feat: add ROCm support for AMD training via cuda/rocm extras

.github/PULL_REQUEST_TEMPLATE.md

@@ -21,7 +21,7 @@
 
 ## Validation
 
-- [ ] `uv run pytest -q`
+- [ ] `python -m pytest -q`
 - [ ] `prek run --all-files`
 - [ ] CI checks pass on the target branch
 - [ ] Documentation updated (if applicable)

.github/workflows/tests.yml

@@ -33,7 +33,7 @@ jobs:
         run: uv python install 3.10
 
       - name: Install workspace and test dependencies
-        run: uv sync --all-packages --all-groups --all-extras
+        run: sh install.sh --all
 
       - name: Run tests
-        run: uv run pytest -q --disable-warnings --maxfail=1
+        run: .venv/bin/python -m pytest -q --disable-warnings --maxfail=1

AGENTS.md

@@ -104,20 +104,22 @@ Python 方法，由 manager 运行时契约定义（`wiki/design/manager/runtime
 
 项目使用 UV 管理依赖，Python 版本要求 3.10.\*。可用的依赖组与 extras 见 `pyproject.toml`。
 
-完整安装（含全部 package、依赖组与 extras）：
+完整开发环境（全部 package、依赖组与 extras；`--gpu rocm` 切换 AMD wheel）：
 
 ```bash
-uv sync --all-packages --all-groups --all-extras
+sh install.sh --all
 ```
 
-按框架安装：
+运行环境（无开发工具链；GPU 厂商自动探测，`--rslrl` 换后端）：
 
 ```bash
-uv sync --all-packages --extra skrl-jax   # SKRL JAX backend
-uv sync --all-packages --extra skrl-torch # SKRL PyTorch backend
-uv sync --all-packages --extra rslrl      # RSLRL（PyTorch）
+sh install.sh
 ```
 
+避免裸 `uv run` / `uv sync`：`cuda`/`rocm` extras 声明为互斥，裸命令会解析到默认 fork（PyPI torch）
+并重装环境。运行命令先 `source .venv/bin/activate`，之后直接 `python` / `pytest`（下文示例即此写法），
+或对单条命令使用 `uv run --no-sync`。依赖变更（`pyproject.toml` / `uv.lock`）后重新执行对应 install 命令。
+
 ## 常用命令
 
 train/play/view CLI 由 [Hydra](https://hydra.cc/) 驱动，参数使用 `key=value` 语法（不是 `--flag`）。可用选项见
@@ -128,72 +130,72 @@ train/play/view CLI 由 [Hydra](https://hydra.cc/) 驱动，参数使用 `key=va
 SKRL（默认）：
 
 ```bash
-uv run scripts/train.py task=cartpole/skrl.ppo
+python scripts/train.py task=cartpole/skrl.ppo
 ```
 
 RSLRL：
 
 ```bash
-uv run scripts/train.py task=cartpole/rslrl.ppo
+python scripts/train.py task=cartpole/rslrl.ppo
 ```
 
 直接覆写框架运行时设置和类型化 RL 参数：
 
 ```bash
-uv run scripts/train.py task=cartpole/skrl.ppo num_envs=64 algo.agent.learning_rate=1e-3
-uv run scripts/train.py task=cartpole/skrl.ppo logging.interval=20 checkpoint.interval=100
+python scripts/train.py task=cartpole/skrl.ppo num_envs=64 algo.agent.learning_rate=1e-3
+python scripts/train.py task=cartpole/skrl.ppo logging.interval=20 checkpoint.interval=100
 ```
 
 ### 环境可视化
 
 不训练只查看环境：
 
 ```bash
-uv run scripts/view.py env=cartpole
+python scripts/view.py env=cartpole
 ```
 
 查看内置机器人（不创建 RL 环境）：
 
 ```bash
-uv run scripts/view.py robot=g1-29dof
+python scripts/view.py robot=g1-29dof
 ```
 
 ### 评估
 
 ```bash
-uv run scripts/play.py env=cartpole
+python scripts/play.py env=cartpole
 ```
 
 指定 policy 文件：
 
 ```bash
-uv run scripts/play.py env=cartpole policy=<path/to/best.[pickle/pt]>
+python scripts/play.py env=cartpole policy=<path/to/best.[pickle/pt]>
 ```
 
 ### ONNX 导出
 
 ```bash
-uv run scripts/export_onnx.py run_dir=<run-dir> output=/tmp/policy.onnx
+python scripts/export_onnx.py run_dir=<run-dir> output=/tmp/policy.onnx
 ```
 
