Consolidate on install.sh, move to stable packages (#474)

Co-authored-by: joecummings <[email protected]>

Author: ebsmothers

.github/packaging/pre_build_cpu.sh

@@ -24,7 +24,7 @@ echo "wheel dir is $WHL_DIR"
 build_vllm() {
     cd "$BUILD_DIR"
 
-    git clone https://github.com/vllm-project/vllm.git --branch $VLLM_BRANCH
+    git clone https://github.com/vllm-project/vllm.git --branch $VLLM_VERSION
     cd "$BUILD_DIR/vllm"
 
     python use_existing_torch.py

.github/packaging/vllm_reqs_12_8.txt

@@ -1,11 +1,4 @@
-# These requirements were generated by running steps 1-3 of scripts/build_wheels.sh
-# then running pip freeze and manually removing the vllm dependency.
-# The intention of this file is to use these known requirements for a fixed
-# vLLM build to supplement a vLLM install from download.pytorch.org without
-# resorting to --extra-index-url https://download.pytorch.org/whl/nightly to find
-# vLLM dependencies (as this results in a ResolutionTooDeep error from pip).
-# See the file .github/workflows/gpu_test.yaml for an E2E forge installation using this approach.
-# TODO: this should be done way less hackily
+# This file was generated by running ./scripts/generate_vllm_reqs.sh
 aiohappyeyeballs==2.6.1
 aiohttp==3.13.1
 aiosignal==1.4.0
@@ -33,8 +26,8 @@ dnspython==2.8.0
 einops==0.8.1
 email-validator==2.3.0
 exceptiongroup==1.3.0
-fastapi==0.119.0
-fastapi-cli==0.0.13
+fastapi==0.119.1
+fastapi-cli==0.0.14
 fastapi-cloud-cli==0.3.1
 fastrlock==0.8.3
 filelock==3.19.1
@@ -94,7 +87,7 @@ prometheus-fastapi-instrumentator==7.1.0
 prometheus_client==0.23.1
 propcache==0.4.1
 protobuf==6.33.0
-psutil==7.1.0
+psutil==7.1.1
 py-cpuinfo==9.0.0
 pybase64==1.4.2
 pycountry==24.6.1
@@ -108,9 +101,9 @@ python-json-logger==4.0.0
 python-multipart==0.0.20
 PyYAML==6.0.3
 pyzmq==27.1.0
-ray==2.50.0
+ray==2.50.1
 referencing==0.37.0
-regex==2025.9.18
+regex==2025.10.23
 requests==2.32.5
 rich==14.2.0
 rich-toolkit==0.15.1
@@ -119,8 +112,8 @@ rpds-py==0.27.1
 safetensors==0.6.2
 scipy==1.15.3
 sentencepiece==0.2.1
-sentry-sdk==2.42.0
-setuptools-scm==9.2.1
+sentry-sdk==2.42.1
+setuptools-scm==9.2.2
 shellingham==1.5.4
 sniffio==1.3.1
 soundfile==0.13.1
@@ -134,11 +127,11 @@ torch==2.9.0+cu128
 tqdm==4.67.1
 transformers==4.57.1
 triton==3.5.0
-typer==0.19.2
+typer==0.20.0
 typing-inspection==0.4.2
 typing_extensions==4.15.0
 urllib3==2.5.0
-uvicorn==0.37.0
+uvicorn==0.38.0
 uvloop==0.22.1
 watchfiles==1.1.1
 websockets==15.0.1

.github/workflows/docs.yml

@@ -26,62 +26,19 @@ jobs:
           activate-environment: test
           python-version: '3.10'
           auto-activate: false
-      - name: Verify conda environment
-        shell: bash -l {0}
-        run: |
-          conda info
-          which python
-          which conda
       - name: Update pip
         shell: bash -l {0}
         run: python -m pip install --upgrade pip
-      - name: Install pytorch
-        shell: bash -l {0}
-        run: pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu130 --force-reinstall
-      - name: Install monarch
-        shell: bash -l {0}
-        run: pip install assets/ci/monarch_no_torch-0.1.0.dev20251010-py3-none-any.whl
       - name: Install torchforge
         shell: bash -l {0}
-        env:
-          GH_TOKEN: ${{ github.token }}
         run: ./scripts/install.sh
       - name: Install docs dependencies
         shell: bash -l {0}
         run: python -m pip install -r docs/requirements.txt
       - name: Build docs
         shell: bash -l {0}
         working-directory: docs
-        run: |
-          # Set up library paths to ensure all dependencies are available
-          # This is critical for monarch and other native dependencies that need libpython3.10.so.1.0
-          export LD_LIBRARY_PATH="${CONDA_PREFIX}/lib:${LD_LIBRARY_PATH:-}"
-
-          # Also set CUDA paths if needed
-          if [ -d "/usr/local/cuda-12.9" ]; then
-            export LD_LIBRARY_PATH="/usr/local/cuda-12.9/compat:${LD_LIBRARY_PATH}"
-            export CUDA_HOME=/usr/local/cuda-12.9
-          fi
-
-          # Verify dependencies can be imported before building docs
-          echo "Verifying dependencies..."
-          python -c "import forge; print('✓ torchforge imported successfully')"
-          python -c "import monarch; print('✓ monarch imported successfully')"
-
-          # Build docs with -W (warnings as errors) and --keep-going to see all issues
-          # Capture exit code but continue to see all errors
-          set +e
-          make html SPHINXOPTS="-W --keep-going"
-          BUILD_EXIT_CODE=$?
-          set -e
-
-          # Report results
-          if [ $BUILD_EXIT_CODE -ne 0 ]; then
-            echo "❌ Documentation build failed with warnings or errors (exit code: $BUILD_EXIT_CODE)"
-            exit $BUILD_EXIT_CODE
-          else
-            echo "✅ Documentation build completed successfully with no warnings or errors"
-          fi
+        run: make html
       - name: Upload docs artifact
         uses: actions/upload-artifact@v4
         with:

