Request: fp16 (not just bf16) full-precision transformer file
#13
by savitarTheGod - opened
Thanks for the LTX-2.5 release!
Currently diffusion_models only offers bf16 for full precision (ltx-2.5-22b-dev-transformer-bf16.safetensors), plus the comfy-int8-convrot and nvfp4 quantized variants. Turing-generation GPUs (RTX 20-series, e.g. 2080 Ti) don't have native bf16 tensor core support — that was introduced with Ampere. On these cards bf16 falls back to a slower path, while fp16 gets full tensor core acceleration.
Would it be possible to also publish an fp16 version of the transformer (dev and/or distilled) alongside the existing bf16 file? This would let older-generation GPU owners run full precision without the int8/nvfp4 quality tradeoff.