fix: support Windows

README.md

@@ -80,8 +80,13 @@ Start with the `Quick Demo` below to run the primary training command. The recom
 </table>
 
 ```bash
-# 0. If uv is not installed
+# 0. Install uv if needed
+# Linux / macOS:
 curl -LsSf https://astral.sh/uv/install.sh | sh
+#
+# Windows:
+# powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
+# choco install make -y
 
 # 1. Clone the repository
 git clone https://github.com/unilabsim/UniLab.git
@@ -90,7 +95,7 @@ cd UniLab
 # 2. Install dependencies
 # Pick the setup command for your platform.
 
-# Linux CUDA or macOS
+# Linux CUDA, macOS, or Windows
 make setup
 
 # Linux AMD / ROCm
@@ -101,7 +106,7 @@ make setup
 
 # Without shell completion setup:
 # uv sync --extra mujoco --extra motrix
-# If `make` is not installed:
+# If `make` is not installed or unavailable:
 # uv sync --extra mujoco --extra motrix && uv run --no-sync unilab-complete install
 
 # 3. Pre-trained checkpoint playback (downloads from Hugging Face on first run)

README_zh.md

@@ -81,7 +81,12 @@
 
 ```bash
 # 0. 如果还没有安装 uv
+# Linux / macOS：
 curl -LsSf https://astral.sh/uv/install.sh | sh
+#
+# Windows：
+# powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
+# choco install make -y
 
 # 1. 克隆仓库
 git clone https://github.com/unilabsim/UniLab.git
@@ -90,7 +95,7 @@ cd UniLab
 # 2. 安装依赖
 # 请按你的平台选择对应的安装命令。
 
-# Linux CUDA 或 macOS
+# Linux CUDA、macOS 或 Windows
 make setup
 
 # Linux AMD / ROCm
@@ -101,7 +106,7 @@ make setup
 
 # 不使用 shell completion 设置时：
 # uv sync --extra mujoco --extra motrix
-# 如果没有安装 `make`：
+# 如果没有安装或无法使用 `make`：
 # uv sync --extra mujoco --extra motrix && uv run --no-sync unilab-complete install
 
 # 3. 预训练 checkpoint 回放（首次运行会从 Hugging Face 下载）

pyproject.toml

@@ -34,6 +34,7 @@ dependencies = [
     "onnxruntime>=1.20 ; python_version >= '3.11'",
     "huggingface_hub>=0.25",
     "ninja ; sys_platform == 'linux'",
+    "imageio-ffmpeg>=0.6.0",
 ]
 
 [project.scripts]
@@ -62,7 +63,7 @@ explicit = true
 
 [tool.uv.sources]
 torch = [
-    { index = "pytorch-cu128", marker = "sys_platform=='linux'" },
+    { index = "pytorch-cu128", marker = "sys_platform=='linux' or sys_platform=='win32'" },
 ]
 
 [tool.uv]

scripts/train_appo.py

@@ -351,7 +351,7 @@ def forward(self, obs: torch.Tensor) -> torch.Tensor:
             on_plan=log_playback_plan,
         )
     if play_video_path is not None:
-        print(f"Saving video to {play_video_path} with mediapy...")
+        print(f"Saving video to {play_video_path} ...")
     print("Done.")
     return play_video_path
 

scripts/train_mlx_ppo.py

@@ -325,28 +325,23 @@ def _play_step(current_obs):
         return mx.nan_to_num(raw_obs, nan=0.0, posinf=0.0, neginf=0.0)
 
