Revert "Fix nan value generated after custom all reduce (#8532)" (#8642)

Author: zhyncs

python/sglang/srt/distributed/device_communicators/custom_all_reduce.py

@@ -184,7 +184,7 @@ def __init__(
             # 8*world_size bytes where world_size is at most 8. Allocating 8MB
             # is enough for 131072 such tuples. The largest model I've seen only
             # needs less than 10000 of registered tuples.
-            self.rank_data = torch.zeros(
+            self.rank_data = torch.empty(
                 8 * 1024 * 1024, dtype=torch.uint8, device=self.device
             )
             self._ptr = ops.init_custom_ar(
@@ -194,14 +194,14 @@ def __init__(
         else:
             # meta data buffers need to be "uncached" for signal on MI200
             self.meta = ops.allocate_meta_buffer(ops.meta_size() + max_size)
-            self.buffer = torch.zeros(max_size, dtype=torch.uint8, device=self.device)
+            self.buffer = torch.empty(max_size, dtype=torch.uint8, device=self.device)
             handle = ops.get_meta_buffer_ipc_handle(self.meta)
             shard_data = (
                 bytes(handle),  # ipc handle to base ptr
                 0,  # offset of base ptr
             )
             handles, offsets = self._gather_ipc_meta(shard_data)
-            self.rank_data = torch.zeros(
+            self.rank_data = torch.empty(
                 8 * 1024 * 1024, dtype=torch.uint8, device=self.device
             )
             self._ptr = ops.init_custom_ar(
@@ -350,14 +350,14 @@ def should_custom_ar(self, inp: torch.Tensor):
     # or, in the context of cuda graphs, register_graph_buffers
     def all_reduce_reg(self, inp: torch.Tensor, out: torch.Tensor = None):
         if out is None:
-            out = torch.zeros_like(inp)
+            out = torch.empty_like(inp)
         ops.all_reduce_reg(self._ptr, inp, out)
         return out
 
     # all reduce, assuming inp tensor is NOT IPC registered
     def all_reduce_unreg(self, inp: torch.Tensor, out: torch.Tensor = None):
         if out is None:
-            out = torch.zeros_like(inp)
+            out = torch.empty_like(inp)
         ops.all_reduce_unreg(self._ptr, inp, self.buffer, out)
         return out
 
@@ -375,7 +375,7 @@ def all_reduce(
         buffer.
         """
         if out is None:
-            out = torch.zeros_like(inp)
+            out = torch.empty_like(inp)
         if registered:
             ops.all_reduce(self._ptr, inp, out, 0, 0)
         else:
@@ -398,7 +398,7 @@ def custom_all_reduce(self, input: torch.Tensor) -> Optional[torch.Tensor]:
             else:
                 # If warm up, mimic the allocation pattern since custom
                 # allreduce is out-of-place.
-                return torch.zeros_like(input)
+                return torch.empty_like(input)
         else:
             if _is_hip:
                 # note: outside of cuda graph context,