Build with nightly Torch instead of the latest (#1681)

Summary:

Changes in the PyTorch C++ layer affect Monarch's tensor engine builds. Without this change, the CI builds will pick up the latest PyTorch through build-requirements.txt. This diff removes this dependency.

We see this affecting #1680

Reviewed By: amirafzali

Differential Revision: D85670693

Author: allenwang28

.github/workflows/build-cpu.yml

@@ -23,6 +23,9 @@ jobs:
       submodules: recursive
       upload-artifact: monarch-cpu-${{ github.sha }}
       script: |
+        # Install torch nightly (CPU version)
+        pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
+
         # Source common setup functions
         source scripts/common-setup.sh
 

.github/workflows/build-cuda.yml

@@ -28,6 +28,9 @@ jobs:
       submodules: recursive
       upload-artifact: monarch-cuda-${{ github.sha }}
       script: |
+        # Install torch nightly first
+        pip install ${{ matrix.torch-spec }}
+
         # Source common setup functions
         source scripts/common-setup.sh
 

.github/workflows/test-gpu-python.yml

@@ -45,6 +45,10 @@ jobs:
         # Setup CUDA environment and library paths
         setup_cuda_environment
 
+        # Install torch nightly before installing the wheel,
+        # so that we can test the wheel against the latest nightly
+        pip install ${{ matrix.torch-spec }}
+
         # Install the built wheel from artifact
         install_wheel_from_artifact
 

README.md

@@ -70,6 +70,9 @@ torchmonarch-nightly is built with torch nightly.
 
 ### Build and Install from Source
 
+If you're building Monarch from source, you should be building it with the nightly PyTorch as well for ABI compatibility.
+
+
 #### On Fedora distributions
 
 ```sh
@@ -87,7 +90,7 @@ rustup default nightly
 conda install libunwind -y
 
 # Install the correct cuda and cuda-toolkit versions for your machine
-sudo dnf install cuda-toolkit-12-0 cuda-12-0
+sudo dnf install cuda-toolkit-12-8 cuda-12-8
 
 # Install clang-dev and nccl-dev
 sudo dnf install clang-devel libnccl-devel
@@ -99,6 +102,7 @@ conda update -n monarchenv --all -c conda-forge -y
 sudo dnf install -y libibverbs rdma-core libmlx5 libibverbs-devel rdma-core-devel
 
 # Install build dependencies
+pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128
 pip install -r build-requirements.txt
 # Install test dependencies
 pip install -r python/tests/requirements.txt
@@ -134,7 +138,11 @@ sudo apt install -y clang
 export CC=clang
 export CXX=clang++
 
+# Install the correct cuda and cuda-toolkit versions for your machine
+sudo apt install -y cuda-toolkit-12-8 cuda-12-8
+
 # Install build dependencies
+pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128
 pip install -r build-requirements.txt
 # Install test dependencies
 pip install -r python/tests/requirements.txt
@@ -153,9 +161,9 @@ pip install --no-build-isolation -e .
 pip list | grep monarch
 ```
 
-#### On MacOS
+#### On non-CUDA machines
 
-You can also build Monarch to run locally on a MacOS system.
+You can also build Monarch to run on non-CUDA machines, e.g. locally on a MacOS system.
 
 Note that this does not support tensor engine, which is tied to CUDA and RDMA (via ibverbs).
 
@@ -172,6 +180,7 @@ rustup toolchain install nightly
 rustup default nightly
 
 # Install build dependencies
+pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cpu
 pip install -r build-requirements.txt
 # Install test dependencies
 pip install -r python/tests/requirements.txt