Feat: Support LoRA and FP8 Adapter of Minimax-H3 and FastH3

[Uxito-Ada]

Summary

Enable Minimax-H3 LoRA adapter, and support the adapter to run with TeleFuser FP8 Linear

Supported adapter:

• MiniMax-H3 Turbo LoRA
• FastVideo fastvideo-lora-v2 to merge with dense FastH3, a FastVideo-finetuned H3

Adapter updates are merged into the BF16 base weights before optional FP8 conversion, so the FP8 quantizes the final adapted weights

Changes

• Detect fastvideo-lora-v2 metadata and apply W += B @ A without an implicit alpha/rank scale
• Stream and merge low-rank tensors plus exact .diff, .diff_b, and .diff_m residuals in float32
• Expose --adapter-path / --lora-path and --adapter-strength through the standard MiniMax H3 FL2VA H100 example
• Document BF16 adapter and FP8 Linear + adapter usage

Performance

Matched workload:

• 1 * H100 80GB
• 1344 x 768
• 124 frames
• 5.175 seconds
• Minimax-H3 Turbo LoRA, and FastH3 adapter

Minimax-H3 Turbo LoRA, compared with the only one community solution on LightX2V:

Compared with LightX2V, TeleFuser improves denoising time by 26.7% and throughput by 36.5%.

FastH3 adapter, compared with FastVideo:

Compared with LightX2V, TeleFuser improves denoising time by 36.4% and throughput by 57.2%.

In addition, it is seen that FP8 does not shrink peak GPU memory so much, which is because that both LightX2V and FastVideo applies CPU offloading, whose whole weights size is actually double of the FP8 runtime. Their CPU offloading is at the cost of high latency.

Generated Output Comparison

All three outputs contain 124 H.264 frames at 1344 x 768 and synchronized stereo AAC audio.

Adapter	Framework	Video
MiniMax-H3 Turbo LoRA	LightX2V BF16 + Sol	turbo_lightx2v_bf16_sol_resident_dit.mp4
MiniMax-H3 Turbo LoRA	TeleFuser FP8 + Sol	turbo_telefuser_fp8_sol_best.mp4
FastH3 adapter	FastVideo BF16 + FA4	fasth3_all_run_03.mp4
FastH3 adapter	TeleFuser FP8 + Sol	fasth3_telefuser_fp8_sol_best.mp4

Testing

• 149 passed: generic LoRA loader, MiniMax H3 model, pipeline, scheduler, parallelism, example, and adapter tests.
• Added hybrid low-rank plus exact-delta merge coverage.
• Added VSA replacement-gate rejection coverage.
• Added adapter CLI forwarding and adapter-before-FP8 ordering coverage.
• Ruff check and format check pass for all modified Python files.
• git diff --check passes.
• Verified the released dense-datafree safetensors header and tensor inventory with this branch.
• Verified the three representative MP4s for codec, resolution, frame count, duration, and audio stream.