fix(runtime): require explicit host-bridge manager devices

docs/sphinx/source/adr/ADR-0012-sole-tensor-manager-and-scoped-backends.md

@@ -61,14 +61,16 @@ owns its explicit packed conversion boundary. MuJoCo remains a production backen
 canonical in-process host-bridge implementation; it is not a test-only fallback.
 
 `env.tensor_runtime` and `env.tensor_runtime_device` are removed. Tensor execution is an invariant
-of the Manager runtime. Placement is derived from declared backend data plane and rank topology.
-No compatibility alias, dual-mode field, or hidden fallback is retained.
-
-For a `HOST_BRIDGE` backend, CPU-authoritative physics does not imply CPU Manager tensors.
-When PyTorch exposes an available GPU and the backend accepts its current `cuda` ordinal,
-the Manager uses that GPU, including ROCm PyTorch's `cuda` namespace. CPU-only capabilities
-or hosts without an available GPU retain CPU placement. `DEVICE_RESIDENT` adapters keep
-their own CUDA-only platform requirements; accepting HIP Torch buffers on a host bridge
+of the Manager runtime. Placement is derived from the declared backend data plane and, for a
+`HOST_BRIDGE`, the process's explicit carrier request. No NumPy/dual-runtime compatibility alias
+or hidden fallback is retained.
+
+For a `HOST_BRIDGE` backend, CPU-authoritative physics keeps the default Manager/TorchEnv
+carriers on CPU. A training process may explicitly request `cuda`/`cuda:<ordinal>` through
+`manager_torch_device` (routed from its resolved learner device), including ROCm PyTorch's
+`cuda` namespace; UniLab validates the request against the backend's declared Torch devices.
+GPU visibility alone is not a placement request. `DEVICE_RESIDENT` adapters keep their own
+CUDA-only placement and platform requirements; accepting HIP Torch buffers on a host bridge
 does not make those physics engines ROCm-compatible.
 
 ### Manager-owned RNG

docs/sphinx/source/en/1-getting_started/2-installation.md

@@ -243,10 +243,13 @@ ROCm notes:
   profile, use `uv run --no-sync` (or `UV_NO_SYNC=1 make ...`) for validation;
   automatic synchronization would reinstall that profile's CUDA wheels.
 - With ROCm PyTorch and an available GPU, a `HOST_BRIDGE` backend accepting
-  `cuda` buffers runs the Manager/TorchEnv tensors on the current GPU. MuJoCo
-  physics remains on CPU; packed transfers connect it to GPU observations,
-  actions, rewards, and resets. CUDA-only physics backends remain unsupported
-  on ROCm. See {doc}`/adr/ADR-0012-sole-tensor-manager-and-scoped-backends`.
+  `cuda` buffers can use the current GPU for Manager/TorchEnv tensors when the
+  training process explicitly requests that learner device (which routes
+  `manager_torch_device`). The direct environment default remains CPU. MuJoCo
+  physics stays on CPU; packed transfers connect it to explicitly requested GPU
+  observations, actions, rewards, and resets. CUDA-only physics backends remain
+  unsupported on ROCm. See
+  {doc}`/adr/ADR-0012-sole-tensor-manager-and-scoped-backends`.
 - When installing from PyPI instead of a source checkout, `make sync-rocm` does
   not apply. Install the torch build validated by the repository from the
   PyTorch ROCm index first, then `unilab`. The published dependency range is

docs/sphinx/source/en/5-reference/5-support_matrix.md

@@ -74,12 +74,12 @@ This table is derived from UniSim's SDK-free public static inventory. It describ
 
 | Backend | Execution / process / data plane | Torch devices | CUDA runtime | Linux+CUDA | macOS | ROCm | Worker | Reset randomization | Fixed variants | Host callbacks | Packed bridge |
 |---|---|---|---|---|---|---|---|---|---|---|---|
-| `mujoco` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | unknown | unknown | Unsupported | Exact |
-| `motrix` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
-| `drake` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
+| `mujoco` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | unknown | unknown | Unsupported | Exact |
+| `motrix` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
+| `drake` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
 | `mjwarp` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
 | `newton` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
-| `superdex` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
+| `superdex` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Exact |
 | `genesis` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | Unsupported | Unsupported | Unsupported | Unsupported |
 
 ### Entrypoint x Task Owner
@@ -125,7 +125,8 @@ configuration. If `CUDA_VISIBLE_DEVICES` is set, backend ordinals address that
 remapped namespace, not host-global physical indices.
 
