @@ You can install vLLM using pip:
42
43
Therefore, it is recommended to install vLLM with a **fresh new** conda environment. If either you have a different CUDA version or you want to use an existing PyTorch installation, you need to build vLLM from source. See below for instructions.
44
45
+.. note::
46
+
47
+ vLLM also publishes a subset of wheels (Python 3.10, 3.11 with CUDA 12) for every commit since v0.5.3. You can download them with the following command:
48
+
49
+ .. code-block:: console
50
+
51
+ $ export VLLM_VERSION=0.5.2 # vLLM's main branch version is currently set to latest released tag
52
+ $ export PYTHON_VERSION=310
53
+ $ pip install https://vllm-wheels.s3.us-west-2.amazonaws.com/nightly/vllm-${VLLM_VERSION}-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl
54
+ $ # You can also access a specific commit
55
+ $ # export VLLM_COMMIT=...
56
+ $ # pip install https://vllm-wheels.s3.us-west-2.amazonaws.com/${VLLM_COMMIT}/vllm-${VLLM_VERSION}-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl
57
+
58
+
59
.. _build_from_source:
60
61
Build from source