[TroyHernandez]

LTX-2.5 on 8 and 12 GB cards, and in less host RAM.

Small cards

• ltx23_load_pipeline(resident_gb = ) keeps the leading NF4 blocks on the GPU and streams the rest from pinned host memory, one block per TorchScript call. Below 7 GB the text connectors stream too.
• load_gemma4_nf4(stream = TRUE) streams the Gemma 4 layers.
• ltx23_memory_profile() picks resident_gb = vram - 4 below 14 GB and stream_text below 12 GB.

Simulated with a ballast tensor on the 16 GB card, 768x512x121: 12 GB and 8 GB both render in ~84 s, the same as 16 GB.

Less host RAM

The one-process flow needs ~31 GB of host RAM (pinned staging plus Gemma 4). Two ways to need less:

1. Split text phase. Encode with Gemma, free it, run the connectors on a connectors-only pipeline with the new ltx23_text_embeds(), free that, then load the render pipeline without connectors and pass connector_embeds.
2. resident_from = "disk". The GPU-resident blocks keep no host copy and are re-read from the artifact each render (~1 GB/s through safetensors).

pin = FALSE on ltx23_load_pipeline now also disables the staging pins. Freed pinned memory stays in torch's host cache, so a pipeline you plan to drop has to load unpinned.

ltx23_memory_profile(host_ram_gb = ) detects available RAM and returns split_text (below 32 GB) and resident_from ("disk" below 22 GB).

Measured under a 20 GB cgroup cap, 768x512x121, render 2:

Setup	Host RAM high-water	Render
One process (before)	~31 GB	84 s
8 GB card, split text, host copies	20.1 GB	82 s
8 GB card, split text, from disk	16.0 GB	90 s
12 GB card, split text, from disk	14.9 GB at load, 10.4 GB rendering	92 s

The text phase peaks at 12.4 GB. Host and disk latents are bit-identical.

Not done

• Not run on real 8 or 12 GB hardware.
• serve.R and recommend("ltx") don't use the new profile fields yet.
• Disk mode still reads the whole transformer into RAM at load before dropping the resident part; that sets the 14.9 GB load peak.

Tests

Full suite: 1611 passed, 0 failed. New tests cover disk-staging parity (exact output match across offload/onload), the host-RAM profile cases, and the loader and ltx23_text_embeds errors.

Version 0.2.2.18.