Adds scripts and instructions for installing Forge from source. (#98)

* commit the install stuff

* ld lib path

* fix install.sh

* remove torchao again

* a few more fixes

* install forge as editable

---------

Co-authored-by: Allen Wang <[email protected]>

Author: allenwang28

.github/workflows/unit_test.yaml

@@ -26,7 +26,7 @@ jobs:
       - name: Install pytorch
         run: python -m pip install torch==2.9.0.dev20250826 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
       - name: Install monarch
-        run: python -m pip install monarch-no-torch==0.1.0.dev20250826 --find-links assets/wheels
+        run: python -m pip install monarch-no-torch==0.1.0.dev20250826 --find-links assets/ci
       - name: Install dependencies
         run: python -m pip install --no-build-isolation -e ".[dev]"
       - name: Run unit tests with coverage

.gitignore

@@ -187,3 +187,5 @@ cover/
 
 # wandb
 wandb/
+
+assets/wheels/vllm*.whl

README.md

@@ -6,6 +6,34 @@
 
 ## Installation
 
+### Basic
+
+Forge requires the latest PyTorch nightly with Monarch, vLLM, and torchtitan. For convenience,
+we have pre-packaged these dependencies as wheels in assets/wheels.
+
+To install Forge easily:
+
+```bash
+conda create -n forge python=3.10
+./scripts/install.sh
+```
+
+You can test with
+```
+python -m apps.grpo.main
+```
+
+If you need to re-build the wheels for whatever reason, you can do so with:
+```bash
+conda create -n forge python=3.10
+./scripts/build_wheels.sh
+```
+
+Since the vLLM wheel is too large for GitHub, we uploaded it as a release:
+```
+$ gh release create v0.0.0 assets/wheels/vllm-*.whl --title "Forge Wheels v0.0.0"
+```
+
 ### Basic (Broken)
 
 ```bash

apps/sft_v2/llama3_8b.yaml

@@ -20,7 +20,7 @@ processes:
   scheduler: local # local | mast (not supported yet)
   num_hosts: 1
   num_procs: 8
-  num_gpus: 8
+  with_gpus: true
 
 optimizer:
   name: AdamW

assets/wheels/monarch_no_torch-0.1.0.dev20250826-py3-none-any.whl renamed to assets/ci/monarch_no_torch-0.1.0.dev20250826-py3-none-any.whl

@@ -13,7 +13,6 @@ dependencies = [
     # PyTorch
     "torchdata>=0.8.0",
     "torchtitan",
-    "torchao",
     # vLLM
     # TODO: pin specific vllm version
     #"vllm==0.10.0",