[feat] Add LingBot-World v2 causal-fast streaming support

[lzx1413]

Description

Add LingBot-World v2 support on top of the shared causal-fast runtime, including offline image-to-video generation, four-GPU bidirectional WebRTC streaming, and v2-focused documentation.

Motivation

LingBot-World v2 uses a different public checkpoint layout and causal-fast defaults from v1. This PR integrates v2 without duplicating the runtime, aligns shared inference behavior with the reference implementation, and reduces streaming memory growth for long sessions.

Type of Change

• Bug fix
• New feature
• Performance improvement
• Code refactoring
• Documentation update

Changes Made

• Add a session-owned VAE encoder cache alongside the existing decoder cache; encode conditioning video incrementally per chunk and reset cache state after the final chunk.
• Align the shared LingBot causal-fast runtime with the reference implementation:
  • tensor-space bicubic image preprocessing
  • VAE/text precision handling
  • configurable timestep selection and stricter checkpoint validation
  • incremental noise and condition-latent generation instead of duration-sized chunk lists
  • centralized frame alignment and derived chunk completion
• Add LingBotWorldV2Pipeline with public checkpoint validation, camera-control enforcement, offline v2 example, and four-GPU v2 streaming example.
• Configure v2 streaming for camera control, Torch SDPA, chunk_size=4, and a two-minute session limit.
• Simplify the WebRTC demo so LingBot generation parameters come from the service PPL configuration instead of duplicated browser-side controls.
• Update English and Chinese README content and add a News entry linking to LingBot-World v2.

Testing

• Focused unit tests pass
• Manual offline and streaming testing performed
• Benchmarks added/updated

Test commands:

.venv/bin/python -m pytest tests/unit/pipelines/lingbot_world_fast tests/unit/pipelines/lingbot_world_v2 -q
# 61 passed

.venv/bin/python -m compileall -q \
  examples/lingbot \
  telefuser/pipelines/lingbot_world_fast/service.py

git diff --check

Manual validation:

• Generated v1 offline output on four H100 GPUs: 81 frames.
• Generated v2 offline output on four H100 GPUs: 77 frames.
• Started the four-GPU v2 WebRTC service and connected the browser demo through TURN relay.

Checklist

• Code follows the project's coding standards (ruff)
• Commit hooks pass
• Pre-commit hooks pass for all files (pre-commit run --all-files not run)
• All repository tests pass (pytest tests/ not run)
• New tests added for new functionality
• Documentation updated (README, CLAUDE.md, docstrings)
• Commit messages are clear and descriptive
• PR title follows the convention: [TYPE] Brief description

Related Issues

N/A

Additional Notes

• The v2 streaming PPL allows sessions up to 120 seconds. At 16 FPS with chunk_size=4, this resolves to 1917 output frames (119.75 seconds) to preserve complete latent chunks.
• The two-minute limit is a configured ceiling; actual memory capacity still depends on resolution and concurrent GPU workloads.
• No custom kernels were modified.

GPU Architecture Support

• SM80 (Ampere, Ada Lovelace)
• SM90 (Hopper H100)
• SM100+ (Blackwell)

Performance Impact

Long-running LingBot sessions no longer retain duration-sized noise and VAE condition chunk lists. Conditioning VAE encoding is incremental and cache state is session-owned, reducing session memory growth and preserving isolation between concurrent sessions.

[Copilot]

Pull request overview

This PR adds LingBot-World v2 support by introducing a v2 pipeline facade on top of the shared LingBot “causal-fast” runtime, while refactoring the shared runtime to generate/encode incrementally per chunk (reducing long-session memory growth) and updating streaming/offline examples plus documentation.

Changes:

• Introduce LingBotWorldV2Pipeline + public-checkpoint validation and add v2 offline + 4-GPU bidirectional streaming examples.
• Refactor LingBot causal-fast runtime to be chunk-incremental (noise generation + VAE conditioning encode cache), add frame_policy, and make timestep selection configurable/validated.
• Simplify the WebRTC demo and update English/Chinese READMEs + project documentation notes.

Reviewed changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/unit/pipelines/lingbot_world_v2/test_service.py	Adds unit tests validating v2 example/service wiring and CLI forwarding.
tests/unit/pipelines/lingbot_world_fast/test_service_action_loop.py	Updates action-loop test expectations for new done semantics.
tests/unit/pipelines/lingbot_world_fast/test_runtime_baseline.py	Updates baseline runtime tests for incremental noise/condition handling and v1/v2 shared contracts.
tests/unit/pipelines/lingbot_world_fast/test_pipeline_call.py	Updates pipeline-call tests to use session status (RELEASED) and done completion.
telefuser/service/api/stream_schema.py	Changes WebRTC offer request fps default to None.
telefuser/pipelines/lingbot_world_v2/pipeline.py	Adds v2 pipeline facade + checkpoint layout/shape validation.
telefuser/pipelines/lingbot_world_v2/init.py	Exports v2 pipeline symbols.
telefuser/pipelines/lingbot_world_fast/session.py	Adds frame_policy, per-session noise generator + VAE encoder cache state, and chunk_count property.
telefuser/pipelines/lingbot_world_fast/service.py	Adds session default config plumbing, duration-based frame calculation, and max_generation_seconds enforcement.
telefuser/pipelines/lingbot_world_fast/pipeline.py	Refactors image preprocessing, per-chunk condition encoding, per-chunk noise generation, and timestep plumbing.
telefuser/pipelines/lingbot_world_fast/denoising.py	Adds validated timestep selection and autocast around DiT calls.
telefuser/pipelines/lingbot_world_fast/control.py	Simplifies deferred control handling to plain callables.
telefuser/pipelines/lingbot_world_fast/init.py	Updates exports after timesteps refactor.
telefuser/models/wan_video_vae.py	Adds session-owned incremental VAE encode cache + cached encode method.
telefuser/models/lingbot_world_fast_dit.py	Tightens checkpoint validation/loading, adjusts dtype handling and output casting.
README.md	Updates news and streaming demo docs to reference LingBot-World v2.
README_zh.md	Chinese README updates mirroring v2 additions.
examples/stream_server/webrtc_bidirectional_demo.py	Removes browser-side LingBot parameter controls and stops always sending fps.
examples/lingbot/stream_lingbot_world_v2.py	Adds v2 streaming service example with PPL-configured defaults + 2-minute cap.
examples/lingbot/stream_lingbot_world_fast.py	Updates v1 streaming example to use service-side defaults config.
examples/lingbot/README.md	Updates LingBot examples README to cover v1 + v2 and fixed frame counts per PPL config.
examples/lingbot/lingbot_world_v2_image_to_video_h100.py	Adds v2 offline camera-controlled example + CLI.
examples/lingbot/lingbot_world_fast_image_to_video_h100.py	Updates v1 offline example to align with new session config defaults and removes CLI frame_num override.
CLAUDE.md	Documents LingBot streaming state constraints and adds pipeline directory entries.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.