Feature Request: Support for smaller text encoders (Gemma 3 4B / quantized Gemma) for <32GB VRAM GPUs for those not using ComfyUI (instead CLI or python code)

[BossLevel]

Basically this: Lightricks/ComfyUI-LTXVideo#303 but for those not using ComfyUI for inference.

RTX4090, 96Gb Ram