 ### 渲染
 
 训练时开启可视化：
 
 ```bash
-uv run scripts/train.py task=cartpole/skrl.ppo render=true
+python scripts/train.py task=cartpole/skrl.ppo render=true
 ```
 
 ### TensorBoard
 
 ```bash
-uv run tensorboard --logdir runs/{env-name}
+tensorboard --logdir runs/{env-name}
 ```
 
 ### 测试
 
 ```bash
-uv run pytest
+python -m pytest
 ```
 
 ## 架构要点

CONTRIBUTING.md

@@ -26,17 +26,16 @@ git lfs pull
 Create the complete development environment from the repository root:
 
 ```bash
-uv sync --all-packages --all-groups --all-extras
+sh install.sh --all       # --gpu rocm switches to the AMD wheels
 ```
 
-Optional training backends can be installed separately when a full environment
-is unnecessary:
-
-```bash
-uv sync --all-packages --extra skrl-torch
-uv sync --all-packages --extra skrl-jax   # Linux only
-uv sync --all-packages --extra rslrl
-```
+A runtime-only environment without the dev toolchain is available via `sh install.sh`: it
+auto-detects the GPU vendor (NVIDIA/AMD) for the torch wheels and `--rslrl`
+selects the RSLRL backend. On Windows, run `install.ps1` in PowerShell. Avoid bare
+`uv sync` / `uv run`: the `cuda`/`rocm` extras are declared conflicting, so a bare
+command resolves the default fork (PyPI torch) and reinstalls the environment. Run
+commands from the activated environment (`source .venv/bin/activate`); the examples
+below assume it is active, or pass `--no-sync` to one-off `uv run` calls.
 
 The workspace contains nine packages. Package-local changes should use the
 smallest required extra; changes involving the simulator, built-in assets, or
@@ -168,15 +167,15 @@ branch unless the change specifically requires them.
 Run the full test suite before opening a pull request:
 
 ```bash
-uv run pytest -q
+python -m pytest -q
 ```
 
 For an iteration on one package, run its tests directly, then run the full
 suite before requesting review:
 
 ```bash
-uv run pytest motrix_env_core/tests -q
-uv run pytest motrix_deploy/tests motrix_deploy_mujoco/tests -q
+python -m pytest motrix_env_core/tests -q
+python -m pytest motrix_deploy/tests motrix_deploy_mujoco/tests -q
 ```
 
 Run the repository's formatting, license-header, and lint hooks:
@@ -190,17 +189,16 @@ prek run --all-files
 The hooks run Copywrite, Ruff, and dprint. For individual checks, use:
 
 ```bash
-uv run ruff check .
-uv run ruff format --check .
-uv run mypy
+ruff check .
+ruff format --check .
+mypy
 ```
 
-When changing documentation, install the `docs` extra and build with warnings
-treated as errors:
+When changing documentation, build with warnings treated as errors (`sh install.sh --all`
+or a runtime install with `--docs` provides the required docs toolchain):
 
 ```bash
-uv sync --all-packages --extra docs
-uv run sphinx-build -W -b html docs/source docs/build/html
+sphinx-build -W -b html docs/source docs/build/html
 ```
 
 The public CI runs the complete workspace test command on Linux for both

README.md

@@ -37,7 +37,7 @@ _Microduck locomotion policies trained with MotrixLab, rendered in MotrixRender
 <div align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" srcset="docs/source/_static/images/architecture-dark.svg">
-    <img src="docs/source/_static/images/architecture-light.svg" alt="MotrixLab architecture: define an environment once, train it with SKRL, RSL-RL or FastSAC on thousands of parallel MotrixSim environments, then deploy the same policy artifact to MuJoCo or Unitree hardware" width="720">
+    <img src="docs/source/_static/images/architecture-light.svg" alt="MotrixLab architecture: define an environment once, train it with SKRL, RSL-RL or FastSAC on thousands of parallel MotrixSim environments running on NVIDIA CUDA or AMD ROCm GPUs, then deploy the same policy artifact to MuJoCo or Unitree hardware" width="100%">
   </picture>
 </div>
 
@@ -60,6 +60,7 @@ _Microduck locomotion policies trained with MotrixLab, rendered in MotrixRender
 | [uv](https://docs.astral.sh/uv/) | Python project and dependency manager — [installation guide](https://docs.astral.sh/uv/getting-started/installation/) |
 | [Git LFS](https://git-lfs.com) | Robot meshes, motion data, and videos are tracked by LFS |
 | OS | Linux x86_64 or Windows x86_64; the JAX training backend is Linux-only |
+| GPU | NVIDIA (CUDA) or AMD (ROCm) — the matching wheels are selected automatically by `sh install.sh` |
 
