[Misc] AMD Build Improvements (#12923)

csrc/moe/moe_align_sum_kernels.cu

@@ -3,7 +3,7 @@
 #include <c10/cuda/CUDAGuard.h>
 
 #include <ATen/ATen.h>
-#include <THC/THCAtomics.cuh>
+#include <ATen/cuda/Atomic.cuh>
 
 #include "../cuda_compat.h"
 #include "../dispatch_utils.h"

vllm/model_executor/models/registry.py

@@ -205,6 +205,14 @@ _VLLM_MODELS = {
     **_FALLBACK_MODEL,
 }
 
+# This variable is used as the args for subprocess.run(). We
+# can modify  this variable to alter the args if needed. e.g.
+# when we use par format to pack things together, sys.executable
+# might not be the target we want to run.
+_SUBPROCESS_COMMAND = [
+    sys.executable, "-m", "vllm.model_executor.models.registry"
+]
+
 
 @dataclass(frozen=True)
 class _ModelInfo:
@@ -502,8 +510,7 @@ def _run_in_subprocess(fn: Callable[[], _T]) -> _T:
 
         # cannot use `sys.executable __file__` here because the script
         # contains relative imports
-        returned = subprocess.run(
-            [sys.executable, "-m", "vllm.model_executor.models.registry"],
+        returned = subprocess.run(_SUBPROCESS_COMMAND,
                                   input=input_bytes,
                                   capture_output=True)