[Bug Fix] Hunyuan image cfg bugfix (#4752)

Signed-off-by: zhuhh97 <zhuhonghan@huawei.com>
Signed-off-by: dph97 <nenbaying@163.com>
Co-authored-by: zhuhh97 <z00806815@gmail.com>
Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
authored and GitHub committed 2026-07-15 13:08:15 +08:00
1 parent 2d58070f60
commit f84f0e72fd
7 files changed
+132 -13

No files matched your search

@@ -589,11 +589,7 @@ class GroupCoordinator:
if dst is None:
dst = self.group_next_rank
torch.distributed.send(
tensor,
self.ranks[dst],
group=(self.device_groups[self.rank_in_group % 2] if self.world_size == 2 else self.device_group),
)
torch.distributed.send(tensor, self.ranks[dst], group=self.device_group)
def recv(self, size: torch.Size, dtype: torch.dtype, src: int | None = None) -> torch.Tensor:
"""Receives a tensor from the src rank."""
@@ -602,11 +598,7 @@ class GroupCoordinator:
src = self.group_prev_rank
tensor = torch.empty(size, dtype=dtype, device=self.device)
torch.distributed.recv(
tensor,
self.ranks[src],
(self.device_groups[(self.rank_in_group + 1) % 2] if self.world_size == 2 else self.device_group),
)
torch.distributed.recv(tensor, self.ranks[src], self.device_group)
return tensor
def destroy(self):
@@ -710,6 +702,22 @@ class PipelineGroupCoordinator(GroupCoordinator):
self.skip_device_group = skip_device_group
assert self.skip_device_group is not None
def send(self, tensor: torch.Tensor, dst: int | None = None) -> None:
if dst is None:
dst = self.group_next_rank
group = self.device_groups[self.rank_in_group % 2] if self.world_size == 2 else self.device_group
torch.distributed.send(tensor, self.ranks[dst], group=group)
def recv(self, size: torch.Size, dtype: torch.dtype, src: int | None = None) -> torch.Tensor:
if src is None:
src = self.group_prev_rank
tensor = torch.empty(size, dtype=dtype, device=self.device)
group = self.device_groups[(self.rank_in_group + 1) % 2] if self.world_size == 2 else self.device_group
torch.distributed.recv(tensor, self.ranks[src], group)
return tensor
def reset_buffer(self):
self.recv_tasks_queue = []
self.receiving_tasks = []