add helpers.so and delete __init__ note

megatron/fused_kernels/__init__.py

@@ -31,8 +31,6 @@ def load(args):
 
     # Check if cuda 11 is installed for compute capability 8.0
     cc_flag = []
-    _, bare_metal_major, _ = _get_cuda_bare_metal_version(
-        cpp_extension.CUDA_HOME)
 
     # Build path
     srcpath = pathlib.Path(__file__).parent.absolute()
@@ -98,17 +96,13 @@ def load(args):
 
 
 def _get_cuda_bare_metal_version(cuda_dir):
-    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "--version"],
+    raw_output = subprocess.check_output([cuda_dir + "/bin/hipcc", "--version"],
                                          universal_newlines=True)
     output = raw_output.split()
-    print("output:", output)
     release_idx = output.index("version:") + 1
-    print("release_idx", release_idx)
     release = output[release_idx].split(".")
-    print("release", release)
     bare_metal_major = release[0]
     bare_metal_minor = release[1][0]
-    print("bare_metal_major, bare_metal_minor", bare_metal_major, bare_metal_minor)
 
     return raw_output, bare_metal_major, bare_metal_minor
 

megatron/initialize.py

@@ -237,7 +237,7 @@ def set_jit_fusion_options():
     if (TORCH_MAJOR > 1) or (TORCH_MAJOR == 1 and TORCH_MINOR >= 10):
         # nvfuser
         torch._C._jit_set_profiling_executor(True)
-        torch._C._jit_set_profiling_mode(True)
+        torch._C._jit_set_profiling_mode(False)
         torch._C._jit_override_can_fuse_on_cpu(False)
         torch._C._jit_override_can_fuse_on_gpu(False)
         torch._C._jit_set_texpr_fuser_enabled(False)