Spaces:
Build error
Build error
| """Die Hälften eines **gesplitteten** MiniMax-H3-Deployments, plus eine Video-only-Variante. | |
| MiniMax-H3 ist in bfloat16 195.9 GiB und ein ZeroGPU-Space wird bei 150 GB Storage evicted, also wird | |
| `MiniMaxH3Blocks` an seinem `text_encoder`-Step aufgeschnitten: das 62.14 GiB Qwen3-VL laeuft im Conditioner-Space, | |
| alles andere hier. `prompt_embeds` + `text_token_tags` ist das ganze Wire-Format dazwischen. | |
| Zusaetzlich zu den Bloecken von akhaliq/MiniMax-H3-Turbo-Lora definiert dieses Modul | |
| `MiniMaxH3MuteGeneratorBlocks`: identisch bis auf den Decode-Step, der nur `MiniMaxH3VideoDecodeStep` enthaelt. | |
| Damit laesst sich pro Request entscheiden, ob der Audio-VAE ueberhaupt angefasst wird. | |
| WICHTIG: Diese Klassen leiten von Block-Klassen des gepinnten diffusers-Commits ab (siehe requirements.txt). | |
| Beim Neupinnen die Klassennamen unten gegen den neuen Head pruefen. | |
| """ | |
| from diffusers.modular_pipelines.minimax_h3.before_encoder import MiniMaxH3Ref2VASetupStep | |
| from diffusers.modular_pipelines.minimax_h3.decoders import ( | |
| MiniMaxH3AfterDenoiseStep, | |
| MiniMaxH3VideoDecodeStep, | |
| ) | |
| from diffusers.modular_pipelines.minimax_h3.encoders import ( | |
| MiniMaxH3Ref2VAReferenceEncoderStep, | |
| MiniMaxH3Ref2VATextEncoderStep, | |
| MiniMaxH3TextEncoderStep, | |
| ) | |
| from diffusers.modular_pipelines.minimax_h3.modular_blocks_minimax_h3 import ( | |
| MiniMaxH3AutoKeyframeVaeEncoderStep, | |
| MiniMaxH3AutoResizeStep, | |
| MiniMaxH3CoreDenoiseStep, | |
| MiniMaxH3DecodeStep, | |
| MiniMaxH3Ref2VACoreDenoiseStep, | |
| _generation_outputs, | |
| ) | |
| from diffusers.modular_pipelines.modular_pipeline import SequentialPipelineBlocks | |
| from diffusers.modular_pipelines.modular_pipeline_utils import OutputParam | |
| def _wire_outputs(num_frames: bool = True) -> list[OutputParam]: | |
| """Das Wire-Format des Splits. `num_frames` deklariert nur die `ref2va`-Haelfte, deren Setup eines aufloest.""" | |
| return [ | |
| OutputParam.template("prompt_embeds"), | |
| OutputParam("text_token_tags", description="The per-row modality tag of every row of `prompt_embeds`."), | |
| OutputParam("height", type_hint=int, description="Resolved height of the generated video in pixels."), | |
| OutputParam("width", type_hint=int, description="Resolved width of the generated video in pixels."), | |
| *( | |
| [OutputParam("num_frames", type_hint=int, description="Resolved number of frames, of the form 17 * n + 5.")] | |
| if num_frames | |
| else [] | |
| ), | |
| ] | |
| class MiniMaxH3ConditionerBlocks(SequentialPipelineBlocks): | |
| """Die Conditioner-Haelfte: Keyframes auf dem Canvas plus das Qwen3-VL, gelesen an Layer 50.""" | |
| model_name = "minimax-h3" | |
| block_classes = [MiniMaxH3AutoResizeStep, MiniMaxH3TextEncoderStep] | |
| block_names = ["resize", "text_encoder"] | |
| def description(self): | |
| return ( | |
| "The conditioner half of a split MiniMax-H3 deployment: puts the keyframes onto the target canvas and " | |
| "encodes MiniMax-H3's presentation of the request into the `prompt_embeds` / `text_token_tags` pair the " | |
| "denoising half consumes. The frame count is the caller's to align." | |
| ) | |
| def outputs(self): | |
| return _wire_outputs(num_frames=False) | |
| class MiniMaxH3GeneratorBlocks(SequentialPipelineBlocks): | |
| """Die Denoising-Haelfte: `MiniMaxH3Blocks` ohne seinen `text_encoder`-Step. Video **und** Soundtrack.""" | |
| model_name = "minimax-h3" | |
| block_classes = [ | |
| MiniMaxH3AutoResizeStep, | |
| MiniMaxH3AutoKeyframeVaeEncoderStep, | |
| MiniMaxH3CoreDenoiseStep, | |
