| # FLUX.1 offloading versus no-offloading VRAM matrix |
|
|
| Fresh ComfyUI process per cell; total physical-GPU memory sampled every 20 ms under the exclusive GPU lock. |
| `--highvram` is the no-model-offloading control; it is not `--gpu-only`. |
|
|
| | Policy | Pipeline | Status | Cold peak MiB | Warm peak MiB | Maximum peak MiB | |
| |---|---|---|---:|---:|---:| |
| | lowvram | Original BF16 | success | 24091 | 23650 | 24091 | |
| | lowvram | Partial INT8 | success | 20351 | 20016 | 20351 | |
| | lowvram | Partial INT4 | success | 16642 | 16604 | 16642 | |
| | lowvram | Paper Mixed | success | 12066 | 12050 | 12066 | |
| | lowvram | Paper W4A4 | success | 10765 | 10797 | 10797 | |
| | lowvram | Whole W8A8 | success | 16298 | 16298 | 16298 | |
| | lowvram | Whole W8A8 + INT8 T5 | success | 16265 | 16265 | 16265 | |
| | highvram | Original BF16 | capacity_failure | 23939 | | 23939 | |
| | highvram | Partial INT8 | success | 21342 | 21342 | 21342 | |
| | highvram | Partial INT4 | success | 21438 | 21438 | 21438 | |
| | highvram | Paper Mixed | success | 20062 | 20158 | 20158 | |
| | highvram | Paper W4A4 | success | 18942 | 18974 | 18974 | |
| | highvram | Whole W8A8 | success | 21342 | 21342 | 21342 | |
| | highvram | Whole W8A8 + INT8 T5 | success | 19966 | 19998 | 19998 | |
| |
| GPU UUID: `GPU-27ff99d6-d77d-2226-a4b5-7a11b8ef6bdd` |
| |
| Settings: 1024×1024, batch 1, 20 steps, Euler/simple, CFG 1.0, Flux guidance 3.5. |
| |
| Source benchmark state: measurement complete, run `20260719T215033Z_1143248`. The benchmark used total physical-GPU memory because WSL did not expose per-process memory for compute rows. |
|
|
