Major update: new models, MoE performance, LoRA, and optimizers

.codespellrc

@@ -0,0 +1,3 @@
+[codespell]
+skip = *.lock,*.json,submodules/*,.venv/*,.git,docs/node_modules/*
+ignore-words-list = dout,te,subtile,parm,mot,numer

.github/workflows/lint.yml

@@ -0,0 +1,29 @@
+name: Lint
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+
+      - uses: astral-sh/setup-uv@cec208311dfd045dd5311c1add060b2062131d57 # v8.0.0
+
+      - name: Install pre-commit
+        run: uv tool install "pre-commit~=4.5"
+
+      - name: Cache pre-commit envs
+        uses: actions/cache@5a3ec84eff668545956fd18022155c47e93e2684 # v4.2.3
+        with:
+          path: ~/.cache/pre-commit
+          key: pre-commit-${{ hashFiles('.pre-commit-config.yaml') }}
+          restore-keys: |
+            pre-commit-
+
+      - name: Run pre-commit checks
+        run: SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure

.gitmodules

@@ -0,0 +1,6 @@
+[submodule "submodules/xorl-client"]
+	path = submodules/xorl-client
+	url = https://github.com/togethercomputer/xorl-client
+[submodule "submodules/xorl-sglang"]
+	path = submodules/xorl-sglang
+	url = https://github.com/togethercomputer/xorl-sglang

.pre-commit-config.yaml

@@ -0,0 +1,34 @@
+default_stages: [pre-commit, pre-push, manual]
+exclude: ^(submodules/|\.venv/)
+
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v5.0.0
+    hooks:
+      - id: check-symlinks
+      - id: destroyed-symlinks
+      - id: trailing-whitespace
+      - id: end-of-file-fixer
+      - id: check-yaml
+        args: [--allow-multiple-documents]
+      - id: check-toml
+      - id: check-ast
+      - id: check-added-large-files
+      - id: check-merge-conflict
+      - id: check-shebang-scripts-are-executable
+      - id: detect-private-key
+      - id: debug-statements
+      - id: no-commit-to-branch
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.11.4
+    hooks:
+      - id: ruff
+        args: [--fix-only]
+      - id: ruff-format
+  - repo: https://github.com/codespell-project/codespell
+    rev: v2.4.1
+    hooks:
+      - id: codespell
+        args: ["--config", ".codespellrc"]
+        additional_dependencies:
+          - tomli

CONTRIBUTING.md

@@ -28,7 +28,28 @@ Add ring attention support for long-context training
 
 ## Code Style
 
-- Python: follow existing style, run `ruff check` and `ruff format` before pushing
+We use [pre-commit](https://pre-commit.com/) to enforce formatting and catch common issues. Set it up once:
+
+```bash
+pip install -e ".[lint]"
+pre-commit install
+```
+
+This runs automatically on every `git commit`. To check the entire codebase manually:
+
+```bash
+pre-commit run --all-files
+```
+
+The hooks include:
+- **ruff** — auto-fixes imports and lint issues
+- **ruff-format** — code formatting (line length 120)
+- **codespell** — catches typos
+- **trailing-whitespace / end-of-file-fixer** — file hygiene
+
+CI runs the same hooks, so if pre-commit passes locally, CI will too.
+
+Additional guidelines:
 - No dead code, no commented-out blocks
 - Type hints on public APIs
 

README.md

@@ -18,10 +18,18 @@
 
 XoRL is a distributed training framework designed for large language models with composable parallelism and flexible training modes.
 
+**The XoRL stack consists of three repos:**
+
+| Repo                        | Description |
+|---|---|
+| **[xorl](https://github.com/togethercomputer/xorl-internal)** | Distributed training framework — local SFT/pretraining and server-mode RL training |
+| **[xorl-client](https://github.com/togethercomputer/xorl-client)** | Lightweight Python SDK for driving the xorl training server (forward/backward, optimizer steps, checkpointing, sampling) |
+| **[xorl-sglang](https://github.com/togethercomputer/xorl-sglang)** | Fork of [SGLang](https://github.com/sgl-project/sglang) with weight-sync APIs, MoE routing export, and numerical alignment for online RL |
+
 **Two training modes:**
 
 - **Local** — `torchrun`-based training for offline SFT and pretraining
-- **Server** — REST API-driven training for online RL loops where an external orchestrator (e.g. [xorl_client](https://github.com/xorl-org/xorl_client)) controls the training loop
+- **Server** — REST API-driven training for online RL loops where [xorl-client](https://github.com/togethercomputer/xorl-client) drives the training loop and [xorl-sglang](https://github.com/togethercomputer/xorl-sglang) serves inference
 
 **Parallelism strategies** — mix and match freely:
 
@@ -40,14 +48,61 @@ XoRL is a distributed training framework designed for large language models with
 ## 🚀 Installation
 
 ```bash
-git clone --recurse-submodules git@github.com:togethercomputer/xorl.git
-cd xorl
-uv sync
+git clone --recurse-submodules git@github.com:togethercomputer/xorl-internal.git
+cd xorl-internal
 ```
 
