[XPU][CI] enhance xpu test support (#20652)

Signed-off-by: Ma, Liangliang <liangliang.ma@intel.com>
Co-authored-by: zhenwei-intel <zhenweiliu@habana.ai>
This commit is contained in:
Liangliang Ma
2025-07-09 16:53:09 +00:00
committed by GitHub
co-authored by zhenwei-intel
parent eb58f5953d
commit a3e4e85ece
5 changed files with 18 additions and 12 deletions
@@ -53,3 +53,6 @@ class XpuCommunicator(DeviceCommunicatorBase):
else:
output_tensor = None
return output_tensor
def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
dist.broadcast(input_, src=src, group=self.device_group)
+6 -4
View File
@@ -240,6 +240,8 @@ class GroupCoordinator:
if current_platform.is_cuda_alike():
self.device = torch.device(f"cuda:{local_rank}")
elif current_platform.is_xpu():
self.device = torch.device(f"xpu:{local_rank}")
elif current_platform.is_out_of_tree():
self.device = torch.device(
f"{current_platform.device_name}:{local_rank}")
@@ -1317,13 +1319,13 @@ def in_the_same_node_as(pg: Union[ProcessGroup, StatelessProcessGroup],
def is_global_first_rank() -> bool:
"""
Check if the current process is the first rank globally across all
Check if the current process is the first rank globally across all
parallelism strategies (PP, TP, DP, EP, etc.).
Unlike group-specific checks like `get_tensor_model_parallel_rank() == 0`
or `get_pp_group().is_first_rank`, this function checks the global rank
across all parallelism dimensions.
Returns:
bool: True if this is the global first rank (rank 0), False otherwise.
Returns True if distributed is not initialized (single process).
@@ -1352,7 +1354,7 @@ def _node_count(pg: Union[ProcessGroup, StatelessProcessGroup]) -> int:
Args:
pg: The process group to analyze
Returns:
int: The total number of nodes
"""