minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[diffusion] chore: remove default post-denoising dit offload in local mode (#15573)

Browse Source
This commit is contained in:
ryang
2025-12-22 17:01:35 +08:00
committed by GitHub
parent e62e17442e
commit 575a49dc0e
1 changed files with 0 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py
-14
View File
@@ -754,20 +754,6 @@ class DenoisingStage(PipelineStage):
torch.mps.current_allocated_memory(), torch.mps.current_allocated_memory(),
) )
# In offline local mode (`sglang generate`), offload transformer weights to CPU
# after denoising to reduce peak VRAM during VAE decoding.
if current_platform.is_cuda_alike() and server_args.is_local_mode:
for model in (self.transformer, self.transformer_2):
if model is not None:
model.to("cpu")
logger.info(
"Offloaded denoiser transformer weights to CPU after denoising to reduce peak VRAM during VAE decoding."
)
try:
torch.cuda.empty_cache()
except Exception:
pass
def _preprocess_sp_latents(self, batch: Req, server_args: ServerArgs): def _preprocess_sp_latents(self, batch: Req, server_args: ServerArgs):
"""Shard latents for Sequence Parallelism if applicable.""" """Shard latents for Sequence Parallelism if applicable."""
if get_sp_world_size() <= 1: if get_sp_world_size() <= 1: