SergioShu/diffueraser-trt-sm120

DiffuEraser TensorRT engines (sm120)

Serialized Torch-TensorRT programs (torch.export + torch_tensorrt.dynamo) of DiffuEraser's networks, built for NVIDIA Blackwell GPUs of compute capability 12.0 with TensorRT 11.1.0. They load only on that architecture and TensorRT version.

• trt11.1.0-hc-any/: hardware-compatible, and both sides of the canvas vary (384..960 each, a multiple of 64): one engine per network for both orientations, as fast as one per orientation and half the size. Use these.
• trt11.1.0-hc/: hardware-compatible, one engine per orientation (tested on RTX PRO 6000 and RTX PRO 4500).
• trt11.1.0/: tuned to the RTX PRO 6000 (188 SMs). On a GPU with fewer SMs (RTX PRO 4500) TensorRT warns that deadlocking is likely and the diffusion networks hang.

File	Network	Size
trt11.1.0{,-hc}/sm120_brushnet_{land,port}.ep	DiffuEraser BrushNet	2.4 / 2.6 GB
trt11.1.0{,-hc}/sm120_unet_{land,port}.ep	DiffuEraser motion UNet, PCM 2-step LoRA fused	5.25 / 5.6 GB
trt11.1.0{,-hc}/sm120_vae-encode_{land,port}.ep	sd-vae-ft-mse encoder	0.11 / 0.12 GB
trt11.1.0{,-hc}/sm120_vae-decode_{land,port}.ep	sd-vae-ft-mse decoder	0.14 / 0.15 GB

Shapes are dynamic along one side of a 960-pixel canvas: land has a width of 960 and a height of 384..960 (a multiple of 64), port the other way round; any takes both. The diffusion networks take 22 frames.

Licenses

These are derivatives of models under these licenses:

• Stable Diffusion 1.5 (the base of BrushNet and the UNet): CreativeML Open RAIL-M, see LICENSE-CreativeML-OpenRAIL-M. Its use-based restrictions (paragraph 5 and Attachment A) apply to these engines and to every user of them.
• DiffuEraser: Apache License 2.0, see LICENSE-Apache-2.0.
• sd-vae-ft-mse (Stability AI): MIT.