ltx23_load_pipeline

Load the LTX-2.3 generation components from a single-file checkpoint

Description

Builds the transformer, connectors, video VAE, audio VAE, and vocoder with the LTX 2.3 configuration and streams the checkpoint weights into them. The Gemma3 text encoder ships separately (see load_gemma3_text_encoder).

Usage

ltx23_load_pipeline(checkpoint_path, device = "cuda", dtype = "bfloat16",
                    transformer_device = "cpu",
                    components = c("dit", "connectors", "vae", "audio_vae", "vocoder"),
                    pin = TRUE, attn_chunk = NULL, phase_offload = TRUE,
                    verbose = TRUE)

Arguments

  • checkpoint_path: Path to the single-file checkpoint (e.g. ltx-2.3-22b-distilled-1.1.safetensors) or to an fp8 artifact directory produced by ltx23_quantize_fp8. With the fp8 artifact, the transformer loads with CPU-resident fp8 weights that stream to device during the forward pass.
  • device: Character. Device for the small components (VAEs, vocoder, connectors) and, with fp8, the transformer residents.
  • dtype: Character. “bfloat16” (checkpoint native) or “float32”.
  • transformer_device: Character. Device for the transformer weights when loading the plain (non-fp8) checkpoint.
  • components: Character vector. Which components to load.
  • pin: Logical. Pin fp8 host memory (fp8 artifact only).
  • attn_chunk: Integer or NULL. Query-chunk size for attention (see ltx23_set_attn_chunk).
  • phase_offload: Logical. Load the small components (connectors, VAEs, vocoder) to the CPU; the pipeline moves each onto the compute device only for its phase.
  • verbose: Logical.

Value

A list with the loaded modules and the checkpoint config, class ltx23_pipeline.