[TroyHernandez]

FLUX 3 Image Klein on 8 and 12 GB cards. flux3_load_pipeline() decides what streams based on the card size (vram_gb, detected by default). Version 0.2.2.17.

Streaming

• stream_text (default below 10 GB): the Qwen3-VL layers stay in pinned host memory and move to the GPU one at a time during the encode, which takes about 1 s. The whole encoder would need about 9 GB at once.
• resident_gb (default vram_gb - 6.5 below 12 GB): that many GB of fp8 transformer weights stay on the GPU. The remaining linears copy in from pinned memory inside each forward, reusing the existing host path in ltx23_fp8_linear.
• 12 GB and up: nothing streams, same as before.

Garbage collection

• FLUX 3 now runs a minor R gc after every transformer block, including the per-mode blocks. Previously it ran a full gc after the joint blocks only.
• The VAE encoder and decoder blocks run a minor gc after each resnet when diffuseR.block_gc is on.
• Streaming mode sets its own allocator gc gates: reserved 0.6, allocated 0.8. torch for R runs R's gc only on an allocator miss, and skips it entirely while reserved memory is below reserved_rate × total. The 12 GB-footprint default works out to 0.92 on an 8 GB card, so no gc ran there even when an allocation failed.

Measured (RTX 5060 Ti 16 GB, fp8; smaller cards simulated by holding memory in an unused tensor)

Card	1024² text-to-image	1 MP edit
8 GB (simulated)	30 s	56 s
12 GB (simulated, no streaming)	27 s	53 s
16 GB	27 s, peak 9.1 GB (was 44 s, 11.7 GB)	53 s, peak 10.7 GB (was 50 s, 13.8 GB)

• The 4 MP edit is unchanged at 285 s and a 14.1 GB peak.
• The streamed 8 GB image matches the resident one at 58.7 dB PSNR.
• tinytest: 1561 passed, 0 failed (21 new in test_flux3_stream.R: the plan, Qwen layer streaming parity, and the fp8 split census and parity).

Not in this PR

• A run on real 8 and 12 GB hardware.
• A recommend() entry for FLUX 3.
• Applying the minor-gc change to klein and Z-Image.