[RL] Support collective rpc api to entrypoint && Support custom input output (#1646)

Signed-off-by: knlnguyen1802 <knlnguyen1802@gmail.com>
Signed-off-by: Samit <285365963@qq.com>
Co-authored-by: Samit <285365963@qq.com>
This commit is contained in:
knlnguyen1802andSamit authored and GitHub committed 2026-03-09 10:12:43 +08:00
1 parent a9f694b08a
commit 92e1a54419
11 files changed
+476 -18

No files matched your search

@@ -63,6 +63,7 @@ class DiffusionWorker:
local_rank: int,
rank: int,
od_config: OmniDiffusionConfig,
skip_load_model: bool = False,
):
self.local_rank = local_rank
self.rank = rank
@@ -79,8 +80,9 @@ class DiffusionWorker:
od_config=self.od_config,
device=self.device,
)
self.load_model(load_format=self.od_config.diffusion_load_format)
self.init_lora_manager()
if not skip_load_model:
self.load_model(load_format=self.od_config.diffusion_load_format)
self.init_lora_manager()
logger.info(f"Worker {self.rank}: Initialization complete.")
def init_device(self) -> None:
@@ -524,10 +526,15 @@ class WorkerWrapperBase:
worker_class = self._prepare_worker_class()
# Create the actual worker instance
# When custom_pipeline_args is provided, skip initial model loading
# since re_init_pipeline will handle it. This avoids allocating memory
# through CuMemAllocator twice, which causes assertion failures in
# sleep mode.
self.worker = worker_class(
local_rank=gpu_id,
rank=gpu_id,
od_config=od_config,
skip_load_model=(self.custom_pipeline_args is not None),
)
# Re-initialize pipeline with custom pipeline if provided