     output_dir = run_dir if run_dir is not None else task_log_root
-    try:
-        play_video_path = env.run_playback_mode(
-            play_render_mode=getattr(cfg.training, "play_render_mode", "auto"),
-            play_steps=getattr(cfg.training, "play_steps", None),
-            output_video=output_dir / "play_video.mp4",
-            initialize=lambda: obs,
-            step=_play_step,
-            camera_kwargs={
-                "cam_distance": getattr(cfg.training, "cam_distance", 2.0),
-                "cam_elevation": getattr(cfg.training, "cam_elevation", -20.0),
-                "cam_azimuth": getattr(cfg.training, "cam_azimuth", 90.0),
-                "cam_lookat": getattr(cfg.training, "cam_lookat", None),
-                "cam_tracking": getattr(cfg.training, "cam_tracking", False),
-                "cam_tracking_env_idx": getattr(cfg.training, "cam_tracking_env_idx", 0),
-                "cam_tracking_extra_envs": getattr(cfg.training, "cam_tracking_extra_envs", 2),
-            },
-            on_plan=lambda plan: log_playback_plan(plan, prefix="[MLX PPO] "),
-        )
-    except ImportError:
-        print("mediapy is required for play video export. Install with `pip install mediapy`.")
-        env.close()
-        return None
+    play_video_path = env.run_playback_mode(
+        play_render_mode=getattr(cfg.training, "play_render_mode", "auto"),
+        play_steps=getattr(cfg.training, "play_steps", None),
+        output_video=output_dir / "play_video.mp4",
+        initialize=lambda: obs,
+        step=_play_step,
+        camera_kwargs={
+            "cam_distance": getattr(cfg.training, "cam_distance", 2.0),
+            "cam_elevation": getattr(cfg.training, "cam_elevation", -20.0),
+            "cam_azimuth": getattr(cfg.training, "cam_azimuth", 90.0),
+            "cam_lookat": getattr(cfg.training, "cam_lookat", None),
+            "cam_tracking": getattr(cfg.training, "cam_tracking", False),
+            "cam_tracking_env_idx": getattr(cfg.training, "cam_tracking_env_idx", 0),
+            "cam_tracking_extra_envs": getattr(cfg.training, "cam_tracking_extra_envs", 2),
+        },
+        on_plan=lambda plan: log_playback_plan(plan, prefix="[MLX PPO] "),
+    )
     if play_video_path is not None:
         print(f"[MLX PPO] Play video saved: {play_video_path}")
     else:

src/unilab/algos/torch/appo/learner.py

@@ -21,6 +21,8 @@
 from rsl_rl.utils import resolve_optimizer
 from tensordict import TensorDict
 
+from unilab.algos.torch.common.compile import get_torch_compile_for_cuda
+
 _LOG_2_PI = math.log(2.0 * math.pi)
 _NORMAL_ENTROPY_OFFSET = 0.5 * (1.0 + _LOG_2_PI)
 
@@ -209,7 +211,7 @@ def __init__(
         self._update_counter = 0
         self.last_update_metrics: dict[str, float] = {}
         self.enable_compile = (
-            bool(enable_compile) and self._device_type == "cuda" and hasattr(torch, "compile")
+            bool(enable_compile) and get_torch_compile_for_cuda(device, warn=True) is not None
         )
 
         # Optimizer
@@ -221,8 +223,8 @@ def __init__(
             self._compile_training_methods()
 
     def _compile_training_methods(self) -> None:
-        compile_fn = getattr(torch, "compile", None)
-        if compile_fn is None or self._device_type != "cuda":
+        compile_fn = get_torch_compile_for_cuda(self._device_type, warn=True)
+        if compile_fn is None:
             return
 
         self._minibatch_loss_fn = compile_fn(

src/unilab/algos/torch/common/compile.py

@@ -0,0 +1,48 @@
+from __future__ import annotations
+
+import importlib.util
+import sys
+from collections.abc import Callable
+from typing import Any, cast
+
+import torch
+
+_WARNED_REASONS: set[str] = set()
+
+
+def _warn_once(reason: str) -> None:
+    if reason in _WARNED_REASONS:
+        return
+    _WARNED_REASONS.add(reason)
+    message = f"WARNING: torch.compile is unavailable for CUDA; using eager mode ({reason})."
+    try:
+        from rich.console import Console
+
+        Console(stderr=True).print(message, style="yellow")
+    except Exception:  # pragma: no cover - best-effort diagnostic only
+        print(message, file=sys.stderr)
+
+
+def get_torch_compile_for_cuda(
+    device: torch.device | str, *, warn: bool = False
+) -> Callable[..., Any] | None:
+    """Return ``torch.compile`` when CUDA Inductor dependencies are available."""
+    compile_fn = getattr(torch, "compile", None)
+    if torch.device(device).type != "cuda":
+        return None
+    if compile_fn is None:
+        if warn:
+            _warn_once("torch.compile is not present in this PyTorch build")
+        return None
+    if (
+        getattr(compile_fn, "__module__", "") == "torch"
+        and importlib.util.find_spec("triton") is None
+    ):
+        if warn:
+            _warn_once(
+                "Triton is not installed; this environment cannot use CUDA Inductor. "
+                "PyTorch's Windows torch.compile documentation currently covers "
+                "CPU/XPU Inductor, not the CUDA/Triton path"
+            )
+        return None
+    return cast(Callable[..., Any], compile_fn)

