The dynamo exporter in PyTorch 2.13 produces complex shape formulas
(Min, TruncToInt, CEIL_DIV) that TensorRT cannot resolve for dynamic
shapes. The legacy TorchScript exporter produces clean symbolic dims
(batch, height, width). Also skip onnxsim which folds constants based
on dummy input size, breaking TRT large profile builds.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The previous 512x512 dummy input caused onnxsim to fix intermediate
dimensions to 512px, which broke TensorRT engine build for the large
resolution profile (720-1920px). Using 1080x1080 prevents this.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
- Add RIFE v4.25 architecture support in rife_arch.py (Head with
16/4 channels, 5 IFBlocks with c=192/128/96/64/32, scale_list
[16,8,4,2,1], no ensemble)
- Add rife425.pth to export_onnx.py with HF fallback URL
- Update load_rife_config.json: dropdown now only shows rife425 and
rife49 (removed rife47/48/417/426/sudo_rife4 that never worked)
- Set rife425 as default
- Make TORCH_DEVICE fallback to CPU for ONNX export on non-CUDA
- Pin polygraphy==0.49.26 for TensorRT 10.15.1 compatibility
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
rife426.pth and possibly other new checkpoints contain extra modules
used during training (teacher network, caltime scheduler) that are not
part of the IFNet inference architecture. These cause
load_state_dict to fail with "Unexpected key(s)" errors.
Now export_onnx.py filters the state_dict to keep only keys that match
the model architecture before loading, with strict=True to catch any
real missing keys.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Update rife_arch.py with support for new architectures:
- v4.17: Head_417 encoder, IFBlock with 16 extra channels
- v4.26: 5 IFBlocks (block4 extra), lastconv 4*13 channels,
feat_out return, no ensemble, 5 scale levels
- Head/Head_417 classes for v4.17/v4.26 encoders
- warp() fix: MPS padding_mode, dtype mismatch, get_torch_device()
Update export_onnx.py:
- Add rife417.pth and rife426.pth to CKPT_NAME_VER_DICT
- Add CKPT_FALLBACK_URLS with HuggingFace mirrors for all models
- Add export calls for rife49, rife417, rife426, sudo_rife4
Update load_rife_config.json with 3 new ONNX model options.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Ported and adapted from Fannovel16/ComfyUI-Frame-Interpolation:
- gc.collect() after soft_empty_cache() to reduce RAM fragmentation
- Explicit float32 cast in postprocess_frames for fp16 engine safety
- InterpolationStateList class + MakeInterpolationStateList node to
skip/keep specific frame pairs during interpolation
- Per-pair multiplier schedule (comma-separated list) overriding uniform
multiplier, with automatic padding to match the number of pairs
- Automatic tiling with cosine feathering when frame dimensions exceed
the TensorRT engine optimization profile (fixes 1440x2560 with max=1920)
- Padding with replicate mode when frame dimensions are below profile min
- get_input_profile_bounds() on Engine to retrieve min/opt/max shapes
- Batch inference: process multiple (pair, timestep) tasks per GPU call
with dynamic batch axis in ONNX export and engine profile
- Backward compatible: all new features are opt-in via optional inputs
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
huchukato