# U1.5: GPU memory, system RAM, and 8-step generation time

English | 简体中文

These measurements cover [U1.5](u1.5_model.md) text-to-image, editing, and one-image interleaved generation. The runs used the 8-step V2 LoRA, merged into the base weights before generation; no separate adapter weights were active during the measured calls. The [model guide](u1.5_model.md#8-step-v2-lora) has the download and run commands. ## RTX 5090 measurements All nine task and memory-mode combinations completed on a GeForce RTX 5090 (32,607 MiB). Each ran in a fresh process: seed 0 warmed the loaded model, then seeds 1–3 were timed. The table gives the median generation time (and range) for those three calls. Loading, LoRA merging, and file saving are excluded. GPU figures are the highest PyTorch allocated and reserved peaks across the three measured calls, not whole-card use. [Per-call data, image hashes, test versions, and fast-mode budget settings](data/u1.5_5090_runs.json) are available separately. | Task | Mode | Generation after warmup (s) | GPU allocated / reserved (GiB) | | --- | --- | ---: | ---: | | Text-to-image | `fast` | 4.469 (4.469–4.476) | 16.60 / 18.46 | | Text-to-image | `balanced` | 4.461 (4.448–4.465) | 3.68 / 8.65 | | Text-to-image | `low` | 6.740 (6.726–6.740) | 2.96 / 3.56 | | Editing | `fast` | 5.709 (5.706–5.725) | 17.95 / 24.98 | | Editing | `balanced` | 5.730 (5.724–5.743) | 10.20 / 15.17 | | Editing | `low` | 8.223 (8.222–8.238) | 9.48 / 10.08 | | Interleaved | `fast` | 23.989 (23.709–24.570) | 19.17 / 22.39 | | Interleaved | `balanced` | 24.269 (23.985–24.883) | 6.99 / 12.48 | | Interleaved | `low` | 28.509 (28.209–29.134) | 6.27 / 7.43 | The runs used BF16, SDPA, eight steps, CFG 1.0, and batch size 1. Text-to-image and editing produced 2048×2048 images; editing used [this input](../examples/editing/data/images/1.webp). Interleaved generation used 1536×1536, `think_mode=False`, and `max_images=1`, so these results do not cover longer sequences. The same prompts were used across modes. All 36 calls produced decodable RGB PNGs; equal task-and-seed pairs had matching PNG hashes across modes. Interleaved timing includes text generation. The runs were recorded on November 13, 2026, without an allocator cap. Process CPU RSS peaked around 52.2 GiB with `balanced` and `low`; plan for 64 GB of system RAM to leave room for the OS and other processes. This is not a measured minimum RAM requirement. For these inputs, `balanced` and `fast` took about the same time, but `balanced` used less GPU memory. `low` used less memory and took longer. PyTorch peaks do not establish minimum card capacity. ## H100 allocator-cap runs During the H100 80 GB tests, other processes occupied roughly 37–40 GiB of GPU memory. The launcher set PyTorch allocator caps of 24, 32, or 48 GiB with `torch.cuda.set_per_process_memory_fraction` and also ran without a cap. Each combination used a fresh process with seed 0, BF16, SDPA, eight steps, and CFG 1.0; tasks and resolutions match the 5090 runs above. The table groups caps with identical measured peaks. Every listed cap and task was run separately. Values are peak PyTorch **reserved** GiB during generation. | H100 allocator cap | Mode | Text-to-image | Editing | Interleaved | | --- | --- | ---: | ---: | ---: | | 24 and 32 GiB | `balanced` | 7.0 | 14.0 | 11.7 | | 24 and 32 GiB | `low` | 3.0 | 10.0 | 7.4 | | 48 GiB and uncapped | `full` | 34.7 | 40.3 | 37.6 | All listed runs completed. `full` reserved 33.0 GiB immediately after loading, so it was not tested under the 24 or 32 GiB caps. Process CPU RSS was about 52 GiB with offload and 14 GiB with `full`. **Versions:** H100 driver 550.90.07, Python 3.11.13, PyTorch 2.8.0+cu128, Transformers 5.15.0; base revision `a2462cc1b549bace2c0896143bbbcbb3fa217c76` and V2 LoRA revision `f33b8fe0216e6f10e5aed58ebbfdd05f02826732`. The H100 data are single calls without warmup and use a different base revision from the 5090 runs. For memory-mode behavior, see the [examples](../examples/README.md#single-gpu-memory-modes). Leave room beyond PyTorch's reported peaks for the CUDA context, other allocations, and other processes.