[LTX-2] Modular pipelines: let callers skip the audio decode for video-only outputs #15044

Is your feature request related to a problem? Please describe.

The LTX-2 / LTX-2.5 modular pipelines always run LTX2AudioDecoderStep (audio VAE decode + vocoder), even when the caller only wants the video. For some IC-LoRAs, decoding audio makes no sense: Lightricks/LTX-2.5-22b-IC-LoRA-Alpha-Gen turns a reference video into its alpha matte, and the soundtrack decoded alongside it has no meaning; it is thrown away.

Today the only ways to avoid it are:

• output_type="latent", which also skips the video decode;
• removing the block (sub_blocks.pop("audio_decode") in each decoder chain), which changes the blocks for every later call. That is not usable when one loaded pipeline serves both video-only and audio-visual requests.

Describe the solution you'd like.

A per-call input, decode_audio (default True): pipe(..., decode_audio=False) skips the audio VAE and vocoder and returns audio=None. The video is unchanged: the audio latents are still denoised jointly with it.

Following the modular guide (a new block rather than a branch inside an existing one), LTX2AudioDecoderStep stays as is and is wrapped in a ConditionalPipelineBlocks, LTX2AutoAudioDecoderStep. That wrapper runs it unless decode_audio=False, the same pattern as LTX2AutoPromptEnhancerStep with enable_prompt_enhancement. It replaces LTX2AudioDecoderStep in the LTX-2 and LTX-2.5 decoder chains under the same audio_decode name, so workflow layouts do not change.

Describe alternatives you've considered.

• An if not decode_audio: branch inside LTX2AudioDecoderStep. It is simpler, but it is the specializing branch the modular guide advises against.
• A presence-based AutoPipelineBlocks trigger (skip_audio_decode) selecting a no-op block. Any non-None value would trigger it, False included.
• A video-only variant blockset without audio_decode. That gives no per-call choice.
• Skipping audio generation altogether. This is not possible without changing the video: the transformer always takes audio_hidden_states, and the video attends to the audio tokens.

Additional context.

I have a PR ready and will link it here. On the tiny test checkpoint, decode_audio=False returns audio=None and bit-identical videos for text2video and in_context.