Skip to content

[diffusion] split and batch LTX-2.3 one-stage guidance passes under CFG parallel #43684

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mickqian merged 6 commits into from
Oct 12, 2026
+354 −227
Merged

[diffusion] split and batch LTX-2.3 one-stage guidance passes under CFG parallel #43684

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
077a38d
[diffusion] perf: split LTX-2.3 one-stage guidance passes across CFG-…
niehen6174 Oct 11, 2026
ece501c
[diffusion] perf: batch LTX-2.3 one-stage guidance passes per rank
niehen6174 Oct 11, 2026
dea8e94
[diffusion] fix: skip LTX-2 video postprocess on non-output ranks
niehen6174 Oct 11, 2026
8843fc6
Merge main and preserve primary-rank LTX video postprocessing
mickqian Oct 11, 2026
4740c5d
Preserve decoded frames on all ranks for JoyEcho memory banks
mickqian Oct 11, 2026
8c3b1ae
Fix decoded frame retention at the JoyEcho request boundary
mickqian Oct 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions python/sglang/multimodal_gen/runtime/pipelines_core/stages/encode_while_decoding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@ def can_encode_while_decoding(
batch.save_output
and batch.return_file_paths_only
and not batch.return_raw_frames
and not (batch.extra or {}).get("retain_decoded_frames", False)
and batch.data_type == DataType.VIDEO
and not batch.enable_frame_interpolation
and not batch.enable_upscaling
Expand Down
1 change: 1 addition & 0 deletions ...lang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/joy_echo/setup.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,6 +48,7 @@ def _maybe_reset_multishot_session(self, batch: Req) -> None:
logger.info("JoyEcho memory bank reset for new multi-shot session")

def forward(self, batch: Req, server_args: ServerArgs) -> Req:
batch.extra["retain_decoded_frames"] = batch.enable_memory_bank
if not batch.enable_memory_bank:
return batch

Expand Down
12 changes: 10 additions & 2 deletions ...g/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/ltx_2/decoding_av.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,7 +3,11 @@
import torch
from diffusers.video_processor import VideoProcessor

from sglang.multimodal_gen.runtime.distributed import get_local_torch_device
from sglang.multimodal_gen.runtime.distributed import (
get_local_torch_device,
get_replica_group,
model_parallel_is_initialized,
)
from sglang.multimodal_gen.runtime.managers.memory_managers.component_manager import (
ComponentUse,
)
Expand Down Expand Up @@ -152,7 +156,11 @@ def forward(self, batch: Req, server_args: ServerArgs) -> OutputBatch:
raise
if output_paths is not None:
output_batch.output_file_paths = output_paths
else:
elif (
batch.extra.get("retain_decoded_frames", False)
or not model_parallel_is_initialized()
or get_replica_group().rank_in_group == 0
):
output_batch.output = self.video_processor.postprocess_video(
video, output_type="np"
)
Expand Down
Loading
Loading