MiniMax-H3 runs end-to-end on Intel XPU (incl. a custom INT8 path), but output is wrong (blocky video, garbled audio) — requesting native model support

[hanyan009]

Problem Description

MiniMax-H3 has no native support in LightX2V. We did a from-scratch native port (DiT, T2VA packing, dual video/audio rectified-flow scheduler, both VAEs) by reading the reference Python implementation in ComfyUI_RH_MinMaxH3, the only existing implementation of this architecture. Note that repo already ships an INT8-quantized checkpoint plus layerwise DiT offload / AdaLN weight-release tricks, and claims to run MiniMax-H3 on a single 24GB GPU — so the model is already known to be runnable at this memory budget, this isn't a from-scratch feasibility question.

The pipeline runs end-to-end without runtime errors on Intel XPU (1× Intel Arc B65, 32GB discrete VRAM each) — real text prompt → Qwen3-VL encode → 50-block DiT denoise loop → VAE decode → real .mp4/.wav output. But the generated content is wrong: the video shows visible block/tile artifacts, and the audio is garbled.

output_t2va_video.mp4

Confirmed working / Observed issue

• No crashes, no NaN/Inf anywhere; correct output shapes/dtypes; sane memory (GPU memory flat during block-offloaded inference, ~14.5GB peak)
• But: video has clear block/patch tiling artifacts (not a coherent scene), audio is garbled — reproducible

Request

Requesting official native support for MiniMax-H3. Given a working quantized+offloaded reference already exists (ComfyUI_RH_MinMaxH3, 24GB), this looks achievable on LightX2V too — but a review/implementation from maintainers familiar with the codebase's conventions would catch the remaining bugs far faster than we can by guessing against reference source alone. Happy to share our current implementation as a starting point.

[gushiqiao]

We already have an initial implementation in PR #1326 , and we are continuing to optimize both inference latency and GPU memory usage.

In the current version, enabling block offload allows the BF16 model to run with less than 24 GB of GPU memory. If you would like to use INT8 quantization, you can convert the model with the following command:

python tools/convert/converter.py
--source /path/to/MiniMax-H3/transformer
--output /path/to/MiniMax-H3/quantized/int8
--output_name minimax_h3_int8
--model_type h3
--quantized
--linear_type int8
--device cuda
--single_file