vonkaiser/LTX-2.5-FP8-NVFP4

LTX-2.5 FP8 + NVFP4 Quantized Weights

Pre-quantized components from Lightricks/LTX-2.5 for GPU-resident deployment on Blackwell (RTX PRO 6000, SM120).

Components

File	Source	Precision	Size
transformer/ltx-2.5-22b-distilled-fp8.safetensors	LTX-2.5 22B distilled DiT	FP8 weight-only (float8_e4m3fn + float32 scales)	~21 GB
text_encoders/gemma4-12b-with-proj-nvfp4-torchao.safetensors	Gemma 4 12B + LTX projections	NVFP4 torchao (with .torchao_nvfp4 markers)	~7.4 GB

Quantization Details

Transformer (FP8)

• Each 2D bf16 weight matrix cast to float8_e4m3fn with per-tensor absmax scaling
• Scale stored as {key}_scale (float32) alongside each FP8 weight
• 1775 weight matrices quantized, 2574 dense tensors kept as-is
• Compatible with ltx-pipelines QuantizationKind.FP8_SCALED_MM (auto-detected from filename)

Text Encoder (NVFP4)

• torchao NVFP4DynamicActivationNVFP4WeightConfig with use_dynamic_per_tensor_scale=True, use_triton_kernel=True
• Each quantized linear stored as 4 keys: .weight (uint8), .weight_scale (float8_e4m3fn), .weight_scale_2 (float32), .torchao_nvfp4 (JSON metadata)
• 334 linear layers quantized, 352 dense tensors preserved
• Same on-disk format as vonkaiser/MiniMax-H3-NVFP4 text encoder

VRAM Budget (all resident, no CPU offload)

Component	Precision	Est. VRAM
Transformer	FP8	~21 GB
Text Encoder	NVFP4	~7.4 GB
Video VAE (DiffVAE, bf16)	BF16	~2 GB
Audio VAE (bf16)	BF16	~1 GB
Spatial Upscaler (bf16)	BF16	~1 GB
Duration Head (bf16)	BF16	~0.1 GB
Total		~32.5 GB
Free for activations		~64.5 GB

Source

Quantized from Lightricks/LTX-2.5 bf16 checkpoints:

• diffusion_models/ltx-2.5-22b-distilled-transformer-bf16.safetensors
• text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors

License

Same as source: ltx-2-community-license-agreement