.github/workflows/gpu_test.yaml

@@ -1,9 +1,6 @@
-name: GPU Tests
+name: Unit Tests (GPU)
 
 on:
-  schedule:
-    # Runs at midnight every day
-    - cron:  '0 0 * * *'
   push:
     branches: [ main ]
   pull_request:
@@ -27,7 +24,7 @@ jobs:
     runs-on: linux.g5.12xlarge.nvidia.gpu
     strategy:
       matrix:
-        python-version: ['3.10']
+        python-version: ['3.10', '3.11', '3.12']
     steps:
       - name: Check out repo
         uses: actions/checkout@v4
@@ -40,26 +37,10 @@ jobs:
           python-version: ${{ matrix.python-version }}
       - name: Update pip
         run: python -m pip install --upgrade pip
-      - name: Install pinned torch nightly
-        run: python -m pip install --pre torch==2.9.0.dev20250905 --no-cache-dir --index-url https://download.pytorch.org/whl/nightly/cu129
-      - name: Download and install vLLM and its dependencies
-        # TODO: this honestly could not be hackier if I tried
-        run: |
-          python -m pip install -r .github/packaging/vllm_reqs_12_9.txt
-          python -m pip install vllm==0.10.1.dev0+g6d8d0a24c.d20251009.cu129 --no-cache-dir --index-url https://download.pytorch.org/whl/preview/forge
-      - name: Install Monarch
-        run: pip install torchmonarch==0.1.0rc1
-      - name: Install torchtitan and torchstore
-        run: |
-          python -m pip install git+https://github.com/pytorch/torchtitan.git
-          python -m pip install git+https://github.com/meta-pytorch/torchstore.git
-      - name: Install dependencies
-        run: python -m pip install --no-build-isolation -e ".[dev]"
+      - name: Install torchforge
+        run: ./scripts/install.sh
       - name: Run unit tests with coverage
         # TODO add all tests
-        run: |
-          export LD_PRELOAD=$CONDA/envs/test/lib/libpython3.10.so.1.0
-          export LD_LIBRARY_PATH=$CONDA/envs/test/lib/libpython3.10.so.1.0
-          pytest tests/unit_tests --cov=. --cov-report=xml --durations=20 -vv
+        run: pytest tests/unit_tests --cov=. --cov-report=xml --durations=20 -vv
       - name: Upload Coverage to Codecov
         uses: codecov/codecov-action@v3

.github/workflows/unit_test.yaml

@@ -150,10 +150,10 @@ if [ -f "$VERSIONS_FILE" ]; then
     log_info "Sourcing version information from: $VERSIONS_FILE"
     source "$VERSIONS_FILE"
 
-    if [ -n "$TORCHTITAN_COMMIT" ]; then
-        log_info "Installing torchtitan from commit: $TORCHTITAN_COMMIT"
+    if [ -n "$TORCHTITAN_COMMIT_MAST" ]; then
+        log_info "Installing torchtitan from commit: $TORCHTITAN_COMMIT_MAST"
         pip uninstall -y torchtitan
-        pip install "git+https://github.com/pytorch/torchtitan.git@$TORCHTITAN_COMMIT"
+        pip install "git+https://github.com/pytorch/torchtitan.git@$TORCHTITAN_COMMIT_MAST"
 
         if [ $? -eq 0 ]; then
             log_info "Torchtitan installed successfully"
@@ -162,7 +162,7 @@ if [ -f "$VERSIONS_FILE" ]; then
             exit 1
         fi
     else
-        log_error "TORCHTITAN_COMMIT not found in versions.sh"
+        log_error "TORCHTITAN_COMMIT_MAST not found in versions.sh"
         exit 1
     fi
 else

.meta/mast/env_setup.sh

@@ -1,7 +1,6 @@
 # <img width="35" height="35" alt="image" src="https://github.com/user-attachments/assets/2700a971-e5d6-4036-b03f-2f89c9791609" /> torchforge
 