src/unilab/algos/torch/fast_sac/learner.py

@@ -19,6 +19,7 @@
 import torch.nn.functional as F
 import torch.optim as optim
 
+from unilab.algos.torch.common.compile import get_torch_compile_for_cuda
 from unilab.base.augmentation import SymmetryAugmentation
 
 # ---------------------------------------------------------------------------
@@ -406,7 +407,7 @@ def __init__(
         self.use_autotune = use_autotune
         self.use_amp = bool(use_amp) and self._device_type in ("cuda", "xpu")
         self.use_compile = (
-            bool(use_compile) and self._device_type == "cuda" and hasattr(torch, "compile")
+            bool(use_compile) and get_torch_compile_for_cuda(self.device, warn=True) is not None
         )
         self.amp_dtype = amp_dtype
         self._amp_dtype = self._resolve_amp_dtype(amp_dtype, self._device_type)
@@ -520,8 +521,8 @@ def _should_use_grad_scaler(
         return bool(use_amp) and device_type == "cuda" and amp_dtype == torch.float16
 
     def _compile_training_methods(self) -> None:
-        compile_fn = getattr(torch, "compile", None)
-        if compile_fn is None or torch.device(self.device).type != "cuda":
+        compile_fn = get_torch_compile_for_cuda(self.device, warn=True)
+        if compile_fn is None:
             return
 
         compile_kwargs = {"options": {"triton.cudagraphs": False}}

src/unilab/algos/torch/flash_sac/learner.py

@@ -10,6 +10,7 @@
 import torch.nn as nn
 import torch.optim as optim
 
+from unilab.algos.torch.common.compile import get_torch_compile_for_cuda
 from unilab.algos.torch.common.normalization import EmpiricalNormalization
 from unilab.algos.torch.flash_sac.network import (
     FlashSACActor,
@@ -182,7 +183,7 @@ def __init__(
         self.amp_dtype = amp_dtype
         self._amp_dtype = self._resolve_amp_dtype(amp_dtype, self.device.type)
         self.use_compile = bool(
-            use_compile and hasattr(torch, "compile") and self.device.type == "cuda"
+            use_compile and get_torch_compile_for_cuda(self.device, warn=True) is not None
         )
 
         self.actor = FlashSACActor(
@@ -256,8 +257,8 @@ def __init__(
             self._compile_training_methods()
 
     def _compile_training_methods(self) -> None:
-        compile_fn = getattr(torch, "compile", None)
-        if compile_fn is None or self.device.type != "cuda":
+        compile_fn = get_torch_compile_for_cuda(self.device, warn=True)
+        if compile_fn is None:
             return
 
         compile_kwargs = {"options": {"triton.cudagraphs": False}}

src/unilab/algos/torch/hora/appo.py

@@ -240,7 +240,7 @@ def step_play_obs(obs_np: np.ndarray) -> np.ndarray:
             ),
         )
     if play_video_path is not None:
-        print(f"Saving video to {play_video_path} with mediapy...")
+        print(f"Saving video to {play_video_path} ...")
     print("Done.")
     return play_video_path
 

src/unilab/algos/torch/rsl_rl_ppo.py

@@ -7,6 +7,8 @@
 from rsl_rl.algorithms import PPO
 from tensordict import TensorDict
 
+from unilab.algos.torch.common.compile import get_torch_compile_for_cuda
+
 _LOG_2_PI = math.log(2.0 * math.pi)
 _NORMAL_ENTROPY_OFFSET = 0.5 * (1.0 + _LOG_2_PI)
 