 ### 1. Clone the repository
 
@@ -71,16 +72,29 @@ git lfs pull
 
 ### 2. Install dependencies
 
+Linux:
+
 ```bash
-uv sync --all-packages
+sh install.sh
 ```
 
-This installs all workspace packages together with **PyTorch**, the default training backend used by the built-in FastSAC. Third-party frameworks such as SKRL and RSLRL are optional extras.
+Windows (PowerShell):
+
+```powershell
+.\install.ps1
+# if blocked by the execution policy:
+powershell -ExecutionPolicy Bypass -File install.ps1
+```
+
+This auto-detects your GPU vendor (NVIDIA → CUDA, AMD → ROCm) and installs all workspace packages with the matching PyTorch wheels. Use `--gpu cuda|rocm` to override detection and `--skrl-jax` / `--rslrl` to add training backends — see `sh install.sh --help`.
 
 ### 3. Train your first policy
 
+Activate the installed environment (Windows PowerShell: `.venv\Scripts\Activate.ps1`):
+
 ```bash
-uv run scripts/train.py task=microduck-walk-flat/motrix.fastsac play=true
+source .venv/bin/activate
+python scripts/train.py task=microduck-walk-flat/motrix.fastsac play=true
 ```
 
 While training, the built-in dashboard shows live run progress, episode statistics, throughput, rewards, and system health:
@@ -92,7 +106,7 @@ While training, the built-in dashboard shows live run progress, episode statisti
 Training runs thousands of parallel environment instances; when it finishes, the trained policy is loaded and played in the viewer automatically. Checkpoints and TensorBoard logs are saved under `runs/microduck-walk-flat/`; watch the curves with:
 
 ```bash
-uv run tensorboard --logdir runs/microduck-walk-flat
+tensorboard --logdir runs/microduck-walk-flat
 ```
 
 Training finishes in minutes: mean return and episode length typically converge after about 4,000 iterations:
@@ -106,7 +120,7 @@ Training finishes in minutes: mean return and episode length typically converge
 Replay the latest trained policy without retraining (for example, after stopping training early with Ctrl+C):
 
 ```bash
-uv run scripts/play.py env=microduck-walk-flat
+python scripts/play.py env=microduck-walk-flat
 ```
 
 A trained microduck policy replayed in the viewer:
@@ -124,7 +138,7 @@ MotrixLab ships 50+ built-in simulation environments spanning basic control, qua
 | <img src="docs/source/_static/images/poster/g1-wbt-dance.jpg" alt="g1-wbt-dance" width="240"> | Whole-body tracking (WBT) | `g1-wbt-dance` · `k1-wbt-freekick` · `g1-29dof-wbt-largebox` |
 
 ```bash
-uv run scripts/view.py env=go2-walk-rough
+python scripts/view.py env=go2-walk-rough
 ```
 
 See the [full environment gallery](https://motrixlab.readthedocs.io/en/latest/user_guide/envs/index.html) for all registered environments and their supported training algorithms.
@@ -144,7 +158,7 @@ Seven reusable robot models are registered out of the box and can be combined in
 | <img src="docs/source/_static/images/robots/microduck.png" alt="microduck" width="180"> | `microduck` | Humanoid | 14 |
 
 ```bash
-uv run scripts/view.py robot=go2
+python scripts/view.py robot=go2
 ```
 
 See [Supported Robots](https://motrixlab.readthedocs.io/en/latest/user_guide/robots.html) for configuration details and how to add your own model.

README.zh-CN.md

@@ -37,7 +37,7 @@ _使用 MotrixLab 训练的 microduck 行走策略，由 MotrixRender 实时渲
 <div align="center">
   <picture>
     <source media="(prefers-color-scheme: dark)" srcset="docs/source/_static/images/architecture-dark.svg">
-    <img src="docs/source/_static/images/architecture-light.svg" alt="MotrixLab 架构：环境只需定义一次，即可用 SKRL、RSL-RL 或 FastSAC 在数千个并行 MotrixSim 环境上训练，同一策略产物可部署到 MuJoCo 或 Unitree 硬件" width="720">
+    <img src="docs/source/_static/images/architecture-light.svg" alt="MotrixLab 架构：环境只需定义一次，即可用 SKRL、RSL-RL 或 FastSAC 在数千个并行 MotrixSim 环境上训练（支持 NVIDIA CUDA 与 AMD ROCm GPU），同一策略产物可部署到 MuJoCo 或 Unitree 硬件" width="100%">
   </picture>
 </div>
 