 > Already cloned without `--recurse-submodules`? Run `git submodule update --init --recursive`
 
-See the [installation guide](https://togethercomputer.github.io/xorl/getting-started/installation/) for full setup including optional dependencies (DeepEP, Flash Attention).
+### Option A: uv (recommended)
+
+```bash
+uv sync
+source .venv/bin/activate
+```
+
+### Option B: conda
+
+```bash
+conda create -n xorl python=3.12
+conda activate xorl
+pip install -e .
+```
+
+### Submodules
+
+The repo includes two git submodules under `submodules/` (needed for server / online RL training):
+
+- **[xorl-client](https://github.com/togethercomputer/xorl-client)** — Lightweight Python SDK (no PyTorch dependency) for driving the xorl training server. Provides `ServiceClient`, `TrainingClient`, `SamplingClient`, and `RestClient` with async-first `APIFuture` semantics, automatic request ordering, and Tinker API compatibility.
+- **[xorl-sglang](https://github.com/togethercomputer/xorl-sglang)** — XoRL's fork of [SGLang](https://github.com/sgl-project/sglang) with NCCL-based weight sync endpoints, MoE routing data export (R3), and numerical alignment flags for online RL.
+
+Install individually:
+
+```bash
+pip install -e submodules/xorl-client
+pip install -e "submodules/xorl-sglang/python[all]"
+```
+
+Or use the bundled `pyproject.sglang.toml` which pins PyTorch to 2.9.1 (required by sglang) and installs everything together:
+
+**uv:**
+```bash
+cp pyproject.sglang.toml pyproject.toml
+uv sync
+source .venv/bin/activate
+```
+
+**conda:**
+```bash
+conda create -n xorl-sglang python=3.12
+conda activate xorl-sglang
+cp pyproject.sglang.toml pyproject.toml
+pip install -e .
+```
+
+> **Note:** The default `pyproject.toml` uses PyTorch 2.10.0. sglang requires PyTorch 2.9.1, so the two cannot coexist in the same environment unless you use `pyproject.sglang.toml`.
+
+See the [installation guide](https://togethercomputer.github.io/xorl-internal/getting-started/installation/) for full setup including optional dependencies (DeepEP, Flash Attention).
 
 ## ⚡ Quick Start
 
@@ -78,11 +133,14 @@ See the [quick start guide](https://togethercomputer.github.io/xorl/getting-star
 |---|---|---|
 | Qwen3 | Dense | `Qwen/Qwen3-8B`, `Qwen/Qwen3-32B`, ... |
 | Qwen3-MoE | Mixture-of-Experts | `Qwen/Qwen3-30B-A3B`, `Qwen/Qwen3-235B-A22B`, ... |
+| Qwen3.5 | Dense | `Qwen/Qwen3.5-7B`, ... |
+| Qwen3.5-MoE | Mixture-of-Experts | `Qwen/Qwen3.5-35B-A3B`, `Qwen/Qwen3.5-397B-A17B`, ... |
 
 Models are loaded directly from HuggingFace checkpoints — no preprocessing needed. See the [supported models](https://togethercomputer.github.io/xorl/models/) page for details.
 
 ---
 
 ## 🤝 Contributing
 
+
 See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup, coding conventions, and how to run tests.

docs/astro.config.mjs

@@ -37,14 +37,28 @@ export default defineConfig({
           ],
         },
         {
-          label: "Server Training",
+          label: "Server Training for RL",
           collapsed: true,
           items: [
             { label: "Overview", slug: "server-training/overview" },
-            { label: "Server Architecture", slug: "server-training/architecture" },
-            { label: "API Reference", slug: "server-training/api-reference" },
-            { label: "RL Training", slug: "server-training/rl-training" },
+            {
+              label: "Training Server (xorl)",
+              collapsed: true,
+              items: [
+                { label: "Launching & Configuration", slug: "server-training/training-server/launching" },
+                { label: "API Reference", slug: "server-training/training-server/api-reference" },
+              ],
+            },
             { label: "Inference: xorl-sglang", slug: "server-training/sglang" },
+            {
+              label: "Client SDK (xorl-client)",
+              collapsed: true,
+              items: [
+                { label: "Overview", slug: "server-training/client-sdk/overview" },
+                { label: "Training Loop Patterns", slug: "server-training/client-sdk/training-loop" },
+                { label: "Loss Functions", slug: "server-training/client-sdk/loss-functions" },
+              ],
+            },
             {
               label: "Weight Sync",
               collapsed: true,
@@ -53,6 +67,14 @@ export default defineConfig({
                 { label: "Backend: nccl_broadcast", slug: "server-training/weight-sync/nccl-broadcast" },
               ],
             },
+            {
+              label: "Examples",
+              collapsed: true,
+              items: [
+                { label: "SFT on No Robots", slug: "server-training/examples/sft-no-robots" },
+                { label: "Password Memorization", slug: "server-training/examples/password-memorization" },
+              ],
+            },
           ],
         },
         {