[FEATURE] Add LTX-2.5 distilled T2V/I2V inference

[lzx1413]

Description

Add isolated LTX-2.5 distilled text-to-video and image-to-video inference with synchronized audio generation.

The implementation loads checkpoint components through ModuleManager and composes six stages: text encoding, video conditioning, denoising, latent upsampling, video decoding, and audio decoding. It supports one, two, or four H100 GPUs with Ulysses sequence parallelism, FSDP2 transformer sharding, and selectable dense attention backends.

The default DiffVAE path follows upstream CHUNKED_EAGER decoding with NATTEN cutlass-fna. This PR also adds upstream capture, artifact comparison, checkpoint inspection, and cold/warm benchmark tooling.

Motivation

LTX-2.5 differs substantially from the existing LTX-2.3 implementation in model architecture, checkpoint layout, conditioning, sampling, and audio/video decoding. A separate implementation preserves upstream behavior without coupling the two model families.

The integration also provides reproducible quality and performance gates, framework-native stage composition, and multi-GPU regression coverage.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality not to work as expected)
• Performance improvement
• Code refactoring
• Documentation update
• Other

Changes Made

• Added the isolated LTX-2.5 model stack, including the transformer, Gemma text encoder, audio modules, duration head, spatial upsampler, ConvVAE, and NATTEN-backed DiffVAE.
• Added split-checkpoint loading through ModuleManager and a six-stage distilled T2V/I2V pipeline.
• Added Ulysses sequence parallelism and FSDP2 execution for 2- and 4-GPU inference, including sequence padding, masking, gathering, and selectable attention implementations.
• Added standalone T2V and I2V H100 examples, a frozen default I2V reference image, and full two-GPU entries in examples/run_examples.py.
• Added upstream/TeleFuser capture, artifact comparison, checkpoint inspection, transformer structure comparison, and cold/warm benchmark harnesses.
• Added CPU-offloaded transformer block streaming with pinned host weights and reusable GPU buffers.
• Updated English and Chinese model indexes, News entries, example documentation, and the reusable example README template.
• Standardized file-based I2V entry points on input_image_path.

Testing

• Focused unit tests pass
• Manual full-resolution H100 testing performed
• Benchmarks added and executed

Test commands:

pytest tests/unit/models/ltx25 \
  tests/unit/pipelines/ltx25_distilled \
  tests/unit/validation/test_ltx25_capture_utils.py \
  tests/unit/validation/test_benchmark_ltx25_telefuser.py \
  tests/unit/validation/test_capture_ltx25_telefuser.py \
  tests/unit/validation/test_compare_ltx25_benchmarks.py

pytest tests/unit/test_run_examples.py -q
pytest tests/unit/pipelines/minimax_h3/test_examples.py -q

Verification results:

• Initial LTX model, pipeline, and validation suites: 61 passed, 1 skipped.
• Formal single-H100 T2V and I2V comparisons passed against frozen upstream captures.
• Full 121-frame two-H100 T2V and I2V runner regressions passed, including video and audio comparison.
• T2V: 61.90 dB PSNR and 0.999685 SSIM.
• I2V: 62.95 dB PSNR and 0.999671 SSIM.
• Commit-time Ruff and formatting hooks passed.

Checklist

• Code follows the project's coding standards
• Commit-time pre-commit hooks pass
• Full pre-commit run --all-files executed
• Full repository test suite (pytest tests/) executed
• New tests added for new functionality
• Documentation updated
• Commit messages are clear and descriptive
• PR title follows the project convention

Related Issues

N/A

Additional Notes

• The examples require the official split LTX-2.5 checkpoint layout; consolidated checkpoints are not accepted.
• Formal DiffVAE parity requires a compatible NATTEN/libnatten installation. The fallback path is not the formal quality or performance baseline.
• Multi-GPU denoising remains GPU-resident because FSDP2 and transformer CPU offload are mutually exclusive.
• BF16 sequence-parallel output is not expected to be bitwise identical to single-GPU output because rank-local tensor shapes may select different GEMM kernels.

GPU Architecture Support

• SM80 (Ampere)
• SM90 (Hopper H100)
• SM100+ (Blackwell)

Performance Impact

Matched single-H100 BF16 p50 end-to-end results:

Workload	Mode	Upstream cold / warm	TeleFuser cold / warm
T2V 1536x1024, 121 frames	CPU offload	76.78 / 77.29 s	64.44 / 60.09 s
I2V 896x512, 121 frames	CPU offload	65.02 / 64.52 s	51.98 / 44.08 s
T2V 1536x1024, 121 frames	No offload	58.28 / 55.29 s	46.79 / 46.93 s
I2V 896x512, 121 frames	No offload	48.45 / 42.34 s	30.24 / 30.29 s

The no-offload TeleFuser profile reserved 79.14 GB at peak.