- LTX Video VAE detection (128 channels, 32x spatial compression, 8x temporal)
- z-image / z-image-turbo pixel-space VAE passthrough (no decode needed)
- fp16 default on all AMD GPUs (Strix Halo, RDNA 2/3, CDNA)
- Per-architecture tuning with automatic detection (gfx1151/gfx1100/gfx1030/CDNA)
- Temporal tiling for long videos: overlapping chunks with linear blend,
zero frame loss, progress bar in UI (disabled by default)
- Chunked IO bypass for LTX — fixes OOM with max_split_size_mb caps
- Fixed memory estimation: correct for all dims (including 5D temporal) and dtypes
- APU-aware memory management (psutil for unified memory on Strix Halo)
- Tile size floor for high-compression VAEs (min 64 latent pixels)
- 25 unit tests covering all new features
- Added optional in-memory cache to ROCmOptimizedCheckpointLoader to avoid redundant disk loads during repeated API runs.
- Implemented cache eviction strategy to manage memory efficiently on gfx1151 architecture.
- Updated INPUT_TYPES to include parameters for cache usage and forced reload.
- Added integration and unit tests to verify cache behavior and functionality.
- Documented caching mechanism in CHECKPOINT_CACHE_API.md and updated LOADER_NODES_GUIDE.md for clarity.
- Added instrumentation decorator to all ROCM nodes
- Fixed instrumentation to work with ComfyUI node structure
- Collected baseline performance data for ROCMOptimizedVAEDecode
- Created test infrastructure for performance optimization
- Baseline: avg 0.427ms, median 0.199ms, 100% fast executions
iGavroche