[WIP] Add nightly build

.github/packaging/pre_build_nightly.sh

@@ -0,0 +1,75 @@
+#!/bin/bash
+set -uo pipefail # TODO revert back to set -euo
+
+# Script runs relative to forge root
+CURRENT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+FORGE_ROOT="$(cd "$CURRENT_DIR/../.." && pwd)"
+VERSIONS_FILE="$FORGE_ROOT/assets/versions.sh"
+source "$VERSIONS_FILE"
+
+echo "Installing nightly dependencies for forge build"
+echo "PyTorch nightly already installed by test-infra via channel: nightly"
+
+# 1. Verify PyTorch nightly is installed
+python -c "import torch; print(f'PyTorch version: {torch.__version__}')"
+
+# 2. Install torchtitan nightly
+echo "Installing torchtitan nightly..."
+pip install --pre torchtitan \
+  --extra-index-url https://download.pytorch.org/whl/nightly/cu128
+
+# 3. Install torchmonarch-nightly
+echo "Installing torchmonarch-nightly..."
+pip install torchmonarch-nightly
+
+# 4. Install torchstore from main branch WITHOUT dependencies
+# Following monarch_forge.sh:580-588 pattern
+echo "Installing torchstore dependencies..."
+pip install pygtrie
+
+echo "Installing torchstore from main branch..."
+TORCHSTORE_DIR="/tmp/torchstore-build"
+mkdir -p "$TORCHSTORE_DIR"
+cd "$TORCHSTORE_DIR"
+git clone https://github.com/meta-pytorch/torchstore.git
+echo "DEBUG: git clone succeeded"
+
+cd torchstore
+echo "DEBUG: Changed into torchstore subdirectory"
+git checkout main
+pip install --no-deps .
+echo "DEBUG: torchstore installation succeeded"
+cd -
+
+# 5. Build vLLM from source (following internal pt2.sh:561-578 pattern)
+# Note: Cannot use pip install vllm==0.10.0 because PyPI version requires torch==2.7.0
+# vLLM has C++/CUDA extensions that must compile against our PyTorch nightly
+echo "Building vLLM ${VLLM_VERSION} from source against PyTorch nightly..."
+BUILD_DIR="/tmp/vllm-build"
+mkdir -p "$BUILD_DIR"
+cd "$BUILD_DIR"
+
+git clone https://github.com/vllm-project/vllm.git --branch "$VLLM_VERSION"
+cd vllm
+
+# Use existing torch (PyTorch nightly already installed)
+# This script patches vLLM's setup.py to use the installed PyTorch instead of downloading
+python use_existing_torch.py
+pip install -r requirements/build.txt
+
+# Clean up existing builds if needed
+rm -rf build/ *.egg-info/
+
+# Build and install vLLM (compiles C++/CUDA extensions against installed PyTorch)
+pip install --no-build-isolation .
+
+cd -
+
+# 6. Set nightly version in __init__.py
+echo "Setting nightly version..."
+NIGHTLY_VERSION="${BUILD_VERSION:-$(date +%Y.%m.%d)}"
+sed -i "s/__version__ = \".*\"/__version__ = \"${NIGHTLY_VERSION}\"/" "$FORGE_ROOT/src/forge/__init__.py"
+
+echo "Nightly dependency installation complete!"
+echo "Installed package versions:"
+pip list | grep -E "torch|torchtitan|vllm|monarch"

.github/workflows/nightly.yaml

@@ -0,0 +1,53 @@
+name: Build and upload forge nightly
+
+on:
+  schedule:
+    - cron: '0 0 * * *'  # Daily at midnight UTC
+  push:
+    branches:
+      - nightly
+      - add-nightly-build  # TODO: remove this line after PR is accepted. This is for testing.
+  workflow_dispatch:
+
+permissions:
+  id-token: write
+  contents: read
+
+jobs:
+  build:
+    name: forge-cu128-nightly
+    uses: pytorch/test-infra/.github/workflows/build_wheels_linux.yml@main
+    strategy:
+      fail-fast: false
+    with:
+      repository: meta-pytorch/forge
+      # TODO: reset to "" after PR is accepted. This is to make sure during testing, the workflow
+      # checkout the current branch.
+      ref: ${{ github.sha }}
+      test-infra-repository: pytorch/test-infra
+      test-infra-ref: main
+      run-smoke-test: false
+      wheel-nightly-policy: gha_workflow_nightly_build_wheels
+      wheel-upload-path: whl/nightly/forge/
+      package-name: forge
+      build-matrix: |
+        {
+          "include": [
+            {
+              "python_version": "3.12",
+              "gpu_arch_type": "cpu",
+              "gpu_arch_version": "12.8",
+              "desired_cuda": "cu128",
+              "container_image": "pytorch/manylinux2_28-builder:cuda12.8",
+              "package_type": "manywheel",
+              "build_name": "manywheel-py3_10-cuda12_8",
+              "validation_runner": "linux.12xlarge.memory",
+              "upload_to_base_bucket": "no",
+              "use_split_build": false
+            }
+          ]
+        }
+      pre-script: .github/packaging/pre_build_nightly.sh
+      post-script: .github/packaging/post_build_script.sh
+      trigger-event: ${{ github.event_name }}
+      build-platform: 'python-build-package'

README.md

@@ -46,11 +46,25 @@ Optional: By default, the packages installation uses conda. If you want to insta
 
 > **Note:** We are actively working on enabling pure `uv` installation. Currently, Conda is the recommended approach. `uv` support is not fully working at the moment but is being tracked in [issue #494](https://github.com/meta-pytorch/torchforge/issues/494).
 
-After install, you can run the following command and should see output confirming GRPO training is running (you need a minimum 3 GPU devices):
+### Nightly Installation
 
+To install the latest nightly build of torchforge with the newest features and fixes:
+
+```bash
+# Install PyTorch nightly
+pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128
+
+# Install torchforge nightly
+pip install forge --index-url https://download.pytorch.org/whl/nightly/forge/
 ```
-python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml
-```
+
+Nightly builds are updated daily and include the latest features and fixes from the main branch.
+Here is the list of key dependencies:
+1. torchmonarch: download torchmonarch-nightly from PyPI.
+2. vllm: build v0.10.0 from source with torch nightly. TODO: additional refactoring to upgrade to use vLLM nightly.
+3. torchtitan: download the nightly build
+4. torchstore: install from the github main branch since torchstore does not have a nightly build yet.
+
 
 ## Quick Start