19825fa9fa
Release v2.5.13: Fix triton import, OOM on long videos, macOS watermark
Adrien Toupet2025-11-30 09:01:53 -05:00
021fc7b70f
Fix triton.ops import error by using local VAE types
Adrien Toupet2025-11-30 08:27:01 -05:00
b0ac880a4f
Fix OOM crash on float32 conversion for long videos. Gracefully fallback to native dtype if insufficient memory. Fixes#299Adrien Toupet2025-11-30 08:03:07 -05:00
8dd0c3061a
feat: Add CUDNN attention backend support with PyTorch 2.3+ API - Use torch.nn.attention.sdpa_kernel() (new API) with CUDNN_ATTENTION backend - Fallback to torch.backends.cuda.sdp_kernel() when not available Thanks to @eadwu for the original PR #317Adrien Toupet2025-11-27 15:19:55 -05:00
79e7f41216
Revert "Fix: MPS allocator error in model.to() transfer (#305)"
Adrien Toupet2025-11-27 14:52:12 -05:00
9424020687
Fix: MPS allocator error in model.to() transfer (#305) Add fallback to individual parameter movement when bulk model.to(mps) fails with allocator errors. Complements safetensors loading fix.
Adrien Toupet2025-11-14 12:06:13 -05:00
9715d3e37a
Fix: torch.mps AttributeError on Windows Add defensive checks for torch.mps.is_available() to handle PyTorch versions where the method doesn't exist on non-Mac platforms. Resolves AttributeError: module 'torch.mps' has no attribute 'is_available'
Adrien Toupet2025-11-07 23:37:35 -05:00
806bb94df0
feat: unify and improve tooltip documentation across CLI and ComfyUI nodes
Adrien Toupet2025-11-05 15:35:22 -05:00
326489d94c
fix CLI: move Debug import after CUDA allocator config to fix batch processing errors
Adrien Toupet2025-11-05 14:27:06 -05:00
9b79254c39
refactor(cli): improvements and bug fixes + 3b-Q8_0.gguf support
Adrien Toupet2025-11-05 00:28:19 -05:00
3725c1061d
refactor: centralize dimension computation and logging for CLI/ComfyUI
Adrien Toupet2025-11-04 17:01:01 -05:00
1691e657b4
fix(cli): add CUDA device validation before torch initialization - Validate --cuda_device arguments early in pre-parsing phase - Check device IDs exist and are within available GPU range - Fail fast with clear error messages showing available devices
Adrien Toupet2025-11-04 15:52:05 -05:00
32a049dfd9
feat: Add CLI model caching for multi-file processing and unify device handling
Adrien Toupet2025-11-04 15:12:28 -05:00
ad020d3803
feat(cli): improve UX with auto-format detection, FPS tracking, and consistent messaging with ComfyUI implementation
Adrien Toupet2025-11-04 11:54:57 -05:00
ad1d775eaa
fix: tile debug overlay now applies to all batches and adds overlay warning
Adrien Toupet2025-11-04 09:06:13 -05:00
9268346388
feat: CLI Add batch processing, fix multiprocessing issues, and unify model paths
Adrien Toupet2025-11-04 00:12:13 -05:00
77cb6ff684
refactor(cli): inference_cli to match ComfyUI integration
Adrien Toupet2025-10-28 01:01:38 -04:00
5772255aed
refactor: split generation.py and model_manager.py into 4 focused modules
Adrien Toupet2025-10-28 00:04:33 -04:00
e8376ddd6d
refactor: split generation.py and model_manager.py into 4 focused modules
Adrien Toupet2025-10-28 00:04:08 -04:00
f182de79fa
feat: Add temporal_overlap & prepend_frames to ComfyUI with shared logic
Adrien Toupet2025-10-27 21:57:38 -04:00
3a4a4900df
feat: optimize SeedVR2 memory management and color ops
Adrien Toupet2025-10-27 14:58:20 -04:00
003122ebcd
feat: implement lossless arbitrary resolution with padding - replace DivisibleCrop with DivisiblePad to eliminate data loss, track true dimensions for post-processing trim, change default resolution to 1080p with step=2 for flexibility
Adrien Toupet2025-10-27 01:14:57 -04:00
b0f03e54e5
fix: make FSDP imports conditional for AMD ROCm compatibility
Adrien Toupet2025-10-26 23:49:52 -04:00
Adrien Toupet