 #### A PyTorch-native agentic RL library that lets you focus on algorithms—not infra.
-[![Unit Tests](https://github.com/meta-pytorch/forge/actions/workflows/unit_test.yaml/badge.svg?branch=main)](https://github.com/meta-pytorch/forge/actions/workflows/unit_test.yaml?query=branch%3Amain)
 [![GPU Tests](https://github.com/meta-pytorch/forge/actions/workflows/gpu_test.yaml/badge.svg?branch=main)](https://github.com/meta-pytorch/forge/actions/workflows/gpu_test.yaml?query=branch%3Amain)
 [![Documentation](https://img.shields.io/badge/Docs-meta--pytorch.org-blue?style=flat&logo=readthedocs&logoColor=white)](https://meta-pytorch.org/torchforge/)
 [![Discord](https://img.shields.io/badge/Discord-OpenEnv-7289da?style=flat&logo=discord&logoColor=white)](https://discord.gg/YsTYBh6PD9)
@@ -33,14 +32,11 @@ You can also find our notebook tutorials (coming soon)
 
 ### Basic
 
-torchforge requires the latest PyTorch nightly with [Monarch](https://github.com/meta-pytorch/monarch), [vLLM](https://github.com/vllm-project/vllm), and [torchtitan](https://github.com/pytorch/torchtitan). For convenience,
-we have pre-packaged these dependencies as wheels in assets/wheels. (Note that the basic install script
+torchforge requires PyTorch 2.9.0 with [Monarch](https://github.com/meta-pytorch/monarch), [vLLM](https://github.com/vllm-project/vllm), and [torchtitan](https://github.com/pytorch/torchtitan). (Note that the basic install script
 uses [DNF](https://docs.fedoraproject.org/en-US/quick-docs/dnf/), but could be easily extended to other Linux OS.)
 
-torchforge requires the Github CLI (gh) to download a compatible vLLM package. See [here](https://github.com/cli/cli#installation) for gh install instructions before continuting. Please login to gh with your Github account before continuing with `gh auth login`. You may use either https or ssh as the protocol for authentication.
-
 ```bash
-conda create -n forge python=3.10
+conda create -n forge python=3.12
 conda activate forge
 ./scripts/install.sh
 ```
@@ -53,11 +49,6 @@ After install, you can run the following command and should see output confirmin
 python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml
 ```
 
-If you need to re-build the wheels for whatever reason, you can do so with:
-```bash
-./scripts/build_wheels.sh
-```
-
 ## Quick Start
 
 To run SFT on a Llama3 8B model, run

README.md

@@ -59,17 +59,15 @@ def policy_version(self) -> int | None:
 
     @property
     def request_tensor(self) -> torch.Tensor:
-        request_tokens: torch.Tensor = self.completion.prompt_ids
-        tensor = torch.tensor(request_tokens, dtype=torch.long)
+        tensor: torch.Tensor = self.completion.prompt_ids.to(torch.long)
         if tensor.shape[0] < self.request_len:  # left pad
             diff = self.request_len - tensor.shape[0]
             tensor = F.pad(tensor, (diff, 0), value=self.pad_id)
         return tensor
 
     @property
     def response_tensor(self) -> torch.Tensor:
-        response_tokens: torch.Tensor = self.completion.token_ids
-        tensor = torch.tensor(response_tokens, dtype=torch.long)
+        tensor: torch.Tensor = self.completion.token_ids.to(torch.long)
         if tensor.shape[0] < self.response_len:  # right pad
             diff = self.response_len - tensor.shape[0]
             tensor = F.pad(tensor, (0, diff), value=self.pad_id)

apps/grpo/main.py

@@ -7,13 +7,12 @@
 # Version Configuration for Forge Wheel Building
 # This file contains all pinned versions and commits for dependencies
 
-# PyTorch version
-PYTORCH_VERSION="2.9.0.dev20250905"
+# Stable versions of upstream libraries for OSS repo
+PYTORCH_VERSION="2.9.0"
+VLLM_VERSION="v0.10.0"
+MONARCH_VERSION="0.1.0rc8"
+TORCHTITAN_VERSION="0.2.0"
+TORCHSTORE_VERSION="0.0.1.rc3"
 
-# vLLM branch
-VLLM_BRANCH="v0.10.0"
-
-# Commit hashes
-MONARCH_COMMIT="195503223b5c2896846171f60ac99dc6868f8f2c"
-TORCHTITAN_COMMIT="d0e25450bcac2332359b13fbda430dc701f073d4"
-TORCHSTORE_COMMIT="662299faf4fd50ee30bd9aa3f4ce8c0e2db1d310"
+# Torchtitan commit hash for launching on MAST
+TORCHTITAN_COMMIT_MAST="d0e25450bcac2332359b13fbda430dc701f073d4"