 On macOS and ROCm, use a CPU-authoritative host-bridge backend. On ROCm,
-Manager/TorchEnv uses the current GPU when ROCm PyTorch is available and the
-backend accepts that `cuda` device; CPU physics does not require CPU Manager
-tensors. GPU Torch buffers on a host bridge do not imply GPU physics or a
-device-resident backend lifecycle.
+Manager/TorchEnv uses the current GPU only when the training process explicitly
+requests that learner device and the backend accepts the routed `cuda`
+`manager_torch_device`; direct environment construction remains CPU. GPU Torch
+buffers on a host bridge do not imply GPU physics or a device-resident backend
+lifecycle.

docs/sphinx/source/zh_CN/1-getting_started/2-installation.md

@@ -218,10 +218,12 @@ ROCm 说明：
 - 如果手工安装了 ROCm wheel，但项目仍使用默认依赖配置档，验证时使用
   `uv run --no-sync`（或 `UV_NO_SYNC=1 make ...`）；自动同步会重新安装默认配置档的
   CUDA wheel。
-- 安装 ROCm PyTorch 且 GPU 可用时，接受 `cuda` buffer 的 `HOST_BRIDGE` 后端会让
-  Manager/TorchEnv tensor 使用当前 GPU。MuJoCo 物理仿真仍在 CPU 上，通过 packed
-  传输连接 GPU 上的 observation、action、reward 与 reset。CUDA-only 物理后端仍不
-  支持 ROCm。见 {doc}`/adr/ADR-0012-sole-tensor-manager-and-scoped-backends`。
+- 安装 ROCm PyTorch 且 GPU 可用时，接受 `cuda` buffer 的 `HOST_BRIDGE` 后端只有在
+  训练进程显式请求该 learner 设备（并路由 `manager_torch_device`）时才会让
+  Manager/TorchEnv tensor 使用当前 GPU；直接构造环境的默认仍是 CPU。MuJoCo 物理
+  仿真仍在 CPU 上，通过 packed 传输连接显式请求的 GPU observation、action、reward
+  与 reset。CUDA-only 物理后端仍不支持 ROCm。见
+  {doc}`/adr/ADR-0012-sole-tensor-manager-and-scoped-backends`。
 - 从 PyPI 安装（不克隆仓库）时，`make sync-rocm` 不适用；先从 PyTorch ROCm 索引
   安装仓库验证过的 torch build，再安装 `unilab`。发布的依赖范围是
   `torch>=2.9,<2.15`，pip 会保留已安装的 ROCm build，不会替换为 CUDA wheel：

docs/sphinx/source/zh_CN/5-reference/5-support_matrix.md

@@ -65,12 +65,12 @@ uv run scripts/generate_support_matrix.py --write
 
 | Backend | Execution / process / data plane | Torch devices | CUDA runtime | Linux+CUDA | macOS | ROCm | Worker | Reset randomization | Fixed variants | Host callbacks | Packed bridge |
 |---|---|---|---|---|---|---|---|---|---|---|---|
-| `mujoco` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | unknown | unknown | 不支持 | 支持 |
-| `motrix` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
-| `drake` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
+| `mujoco` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | unknown | unknown | 不支持 | 支持 |
+| `motrix` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
+| `drake` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
 | `mjwarp` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 不支持 |
 | `newton` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 不支持 |
-| `superdex` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
+| `superdex` | Host bridge / in-process / host bridge | CPU / CUDA | Required only when the learner requests CUDA state/control buffers | Supported: CPU-authoritative physics with optional CUDA Torch buffers | CPU-authoritative host bridge only; no CUDA physics claim | CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`) | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 支持 |
 | `genesis` | Device-resident / in-process / direct | CUDA | Required for the entire tensor lifecycle | Supported: Linux CUDA only | Unsupported; no CPU, MPS, or ROCm fallback | Unsupported; no CPU, MPS, or ROCm fallback | In-process; no external Python worker | 不支持 | 不支持 | 不支持 | 不支持 |
 
 ### Entrypoint x Task Owner
@@ -114,7 +114,8 @@ NVIDIA 驱动与容器运行时不匹配，再调整任务配置。设置了
 `CUDA_VISIBLE_DEVICES` 时，后端序号指向重映射后的命名空间，而不是宿主机
 全局物理索引。
 
-在 macOS 与 ROCm 上请使用 CPU-authoritative host-bridge 后端。在 ROCm 上，
-ROCm PyTorch 的 GPU 可用且后端接受当前 `cuda` 设备时，Manager/TorchEnv 使用
-当前 GPU；CPU 物理仿真不要求 Manager tensor 使用 CPU。host bridge 使用 GPU
-Torch buffer 不代表 GPU physics，也不代表 device-resident 后端生命周期。
+在 macOS 与 ROCm 上请使用 CPU-authoritative host-bridge 后端。在 ROCm 上，只有
+训练进程显式请求该 learner 设备且后端接受被路由的 `cuda`
+`manager_torch_device` 时，Manager/TorchEnv 才使用当前 GPU；直接构造环境仍为
+CPU。host bridge 使用 GPU Torch buffer 不代表 GPU physics，也不代表
+device-resident 后端生命周期。

scripts/generate_support_matrix.py

@@ -25,16 +25,17 @@
 
 _SUPPORT_LEVEL_ENUM = re.compile(r"\bSupportLevel\.([A-Z][A-Z_]+)\b")
 
-_ROCM_HOST_BRIDGE_LABEL = "Supported: CPU-authoritative physics with ROCm Torch buffers (`cuda`)"
+_ROCM_HOST_BRIDGE_LABEL = "CPU-authoritative host bridge with explicit ROCm Torch buffers (`cuda`)"
 
 
 def _with_support_level_values(content: str) -> str:
     """Render hardened UniSim lifecycle enum values as their public strings."""
     rendered = _SUPPORT_LEVEL_ENUM.sub(lambda match: match.group(1).lower(), content)
     # UniSim's static inventory is deliberately conservative: it only states
     # that no ROCm-native CUDA-only fallback exists. Host bridges accepting
-    # `cuda` Torch buffers also accept ROCm PyTorch's CUDA-namespace API while
-    # physics remains CPU-authoritative; make that repository guarantee visible.
+    # `cuda` Torch buffers also accept ROCm PyTorch's CUDA-namespace API,
+    # but only when the process explicitly requests those Manager carriers;
+    # make that opt-in repository guarantee visible.
     return rendered.replace(
         "CPU-authoritative host bridge only; no ROCm CUDA-only fallback",
         _ROCM_HOST_BRIDGE_LABEL,

src/unilab/base/process_device.py

@@ -16,8 +16,9 @@
 from typing import Any, cast
 
 # These backends consume an explicit integer device id while materializing
-# their simulator.  MuJoCo/Motrix/Drake either run on the host or own their
-# device selection internally and must not receive a synthetic override.
+# their simulator.  MuJoCo/Motrix/Drake run CPU-authoritative physics; their
+# optional accelerator Manager buffers are selected separately through
+# ``manager_torch_device``.
 BACKEND_ENV_DEVICE_FIELDS: dict[str, str] = {
     "isaacgym": "isaacgym_device_id",
     "isaacsim": "isaacsim_device_id",
@@ -41,6 +42,11 @@
 # integer payload sent to that worker before construction.
 _EXTERNAL_CUDA_IPC_BACKENDS = {"isaacgym", "isaacsim"}
 
+# HOST_BRIDGE backends accept optional accelerator Torch carriers across their packed
+# boundary. A process may explicitly request those carriers on its learner Torch
+# device; unset non-accelerator requests retain the owner/default CPU placement.
+_HOST_BRIDGE_TORCH_BACKENDS = {"mujoco", "motrix", "drake", "superdex"}
+
 
 # Set once ``bind_genesis_process_device`` has pinned CUDA_VISIBLE_DEVICES for
 # this process.  Genesis/Quadrants binds its CUDA runtime to the first visible
@@ -170,6 +176,35 @@ def apply_backend_env_device_override(
     return result
 
 
+def apply_manager_torch_device_override(
+    env_cfg_override: Mapping[str, Any] | None,
+    backend_type: str,
+    *,
+    learner_device: str | None = None,
+) -> dict[str, Any]:
+    """Return an env override selecting optional HOST_BRIDGE Torch carriers.
+
+    The input mapping is never mutated. Explicit CPU requests force CPU carriers.
+    Unset and non-CUDA learner requests retain owner/default placement. A CUDA request
+    is validated against backend capabilities when the environment binds.
+    DEVICE_RESIDENT backends own placement and never receive this synthetic field.
+    """
+
+    result = dict(env_cfg_override) if env_cfg_override is not None else {}
+    if _normalize_backend(backend_type) not in _HOST_BRIDGE_TORCH_BACKENDS:
+        return result
+    if learner_device is None:
+        return result
+    device = str(learner_device).strip()
+    if not device or device.lower() == "cpu":
+        result["manager_torch_device"] = "cpu"
+        return result
+    if device.split(":", 1)[0].lower() != "cuda":
+        return result
+    result["manager_torch_device"] = device
+    return result
+
+
 def resolve_backend_process_device(backend_type: str, learner_device: str | None) -> str | None:
     backend = _normalize_backend(backend_type)
     if (
@@ -399,6 +434,7 @@ def _reset_genesis_device_pin_for_tests() -> None:
 __all__ = [
     "BACKEND_ENV_DEVICE_FIELDS",
     "apply_backend_env_device_override",
+    "apply_manager_torch_device_override",
     "bind_backend_process_device",
     "bind_backend_process_device_for_backend",
     "bind_genesis_process_device",