Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(39)

Closed

(690)

fix(minimax-h3): honor disabled spatial tiling in the video VAE
#1738Shenrui-Ma opened Oct 12, 2026
OPD Training
#1737Artiprocher opened Oct 12, 2026Collaborator
Support OpenVDN/vdn-minimax-h3
#1736mi804 opened Oct 10, 2026Collaborator
Make cached dataset indexing deterministic
#1729Andzy324 opened Oct 5, 2026
Fix G384 Siglip encoder forwarding and checkpoint loading
#1728Andzy324 opened Oct 5, 2026
Keep local model paths when building training configs
#1727Andzy324 opened Oct 4, 2026
Add Wan DiT LoRA conversion between DiffSynth and Diffusers
#1726Andzy324 opened Oct 4, 2026
fix(joyai): use the pipeline's model names in onload_model_names
#1725fusheng-ji opened Oct 2, 2026
fix(build): make the project resolvable with uv
#1723fusheng-ji opened Oct 1, 2026
fix(wan): raise a clear error for sliding windows with S2V models
#1722fusheng-ji opened Oct 1, 2026
fix(wan): trim vace_video to num_frames
#1721fusheng-ji opened Sep 30, 2026
fix(qwen-image): implement tiled encode and decode in QwenImageVAE
#1720fusheng-ji opened Sep 30, 2026
fix(lora): apply the alpha stored in safetensors metadata
#1719fusheng-ji opened Sep 30, 2026
fix(flux2): cast query and key to the value dtype before attention
#1718fusheng-ji opened Sep 30, 2026
feat(attention): log the selected attention implementation at import
#1715fusheng-ji opened Sep 30, 2026
raise a clear error for incomplete sharded model files
#1711MohammadHijjawi97 opened Sep 25, 2026
V100 t4 16gb
#1710haotemans opened Sep 25, 2026
fix(wan): drop hardcoded device='cuda' in WanToDance music encoder construction
#1708li-lizhe opened Sep 23, 2026
2
fix(loader): raise when no local file matches origin_file_pattern
#1705Linxiushen opened Sep 21, 2026
feat(loader): optionally stage ZeRO-3 checkpoints on CPU
#1700iMountTai opened Sep 21, 2026
fix(checkpoint): preserve keyword arguments with DeepSpeed
#1699iMountTai opened Sep 21, 2026
perf(minimax-h3): unpack QKV with unbind
#1698iMountTai opened Sep 21, 2026
Lora
#1647mi804 opened Aug 27, 2026Collaborator
wan: add dual-GPU model-parallel path for Wan 2.x LoRA training (depends on #1435)
#1436genno-whittlery opened May 11, 20264 of 5 tasks
8
flux2: dual-GPU model-parallel + transformers 5.8 compat + Mistral-on-CPU data_process
#1434genno-whittlery opened May 11, 20264 of 6 tasks
4

Navigation Menu

Repository navigation

All pull requests

fix(minimax-h3): honor disabled spatial tiling in the video VAE

OPD Training

Support OpenVDN/vdn-minimax-h3

Make cached dataset indexing deterministic

Fix G384 Siglip encoder forwarding and checkpoint loading

Keep local model paths when building training configs

Add Wan DiT LoRA conversion between DiffSynth and Diffusers

fix(joyai): use the pipeline's model names in onload_model_names

fix(build): make the project resolvable with uv

fix(wan): raise a clear error for sliding windows with S2V models

fix(wan): trim vace_video to num_frames

fix(qwen-image): implement tiled encode and decode in QwenImageVAE

fix(lora): apply the alpha stored in safetensors metadata

fix(flux2): cast query and key to the value dtype before attention

feat(attention): log the selected attention implementation at import

raise a clear error for incomplete sharded model files

V100 t4 16gb

fix(wan): drop hardcoded `device='cuda'` in WanToDance music encoder construction

fix(loader): raise when no local file matches origin_file_pattern

feat(loader): optionally stage ZeRO-3 checkpoints on CPU

fix(checkpoint): preserve keyword arguments with DeepSpeed

perf(minimax-h3): unpack QKV with unbind

Lora

wan: add dual-GPU model-parallel path for Wan 2.x LoRA training (depends on #1435)

flux2: dual-GPU model-parallel + transformers 5.8 compat + Mistral-on-CPU data_process

Pull requests