📦 Load CLIP (Triggered)

Load CLIP text encoder with memory management and execution order control. [Parameters] trigger: Connect to any output to force execution order. keep_on_vram: Pin CLIP to GPU VRAM permanently. Protected from auto-eviction. If you change the CLIP, the old one is automatically unpinned. use_dram: Check DRAM cache (CPU RAM) for previously loaded CLIP. Much faster than disk reload. replace_cached: When switching CLIPs, evict this loader's previous CLIP from DRAM. Saves RAM. device: Load CLIP on GPU (default) or CPU. Use CPU to save VRAM for diffusion model. cache_to_local_ssd: RunPod only — copy to local NVMe for faster loading. auto_free_vram: Before loading from disk, if free VRAM < threshold, unpin all models and free VRAM. Only triggers for different model (same model reuses cache). min_free_vram_gb: VRAM threshold in GB for auto_free_vram. auto_free_dram: Before loading from disk, if free RAM < threshold, clear DRAM cache. Only triggers on DRAM miss. min_free_dram_gb: RAM threshold in GB for auto_free_dram. [Recipes] stable_diffusion: clip-l stable_cascade: clip-g sd3: t5 xxl / clip-g / clip-l stable_audio: t5 base mochi: t5 xxl cosmos: old t5 xxl lumina2: gemma 2 2B wan: umt5 xxl hidream: llama-3.1 or t5 omnigen2: qwen vl 2.5 3B qwen_image: Qwen VL models