@@ -60,6 +60,7 @@ _使用 MotrixLab 训练的 microduck 行走策略，由 MotrixRender 实时渲
 | [uv](https://docs.astral.sh/uv/) | Python 项目与依赖管理工具 — [安装指南](https://docs.astral.sh/uv/getting-started/installation/) |
 | [Git LFS](https://git-lfs.com) | 机器人网格、运动数据与视频由 LFS 管理 |
 | 操作系统 | Linux x86_64 或 Windows x86_64；JAX 训练后端仅支持 Linux |
+| GPU | NVIDIA（CUDA）或 AMD（ROCm）——`sh install.sh` 自动选择对应 wheel |
 
 ### 1. 克隆仓库
 
@@ -71,16 +72,29 @@ git lfs pull
 
 ### 2. 安装依赖
 
+Linux：
+
 ```bash
-uv sync --all-packages
+sh install.sh
 ```
 
-该命令会安装全部 workspace package，以及内置 FastSAC 所需的默认训练后端 **PyTorch**。SKRL、RSLRL 等第三方训练框架为可选 extras。
+Windows（PowerShell）：
+
+```powershell
+.\install.ps1
+# 若提示执行策略受限：
+powershell -ExecutionPolicy Bypass -File install.ps1
+```
+
+脚本会自动探测 GPU 厂商（NVIDIA → CUDA，AMD → ROCm），安装全部 workspace package 及对应的 PyTorch wheel。可用 `--gpu cuda|rocm` 显式指定 GPU，`--skrl-jax` / `--rslrl` 追加训练后端——详见 `sh install.sh --help`。
 
 ### 3. 训练第一个策略
 
+先激活安装好的环境（Windows PowerShell：`.venv\Scripts\Activate.ps1`）：
+
 ```bash
-uv run scripts/train.py task=microduck-walk-flat/motrix.fastsac play=true
+source .venv/bin/activate
+python scripts/train.py task=microduck-walk-flat/motrix.fastsac play=true
 ```
 
 训练过程中，内置面板会实时显示运行进度、回合统计、吞吐、奖励与系统健康状态：
@@ -92,7 +106,7 @@ uv run scripts/train.py task=microduck-walk-flat/motrix.fastsac play=true
 训练会启动数千个并行环境实例；训练结束后会自动加载策略并在查看器中回放。checkpoint 与 TensorBoard 日志保存在 `runs/microduck-walk-flat/` 目录下，通过以下命令查看训练曲线：
 
 ```bash
-uv run tensorboard --logdir runs/microduck-walk-flat
+tensorboard --logdir runs/microduck-walk-flat
 ```
 
 microduck 的训练数分钟内即可完成：平均回报与回合长度通常在约 4,000 次迭代后收敛：
@@ -106,7 +120,7 @@ microduck 的训练数分钟内即可完成：平均回报与回合长度通常
 无需重新训练即可回放最近一次训练得到的策略（例如提前 Ctrl+C 中断训练之后）：
 
 ```bash
-uv run scripts/play.py env=microduck-walk-flat
+python scripts/play.py env=microduck-walk-flat
 ```
 
 训练好的 microduck 策略在查看器中的回放效果：
@@ -124,7 +138,7 @@ MotrixLab 内置 50+ 个仿真环境，覆盖基础控制、四足、人形、
 | <img src="docs/source/_static/images/poster/g1-wbt-dance.jpg" alt="g1-wbt-dance" width="240"> | 全身动作跟踪（WBT） | `g1-wbt-dance` · `k1-wbt-freekick` · `g1-29dof-wbt-largebox` |
 
 ```bash
-uv run scripts/view.py env=go2-walk-rough
+python scripts/view.py env=go2-walk-rough
 ```
 
 完整环境列表与各环境支持的训练算法见[环境总览](https://motrixlab.readthedocs.io/zh-cn/latest/user_guide/envs/index.html)。
@@ -144,7 +158,7 @@ uv run scripts/view.py env=go2-walk-rough
 | <img src="docs/source/_static/images/robots/microduck.png" alt="microduck" width="180"> | `microduck` | 人形机器人 | 14 |
 
 ```bash
-uv run scripts/view.py robot=go2
+python scripts/view.py robot=go2
 ```
 
 机器人配置细节与自定义新模型的方法见[支持的机器人](https://motrixlab.readthedocs.io/zh-cn/latest/user_guide/robots.html)。

configs/task/microduck-walk-flat/motrix.fastsac.yaml

@@ -16,4 +16,8 @@ algo:
   agent:
     alpha_init: 0.01
     target_entropy_ratio: -0.1
+    batch_size: 2048
     num_updates: 4
+  trainer:
+    async_options:
+      utd_mode: strict