Spaces:
Running on Zero
Running on Zero
| """The denoising half of a split MiniMax-H3 deployment, for the `ref2va` partition. | |
| `MiniMaxH3Blocks` is cut at its `text_encoder` step: the 62 GiB Qwen3-VL conditioner runs in | |
| `multimodalart/qwen3vl-conditioner` (a public embedding service this Space calls over the gradio API), everything | |
| else — the `transformer_ref` partition, both autoencoders, the two schedulers — runs here. `prompt_embeds` + | |
| `text_token_tags` is the whole wire format between the two halves. | |
| Same layout as the character-swap H3 Space (`hugging-apps/minimax-h3-character-swap-lora`), which ships the same | |
| subclass against the same diffusers classes. | |
| """ | |
| import torch | |
| from diffusers.modular_pipelines.minimax_h3.before_encoder import MiniMaxH3Ref2VASetupStep | |
| from diffusers.modular_pipelines.minimax_h3.encoders import MiniMaxH3Ref2VAReferenceEncoderStep | |
| from diffusers.modular_pipelines.modular_pipeline import SequentialPipelineBlocks | |
| from diffusers.modular_pipelines.minimax_h3.modular_blocks_minimax_h3 import ( | |
| MiniMaxH3DecodeStep, | |
| MiniMaxH3Ref2VACoreDenoiseStep, | |
| ) | |
| from diffusers.modular_pipelines.modular_pipeline_utils import OutputParam | |
| class MiniMaxH3Ref2VAGeneratorBlocks(SequentialPipelineBlocks): | |
| """The denoising half of a split `ref2va`: the `ref2va` branch with its `text_encoder` step removed. | |
| `reference_encoder` stays here, next to the two autoencoders it runs: its output shapes are where every | |
| reference block's geometry in the packed layout comes from. | |
| """ | |
| model_name = "minimax-h3" | |
| # `after_denoise` is deliberately absent: in diffusers 0.41.0 `MiniMaxH3Ref2VACoreDenoiseStep` already ends with | |
| # its own `MiniMaxH3AfterDenoiseStep`, so a second copy here would run the unpack twice and crash on the | |
| # audio reshape (`shape '[2, 207, 207]' is invalid for input of size 13248`). The pre-merge PR commit the | |
| # sibling deployment pinned did not nest it, which is where the duplication came from. | |
| block_classes = [ | |
| MiniMaxH3Ref2VASetupStep, | |
| MiniMaxH3Ref2VAReferenceEncoderStep, | |
| MiniMaxH3Ref2VACoreDenoiseStep, | |
| MiniMaxH3DecodeStep, | |
| ] | |
| block_names = ["setup", "reference_encoder", "denoise", "decode"] | |
| def description(self): | |
| return ( | |
| "The denoising half of a split MiniMax-H3 `ref2va` deployment: the `ref2va` branch of `MiniMaxH3Blocks` " | |
| "without its text-encoder step, so `prompt_embeds` and `text_token_tags` come in as inputs and the " | |
| "62 GiB Qwen3-VL conditioner is never loaded here. The transformer is the `transformer_ref` partition." | |
| ) | |
| def outputs(self): | |
| # Inlined from `MiniMaxH3Blocks.outputs` (diffusers 0.41.0): the pre-merge PR exposed this list through a | |
| # `_generation_outputs()` helper that the release dropped, so the same three params are spelled out here. | |
| return [ | |
| OutputParam.template("videos", description="The generated video."), | |
| OutputParam( | |
| "audio", | |
| type_hint=torch.Tensor, | |
| description="The generated soundtrack, of shape `(1, 2, num_samples)`.", | |
| ), | |
| OutputParam("sampling_rate", type_hint=int, description="Sample rate of the generated soundtrack in Hz."), | |
| ] | |