feat(multi-gpu): offload text encoders to idle GPUs

[lstein]

⚠️ Merge order

This PR depends on #9263 and targets its branch (lstein/feat/multi-gpu), not main. It should be reviewed and merged only after #9263 has been reviewed, accepted, and merged. Once #9263 lands, this PR's base can be retargeted to main.

It was inspired by #9310 (split-GPU text encoder) and supersedes that PR — it delivers the same idea, but reworked to compose with the multi-GPU parallel-generation architecture from #9263 and to reuse that branch's existing per-device caches, device-aware VRAM accounting, and shared CPU-weights store rather than re-adding them.

Summary

On a multi-GPU machine, #9263 runs one generation session per GPU. When fewer sessions are running than there are GPUs, the spare GPUs sit idle. This PR uses that idle capacity: a session's text/prompt encoder runs on a currently-idle GPU instead of the GPU running its denoise pipeline.

• Avoids evicting the denoise model from VRAM just to make room for the encoder.
• Lets a cached encoder be reused across generations, making repeated single-session generations noticeably smoother.
• Purely a placement optimization — generated images are unchanged.

Controlled by a new offload_text_encoders_to_idle_gpus setting (default on). With a single device, or under full multi-GPU load (no idle GPU), encoders run on the session's own GPU exactly as before.

How it works

• GENERATION_DEVICE_POOL arbiter (backend/util/device_pool.py) with a per-device exclusive-use lock. A native session blocking-acquires its own GPU's lock for the whole run; an encoder node try-borrows an idle GPU's lock for the duration of that node. A borrowed encoder and a native session are therefore mutually exclusive on a GPU, and the design is deadlock-free (borrows are non-blocking try-acquires; a session only ever blocks on its own device).
• DefaultSessionRunner temporarily re-pins the worker thread to the borrowed GPU for the whole encoder node. The encoder loads into and runs on that GPU; its conditioning is stored on the CPU (as encoder nodes already do) and the denoiser picks it up on its own GPU afterward — so the cross-GPU handoff needs no node changes.
• Per-node opt-in via @invocation(idle_gpu_offloadable=True), mirroring the existing bottleneck ClassVar marker (no API-schema impact). Applied to the text/prompt encoder nodes: compel (+ SDXL/refiner), flux_text_encoder, sd3_text_encoder, qwen_image_text_encoder, anima_text_encoder, cogview4_text_encoder, flux2_klein_text_encoder, z_image_text_encoder, and flux_redux.

Why the per-device lock

An earlier iteration routed the encoder into the idle GPU's cache without exclusivity. Because two sessions using the same model/prompt resolve to the same encoder cache key, they ended up sharing one model object and running concurrent forward passes + in-place LoRA patching on it — producing garbled images. The per-device lock makes a borrow and a native session mutually exclusive on a GPU, which fixes this; prevent_auto_evict from #9310 is intentionally not ported, so a borrowed encoder yields its GPU's VRAM (via normal LRU) the moment that GPU is claimed for a real session.

Tests

• tests/backend/util/test_device_pool.py — arbiter lock semantics (borrow exclusion, session/borrow mutual exclusion, startup-race ordering, deterministic selection) plus a multi-threaded regression test asserting a GPU is never used by a session and a borrow at the same time.
• tests/app/services/session_processor/test_encoder_offload.py — the runner offload context manager (re-pin/restore, no-offload-when-busy, flag-off, restore-on-exception), the idle_gpu_offloadable marker wiring on real nodes, and a two-worker concurrency regression exercising the real offload path.

Docs

• configuration/invokeai-yaml.mdx — documents offload_text_encoders_to_idle_gpus.
• development/Guides/creating-nodes.mdx — explains how (and when) a node should set idle_gpu_offloadable=True.

Verification

• Full backend test suite: 2138 passed / 127 skipped. (One unrelated failure, test_torch_cuda_allocator.py::test_configure_torch_cuda_allocator_configures_backend, requires a working CUDA cudaMallocAsync allocator and fails on a CPU-only box; it touches none of this PR's code.)
• ruff check and ruff format --check clean; openapi.json / schema.ts regenerated (only the new config field).
• Manually verified on a dual-GPU machine: single-session offload, parallel sessions with the same model, and parallel sessions with two different models/encoders all produce correct images.

🤖 Generated with Claude Code