[Kernel] Enable LTX DiffVAE fusions on SM100 and SM103

[mglyn]

Purpose

Follow-up to #7308 and its SM103 qualification: independently gate the LTX-2.5 DiffVAE pointwise fusions and TileLang FNA.

SM90       -> pointwise fusions + TileLang FNA
SM100/103  -> pointwise fusions + existing NATTEN
other      -> existing implementations

The pointwise kernels contain no architecture-specific schedule or instructions. Each runtime key is checked bit-for-bit on first use and falls back to the existing model math on a mismatch or launch failure. FNA remains SM90-only because its schedule is tuned for Hopper and regressed in the SM103 experiment.

SM103 evidence

Same-head ablation on one NVIDIA B300 (SM103), LTX-2.5 distilled two-stage, 1920x1088, 121 frames, BF16, batch 1, 8+3 steps, one warmup and three measured requests:

Task	Pointwise fusion	E2E median	Decode median
T2V	disabled	15.427 s	8.913 s
T2V	enabled	12.158 s (-21.2%)	5.641 s (-36.7%)
I2V	disabled	15.589 s	8.907 s
I2V	enabled	12.320 s (-21.0%)	5.645 s (-36.6%)

All 10 compared T2V/I2V output pairs were bitwise identical for returned RGB frames and audio. The fusion sources in this PR are unchanged from that experiment. SM100 shares the architecture-independent pointwise path and runtime correctness guard; this PR does not claim an SM100 latency measurement.

Test Plan

python -m pytest -q -o addopts='' \
  tests/diffusion/models/ltx2/test_ltx2_ops_platform.py \
  tests/diffusion/models/ltx2/test_ltx2_diffvae_ops.py \
  tests/diffusion/models/ltx2/test_ltx2_diffvae_attention.py \
  tests/diffusion/models/ltx2/test_ltx2_vae.py

pre-commit run --files \
  vllm_omni/diffusion/models/ltx2/ops/__init__.py \
  tests/diffusion/models/ltx2/test_ltx2_ops_platform.py

vLLM Version: 0.28.0

vLLM-Omni Commit: a1f8604511d061919e3d93b7608a60fc08c16304

Test Result

• 81 passed, 15 warnings in 26.48s
• All pre-commit hooks passed.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[mglyn]

Author self-review: this is a two-file dispatch-only change. I verified that SM100/SM103 resolve all four pointwise callbacks with fna=None, regardless of TileLang availability; SM90 still selects FNA only when TileLang is available; unsupported capabilities and non-CUDA/compile/grad paths still fall back. The focused LTX DiffVAE suite passed 81/81 and all pre-commit hooks passed. The PR reports measured performance only for SM103; it makes no SM100 latency claim.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/offloader.md, docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @Bounty-hunter @fhfuih @wtomin

Routing: @Bounty-hunter via module of the changed files, semantic router, CODEOWNERS; @fhfuih via module of the changed files, semantic router, CODEOWNERS; @wtomin via module of the changed files, semantic router, CODEOWNERS

@mglyn, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.