Add WanGP PTD RefAsymOPD adapters and output heads

de1028e verified 21 days ago

2.22 kB

# MiniMax H3 Ref2VA PTD RefAsymOPD for WanGP

Eight scheduled rank-128 LoRAs converted from [taobao-content-ai's release](https://huggingface.co/taobao-content-ai/MiniMax-H3-PTD-RefAsymOPD-LoRAs/tree/5501eb8918ef8bcb858ff045b885bf47a2c4502b). Original checkpoint: `MiniMax-H3-Ref2VA-PTD-RefAsymOPD-8Step.safetensors`, SHA256 `6796802b9150110a3cd7f66891c9e6190502a432c30b0fa1aa299724cfc15cc3`.

Select **MiniMax H3 Ref2VA PTD RefAsymOPD 8-Step 33B** in WanGP. Its definition loads all eight files automatically and enables exactly one at each Euler step. It uses the existing full H3 Ref2VA base and the matching PDD heads in `minimax_h3/ptd_refasymopd/`. The **MiniMax H3 Ref2VA Pruned PTD RefAsymOPD 8-Step 20B** variant shares these adapters and heads with the pruned base. Use one to nine image references, eight steps, video flow shift 12, audio shift 3, and guidance 1. Prompt enhancement is optional.

The LoRAs live together under `loras/ptd_refasymopd/` on Hugging Face and under `minimax_h3/ptd_refasymopd/` inside WanGP's configured LoRA root. The PDD heads are ancillary model weights, stored below the checkpoint root. Do not apply all eight LoRAs simultaneously or combine them with another acceleration LoRA.

Each adapter contains 624 BF16 factors for 312 linear modules (50 transformer blocks and two text-refiner blocks), plus rank/alpha scaling values. SwiGLU output factors use WanGP's row ordering. The PDD file stores head zero and relative heads in FP32 to preserve the existing PDD computation without conversion during loading. All source adapter factors are preserved exactly after layout conversion.

The original adapters and base are subject to the [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). Credit: taobao-content-ai for PTD RefAsymOPD and MiniMax for H3. These files are an inference-format conversion, not a new training run.

Validation status: CPU conversion and contract checks passed, and the published files match the verified conversion checksums. Real BF16 and INT8 ConvRot generation validation is pending. Adapter shapes match both the full and pruned bases; generation quality on the pruned base remains unvalidated.

	# MiniMax H3 Ref2VA PTD RefAsymOPD for WanGP

	Eight scheduled rank-128 LoRAs converted from [taobao-content-ai's release](https://huggingface.co/taobao-content-ai/MiniMax-H3-PTD-RefAsymOPD-LoRAs/tree/5501eb8918ef8bcb858ff045b885bf47a2c4502b). Original checkpoint: `MiniMax-H3-Ref2VA-PTD-RefAsymOPD-8Step.safetensors`, SHA256 `6796802b9150110a3cd7f66891c9e6190502a432c30b0fa1aa299724cfc15cc3`.

	Select MiniMax H3 Ref2VA PTD RefAsymOPD 8-Step 33B in WanGP. Its definition loads all eight files automatically and enables exactly one at each Euler step. It uses the existing full H3 Ref2VA base and the matching PDD heads in `minimax_h3/ptd_refasymopd/`. The MiniMax H3 Ref2VA Pruned PTD RefAsymOPD 8-Step 20B variant shares these adapters and heads with the pruned base. Use one to nine image references, eight steps, video flow shift 12, audio shift 3, and guidance 1. Prompt enhancement is optional.

	The LoRAs live together under `loras/ptd_refasymopd/` on Hugging Face and under `minimax_h3/ptd_refasymopd/` inside WanGP's configured LoRA root. The PDD heads are ancillary model weights, stored below the checkpoint root. Do not apply all eight LoRAs simultaneously or combine them with another acceleration LoRA.

	Each adapter contains 624 BF16 factors for 312 linear modules (50 transformer blocks and two text-refiner blocks), plus rank/alpha scaling values. SwiGLU output factors use WanGP's row ordering. The PDD file stores head zero and relative heads in FP32 to preserve the existing PDD computation without conversion during loading. All source adapter factors are preserved exactly after layout conversion.

	The original adapters and base are subject to the [MiniMax H3 Community License Agreement](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). Credit: taobao-content-ai for PTD RefAsymOPD and MiniMax for H3. These files are an inference-format conversion, not a new training run.

	Validation status: CPU conversion and contract checks passed, and the published files match the verified conversion checksums. Real BF16 and INT8 ConvRot generation validation is pending. Adapter shapes match both the full and pruned bases; generation quality on the pruned base remains unvalidated.