@@ -24,17 +26,15 @@ def __init__(
     ) -> None:
         super().__init__(*args, **kwargs)
         self.enable_compile = (
-            bool(enable_compile)
-            and torch.device(self.device).type == "cuda"
-            and hasattr(torch, "compile")
+            bool(enable_compile) and get_torch_compile_for_cuda(self.device, warn=True) is not None
         )
         self._minibatch_loss_fn = self._minibatch_loss_tensors
         if self.enable_compile:
             self._compile_training_methods()
 
     def _compile_training_methods(self) -> None:
-        compile_fn = getattr(torch, "compile", None)
-        if compile_fn is None or torch.device(self.device).type != "cuda":
+        compile_fn = get_torch_compile_for_cuda(self.device, warn=True)
+        if compile_fn is None:
             return
 
         self._minibatch_loss_fn = compile_fn(

src/unilab/assets/hub.py

@@ -12,7 +12,9 @@
 from __future__ import annotations
 
 import logging
+import ntpath
 import os
+import posixpath
 from collections.abc import Sequence
 from pathlib import Path
 
@@ -89,22 +91,23 @@ def resolve_checkpoint_file(
 def _resolve_single(path_str: str, *, repo_id: str = _HF_MOTIONS_REPO_ID) -> str:
     """Resolve one asset file path, downloading if absent."""
     path = Path(path_str)
+    is_absolute_input = path.is_absolute() or ntpath.isabs(path_str) or posixpath.isabs(path_str)
 
     # Already exists locally — fast path.
     if path.exists():
         return str(path)
 
     # Try interpreting as ASSETS_ROOT_PATH-relative.
-    if not path.is_absolute():
+    if not is_absolute_input:
         local = ASSETS_ROOT_PATH / path
         if local.exists():
             return str(local)
-        relative = path_str
+        relative = _hf_relative_path(path_str)
     else:
         # Extract the portion relative to ASSETS_ROOT_PATH so we can
         # request the matching file from the HF repo.
         try:
-            relative = str(path.relative_to(ASSETS_ROOT_PATH))
+            relative = path.relative_to(ASSETS_ROOT_PATH).as_posix()
         except ValueError:
             raise FileNotFoundError(
                 f"Asset file not found and path is not under "
@@ -114,6 +117,11 @@ def _resolve_single(path_str: str, *, repo_id: str = _HF_MOTIONS_REPO_ID) -> str
     return _download_from_hf(relative, repo_id=repo_id)
 
 
+def _hf_relative_path(path_str: str) -> str:
+    """Return a repo-relative HF path with POSIX separators."""
+    return path_str.replace("\\", "/")
+
+
 def _hf_download(hf_hub_download, relative_path: str, *, repo_id: str) -> str:  # type: ignore[no-untyped-def]
     """Call ``hf_hub_download`` with the standard arguments."""
     return str(
@@ -206,6 +214,7 @@ def _resolve_snapshot_dir(directory: str, *, repo_id: str, marker: str) -> Path:
     Returns:
         Absolute ``Path`` to the resolved directory.
     """
+    hf_directory = _hf_relative_path(directory)
     target = ASSETS_ROOT_PATH / directory
     if (target / marker).is_file():
         return target
@@ -221,10 +230,10 @@ def _resolve_snapshot_dir(directory: str, *, repo_id: str, marker: str) -> Path:
             "  uv pip install huggingface_hub"
         ) from None
 
-    logger.info("Downloading %s from HF repo %s ...", directory, repo_id)
+    logger.info("Downloading %s from HF repo %s ...", hf_directory, repo_id)
 
     try:
-        _snapshot_download(snapshot_download, directory, repo_id=repo_id)
+        _snapshot_download(snapshot_download, hf_directory, repo_id=repo_id)
     except Exception:
         current_endpoint = os.environ.get("HF_ENDPOINT", "")
         if current_endpoint and current_endpoint != _HF_OFFICIAL_ENDPOINT:
@@ -236,7 +245,7 @@ def _resolve_snapshot_dir(directory: str, *, repo_id: str, marker: str) -> Path:
             original = os.environ["HF_ENDPOINT"]
             os.environ["HF_ENDPOINT"] = _HF_OFFICIAL_ENDPOINT
             try:
-                _snapshot_download(snapshot_download, directory, repo_id=repo_id)
+                _snapshot_download(snapshot_download, hf_directory, repo_id=repo_id)
             finally:
                 os.environ["HF_ENDPOINT"] = original
         else: