[MUSA] Add visible-device alias support

README.md

@@ -392,7 +392,7 @@ See `src/torchada/_mappings/` for 400+ mapping rules grouped by API domain.
 
 ```
 # pyproject.toml or requirements.txt
-torchada>=0.1.83
+torchada>=0.1.81
 ```
 
 ### Step 2: Conditional Import

README_CN.md

@@ -376,7 +376,7 @@ if torchada.is_gpu_device(device):  # 在 CUDA 和 MUSA 上都能工作
 
 ```
 # pyproject.toml 或 requirements.txt
-torchada>=0.1.83
+torchada>=0.1.81
 ```
 
 ### 步骤 2：条件导入

benchmarks/benchmark_history.json

@@ -3,7 +3,7 @@
   "description": "Historical benchmark results for torchada performance tracking",
   "results": [
     {
-      "version": "0.1.83",
+      "version": "0.1.81",
       "date": "2026-01-29",
       "platform": "MUSA",
       "pytorch_version": "2.7.1",

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "torchada"
-version = "0.1.83"
+version = "0.1.81"
 description = "Adapter package for torch_musa to act exactly like PyTorch CUDA"
 readme = "README.md"
 license = {text = "MIT"}

src/torchada/__init__.py

@@ -24,7 +24,7 @@
     from torch.utils.cpp_extension import CUDAExtension, BuildExtension, CUDA_HOME
 """
 
-__version__ = "0.1.83"
+__version__ = "0.1.81"
 
 from . import cuda, utils
 

src/torchada/_patch.py

@@ -418,7 +418,9 @@ def _configure_cuda_graph_debug_dump_dir() -> Optional[str]:
     try:
         os.makedirs(dump_dir, exist_ok=True)
     except Exception as exc:  # noqa: BLE001
-        warnings.warn(f"TORCHADA_CUDA_GRAPH_DEBUG_DUMP_PATH={path_setting!r} is unusable: {exc!r}")
+        warnings.warn(
+            f"TORCHADA_CUDA_GRAPH_DEBUG_DUMP_PATH={path_setting!r} is unusable: {exc!r}"
+        )
         _cuda_graph_debug_dump_dir = None
         return None
 
@@ -678,7 +680,9 @@ def _discover_factory_functions() -> List[str]:
     if not names:
         names.update(_FALLBACK_FACTORY_FUNCTIONS)
     # ``*_like`` variants accept device= but are not device-injected by torch.
-    names |= {n + "_like" for n in tuple(names) if callable(getattr(torch, n + "_like", None))}
+    names |= {
+        n + "_like" for n in tuple(names) if callable(getattr(torch, n + "_like", None))
+    }
     for extra in _EXTRA_FACTORY_FUNCTIONS:
         if callable(getattr(torch, extra, None)):
             names.add(extra)
@@ -1212,7 +1216,9 @@ class ProfileWrapper:
 
         def __init__(self, *args, activities=None, **kwargs):
             translated_activities = _translate_activities(activities)
-            self._profiler = original_profile(*args, activities=translated_activities, **kwargs)
+            self._profiler = original_profile(
+                *args, activities=translated_activities, **kwargs
+            )
 
         def __enter__(self):
             return self._profiler.__enter__()
@@ -1303,7 +1309,9 @@ def patched_impl(self, *args, **kwargs):
         bound.apply_defaults()
 
         if bound.arguments.get("dispatch_key") in cuda_dispatch_key_map:
-            bound.arguments["dispatch_key"] = cuda_dispatch_key_map[bound.arguments["dispatch_key"]]
+            bound.arguments["dispatch_key"] = cuda_dispatch_key_map[
+                bound.arguments["dispatch_key"]
+            ]
 
         return original_impl(*bound.args, **bound.kwargs)
 
@@ -1566,7 +1574,9 @@ def _accepts_only_qv(func: Callable) -> bool:
     # module-level ``__getattr__`` (PEP 562), which would keep them out of
     # ``dir()``. flash_attn_varlen_func and flash_attn_with_kvcache are the ones
     # sglang's FA3 path forwards ``only_qv`` into.
-    candidate_names = {n for n in dir(flash_attn_interface) if n.startswith("flash_attn")}
+    candidate_names = {
+        n for n in dir(flash_attn_interface) if n.startswith("flash_attn")
+    }
     candidate_names.update(
         (
             "flash_attn_func",
@@ -1874,7 +1884,9 @@ def __getattr__(self, name):
             elif name in self._REMAP_ATTRS:
                 value = getattr(self._musa_module, self._REMAP_ATTRS[name])
             else:
-                raise AttributeError(f"module 'torch.accelerator' has no attribute '{name}'")
+                raise AttributeError(
+                    f"module 'torch.accelerator' has no attribute '{name}'"
+                )
         object.__setattr__(self, name, value)
         return value
 
@@ -2002,7 +2014,9 @@ def _patch_torch_accelerator():
 
     wrapper = _AcceleratorModuleWrapper(_original_torch_accelerator, torch.musa)
 
-    wrapper._set_override("synchronize", _make_patched_accelerator_synchronize(torch.musa))
+    wrapper._set_override(
+        "synchronize", _make_patched_accelerator_synchronize(torch.musa)
+    )
     device_index_cm, stream_cm = _make_accelerator_context_managers(wrapper)
     if not hasattr(_original_torch_accelerator, "device_index"):
         wrapper._set_override("device_index", device_index_cm)
@@ -2030,7 +2044,9 @@ def gdc_wait():
         raise NotImplementedError("tl.extra.cuda.gdc_wait is not supported on MUSA")
 
     def gdc_launch_dependents():
-        raise NotImplementedError("tl.extra.cuda.gdc_launch_dependents is not supported on MUSA")
+        raise NotImplementedError(
+            "tl.extra.cuda.gdc_launch_dependents is not supported on MUSA"
+        )
 
     if not hasattr(tl.extra.cuda, "gdc_wait"):
         tl.extra.cuda.gdc_wait = gdc_wait