Port CUDA sources to MUSA in place

README.md

@@ -298,7 +298,7 @@ See `src/torchada/_mapping.py` for the complete mapping table (380+ mappings).
 
 ```
 # pyproject.toml or requirements.txt
-torchada>=0.1.69
+torchada>=0.1.70
 ```
 
 ### Step 2: Conditional Import

README_CN.md

@@ -293,7 +293,7 @@ if torchada.is_gpu_device(device):  # 在 CUDA 和 MUSA 上都能工作
 
 ```
 # pyproject.toml 或 requirements.txt
-torchada>=0.1.69
+torchada>=0.1.70
 ```
 
 ### 步骤 2：条件导入

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "torchada"
-version = "0.1.69"
+version = "0.1.70"
 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__ = "0.1.69"
+__version__ = "0.1.70"
 
 from . import cuda, utils
 

src/torchada/_mapping.py

@@ -2,11 +2,14 @@
 
 from ._mappings import MAPPING_RULE as _MAPPING_RULE  # noqa: F401
 
-# Extension file suffix mappings: convert .cu/.cuh to .mu/.muh so torch_musa's
-# musa_compile rule (which only adds -x musa for .mu/.muh) treats them correctly.
+# Keep source extensions unchanged during porting. Sources are ported in place
+# (no <dir>_musa mirror), and .cu/.cuh compile as MUSA via the patched
+# _is_musa_file (which selects the -x musa rule), so renaming to .mu/.muh is
+# unnecessary -- and avoiding it removes any need for include/source-path
+# rewriting.
 EXT_REPLACED_MAPPING = {
-    'cuh': 'muh',
-    'cu': 'mu',
+    'cuh': 'cuh',
+    'cu': 'cu',
     'cc': 'cc',
     'cpp': 'cpp',
     'cxx': 'cxx',

src/torchada/_mappings/cuda_runtime.py

@@ -1,8 +1,6 @@
 """cuda_runtime CUDA->MUSA porting rules."""
 
 MAPPING = {
-    '.cuh"': '.muh"',
-    '.cuh>': '.muh>',
     'cudaMalloc': 'musaMalloc',
     'cudaFree': 'musaFree',
     'cudaMemcpy': 'musaMemcpy',

src/torchada/_mappings/data_types.py

@@ -30,4 +30,11 @@
     'CUDA_R_8F_E5M2': 'MUSA_R_8F_E5M2',
     'cuda_fp16.h': 'musa_fp16.h',
     'cuda_bf16.h': 'musa_bf16.h',
+    # torch_musa's torch::headeronly snapshot omits these scalar-type headers;
+    # the c10 variants carry the same definitions. Only the dtype headers move
+    # (Exception.h / ScalarType.h under torch/headeronly resolve on torch_musa).
+    '#include <torch/headeronly/util/Float8_e4m3fn.h>': '#include <c10/util/Float8_e4m3fn.h>',
+    '#include <torch/headeronly/util/Float8_e4m3fnuz.h>': '#include <c10/util/Float8_e4m3fnuz.h>',
+    '#include <torch/headeronly/util/BFloat16.h>': '#include <c10/util/BFloat16.h>',
+    '#include <torch/headeronly/util/Half.h>': '#include <c10/util/Half.h>',
 }