From bc599d0ee8bde1df132d46878d33892bbc975269 Mon Sep 17 00:00:00 2001 From: youkaichao Date: Tue, 17 Sep 2024 15:51:02 -0700 Subject: [PATCH 1/3] improve doc --- docs/source/getting_started/installation.rst | 2 ++ 1 file changed, 2 insertions(+) diff --git a/docs/source/getting_started/installation.rst b/docs/source/getting_started/installation.rst index 50a761b49490c..a08c1a24589b7 100644 --- a/docs/source/getting_started/installation.rst +++ b/docs/source/getting_started/installation.rst @@ -95,6 +95,8 @@ You can also build and install vLLM from source: $ export MAX_JOBS=6 $ pip install -e . + This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default `_, and you'd better use ``export MAX_JOBS=1`` to avoid running out of memory. The side effect is that the build process will be slower. It is okay if you only touch the Python code, as you can just change the code and run the Python script without any re-compilation or re-installation. + .. tip:: If you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image. From 8c0e24d06246c06479b8da96610a634ae3c2982d Mon Sep 17 00:00:00 2001 From: youkaichao Date: Tue, 17 Sep 2024 16:10:19 -0700 Subject: [PATCH 2/3] Update docs/source/getting_started/installation.rst Co-authored-by: Andy Dai <76841985+Imss27@users.noreply.github.com> --- docs/source/getting_started/installation.rst | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/source/getting_started/installation.rst b/docs/source/getting_started/installation.rst index a08c1a24589b7..f2d8118c06586 100644 --- a/docs/source/getting_started/installation.rst +++ b/docs/source/getting_started/installation.rst @@ -95,7 +95,7 @@ You can also build and install vLLM from source: $ export MAX_JOBS=6 $ pip install -e . - This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default `_, and you'd better use ``export MAX_JOBS=1`` to avoid running out of memory. The side effect is that the build process will be slower. It is okay if you only touch the Python code, as you can just change the code and run the Python script without any re-compilation or re-installation. + This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default `_, and you'd better use ``export MAX_JOBS=1`` to avoid compiling multiple cuda files simultaneously and running out of memory. The side effect is that the build process will be much slower. It is okay to build in editable mode for only once if you only touch the Python code, as you can just change the code and run the Python script without any re-compilation or re-installation. .. tip:: If you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image. From 5fe2743be8e8d8d900fcdf3febda8b21aee98da0 Mon Sep 17 00:00:00 2001 From: youkaichao Date: Tue, 17 Sep 2024 16:12:41 -0700 Subject: [PATCH 3/3] further polish --- docs/source/getting_started/installation.rst | 2 +- tests/compile/test_full_graph.py | 5 ++++- 2 files changed, 5 insertions(+), 2 deletions(-) diff --git a/docs/source/getting_started/installation.rst b/docs/source/getting_started/installation.rst index f2d8118c06586..0322503a89a56 100644 --- a/docs/source/getting_started/installation.rst +++ b/docs/source/getting_started/installation.rst @@ -95,7 +95,7 @@ You can also build and install vLLM from source: $ export MAX_JOBS=6 $ pip install -e . - This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default `_, and you'd better use ``export MAX_JOBS=1`` to avoid compiling multiple cuda files simultaneously and running out of memory. The side effect is that the build process will be much slower. It is okay to build in editable mode for only once if you only touch the Python code, as you can just change the code and run the Python script without any re-compilation or re-installation. + This is especially useful when you are building on less powerful machines. For example, when you use WSL, it only `gives you half of the memory by default `_, and you'd better use ``export MAX_JOBS=1`` to avoid compiling multiple files simultaneously and running out of memory. The side effect is that the build process will be much slower. If you only touch the Python code, slow compilation is okay, as you are building in an editable mode: you can just change the code and run the Python script without any re-compilation or re-installation. .. tip:: If you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image. diff --git a/tests/compile/test_full_graph.py b/tests/compile/test_full_graph.py index 6fc445539bbbe..2e309aaa58d48 100644 --- a/tests/compile/test_full_graph.py +++ b/tests/compile/test_full_graph.py @@ -28,7 +28,10 @@ def test_full_graph(model, tp_size): "The future of AI is", ] sampling_params = SamplingParams(temperature=0) - llm = LLM(model=model, enforce_eager=True, tensor_parallel_size=tp_size) + llm = LLM(model=model, + enforce_eager=True, + tensor_parallel_size=tp_size, + disable_custom_all_reduce=True) outputs = llm.generate(prompts, sampling_params)