| MiniMaxH3AfterDenoiseStep, | |
| MiniMaxH3DecodeStep, | |
| ] | |
| block_names = ["resize", "vae_encoder", "denoise", "after_denoise", "decode"] | |
| def description(self): | |
| return ( | |
| "The denoising half of a split MiniMax-H3 deployment: the `t2va` / `fl2va` branch of `MiniMaxH3Blocks` " | |
| "without its text-encoder step, so `prompt_embeds` and `text_token_tags` come in as inputs and the " | |
| "62.14 GiB Qwen3-VL conditioner is never loaded here." | |
| ) | |
| def outputs(self): | |
| return _generation_outputs() | |
| class MiniMaxH3VideoOnlyDecodeStep(SequentialPipelineBlocks): | |
| """`MiniMaxH3DecodeStep` ohne den Audio-Zweig. | |
| Die Audio-Rows sind Teil derselben gepackten Sequenz und werden mitdenoised — das ist eine | |
| Checkpoint-Eigenschaft und laesst sich nicht abwaehlen. Was hier entfaellt, ist der Audio-VAE: | |
| er wird nicht deklariert, also auch nicht auf die Karte bewegt, nicht denormalisiert und nicht decodiert. | |
| Die denoisten Audio-Latents bleiben ungenutzt im PipelineState liegen. | |
| """ | |
| model_name = "minimax-h3" | |
| block_classes = [MiniMaxH3VideoDecodeStep] | |
| block_names = ["video"] | |
| def description(self): | |
| return "Decodes only the video rows of the packed sequence; the audio VAE is never touched." | |
| def outputs(self): | |
| return [OutputParam.template("videos", description="The generated video.")] | |
| class MiniMaxH3MuteGeneratorBlocks(SequentialPipelineBlocks): | |
| """`MiniMaxH3GeneratorBlocks` mit Video-only-Decode. Deklariert keinen `audio_vae`.""" | |
| model_name = "minimax-h3" | |
| block_classes = [ | |
| MiniMaxH3AutoResizeStep, | |
| MiniMaxH3AutoKeyframeVaeEncoderStep, | |
| MiniMaxH3CoreDenoiseStep, | |
| MiniMaxH3AfterDenoiseStep, | |
| MiniMaxH3VideoOnlyDecodeStep, | |
| ] | |
| block_names = ["resize", "vae_encoder", "denoise", "after_denoise", "decode"] | |
| def description(self): | |
| return ( | |
| "The denoising half of a split MiniMax-H3 deployment, video only: same blocks as " | |
| "`MiniMaxH3GeneratorBlocks` but with the audio decode step dropped, so the float32 audio VAE is neither " | |
| "declared nor moved onto the card." | |
| ) | |
| def outputs(self): | |
| return [OutputParam.template("videos", description="The generated video.")] | |
| class MiniMaxH3Ref2VAConditionerBlocks(SequentialPipelineBlocks): | |
| """Die Conditioner-Haelfte eines gesplitteten `ref2va`. Komponentenweise identisch zu `MiniMaxH3ConditionerBlocks`.""" | |
| model_name = "minimax-h3" | |
| block_classes = [MiniMaxH3Ref2VASetupStep, MiniMaxH3Ref2VATextEncoderStep] | |
| block_names = ["setup", "text_encoder"] | |
| def description(self): | |
| return ( | |
| "The conditioner half of a split MiniMax-H3 `ref2va` deployment: resolves the request plan and encodes " | |
| "MiniMax-H3's presentation of it into the `prompt_embeds` / `text_token_tags` pair." | |
| ) | |
| def outputs(self): | |
| return _wire_outputs() | |
| class MiniMaxH3Ref2VAGeneratorBlocks(SequentialPipelineBlocks): | |
| """Die Denoising-Haelfte eines gesplitteten `ref2va`.""" | |
| model_name = "minimax-h3" | |
| block_classes = [ | |
| MiniMaxH3Ref2VASetupStep, | |
| MiniMaxH3Ref2VAReferenceEncoderStep, | |
| MiniMaxH3Ref2VACoreDenoiseStep, | |
| MiniMaxH3AfterDenoiseStep, | |
| MiniMaxH3DecodeStep, | |
| ] | |
| block_names = ["setup", "reference_encoder", "denoise", "after_denoise", "decode"] | |
| def description(self): | |
| return ( | |
| "The denoising half of a split MiniMax-H3 `ref2va` deployment. The transformer is the `transformer_ref` " | |
| "partition." | |
| ) | |
| def outputs(self): | |
| return _generation_outputs() | |