Repository navigation

All pull requests

[ROCm] Run Sol-Attn on gfx1201 (RDNA4) through aiter
#1599Shan2L opened Oct 11, 20263 tasks
[ROCm] Add aiter_sage_attn attention backend for gfx1201 (RDNA4)
#1598Shan2L opened Oct 11, 20263 tasks
perf(dreamzero): accelerate NVIDIA Thor inference with FA4 and native kernels
#1597Lunemere opened Oct 11, 2026Contributor
[ROCm] Opt-in deterministic FP32 GEMM/conv via rocBLAS
#1596Shan2L opened Oct 10, 2026
1
Set PYTORCH_CUDA_ALLOC_CONF before platform init
#1595Shan2L opened Oct 10, 2026
1
[AMD ROCm] 45:5 INT4/INT8 mixed precision sage attention kernels for MiniMax-H3 on gfx1201
#1583jundali77 opened Oct 9, 2026
[MiniMax-H3] Radeon SP: Sol sparse attention (RADEON_CORESW_SOL, _D/_L/_T)
#1582Shan2L opened Oct 8, 2026
docs: fix quantization tutorial links in quickstart
#1578pratikgx opened Oct 8, 2026
fix(config): stop model bundles from choosing runner and remote-code trust
#1576wanjiang263 opened Oct 3, 20262 tasks done
fix(minimax_h3): cache ref2av audio-reference timesteps in the fl2av AdaLN cache
#1574yinli-systems opened Oct 1, 2026
fix(seedvr): fall back to FlashAttention-3 when flash_attn is missing
#1573yinli-systems opened Oct 1, 2026
feat(minimax-h3): Radeon gfx1201 multi-GPU path (Ulysses SP2/4/8 + TP2)
#1572Shan2L opened Oct 1, 2026
1
fix(cosmos3): default trust_remote_code to False in tokenizer load
#1571AmirF194 opened Sep 29, 2026
Add reusable contiguous CPU block offload across models and accelerators
#1565Chernobyllight opened Sep 28, 2026Contributor
feat(minimax-h3): add Radeon core-sw attention path
#1563Shan2L opened Sep 28, 20265 tasks
1
add lightvae for h3
#1558gushiqiao opened Sep 24, 2026Contributor
Fix NPU Weight Corruption During CPU Offload and Simplify Layout Configurations
#1556Chernobyllight opened Sep 24, 2026Contributor
feat: support pipefusion for MiniMax H3
#1554Qin-sx opened Sep 23, 2026Contributor
Support H3-world inference
#1550Musisoul opened Sep 22, 2026Collaborator
feat(minimax-h3): add a single-GPU AMD ROCm recipe (aiter attention, 4-step Turbo LoRA), measured on MI355X
#1534stayinalive181 opened Sep 19, 2026
1
feat(musa): add MiniMax H3 sparse attention
#1529Watebear opened Sep 18, 2026Collaborator
feat(minimax-h3): add VDN inference and parallel
#1528Watebear opened Sep 17, 2026Collaborator
Unify CPU Block Offload Launchers and Host/NUMA Weight Sharing
#1527Chernobyllight opened Sep 17, 2026Contributor
feat(offload): add CPU block offload and host/NUMA weight sharing across Hunyuan Image 3.0, MiniMax-H3, Qwen-Image, and Wan
#1524Chernobyllight opened Sep 17, 2026Contributor
Ar ws
#1512huochaitiantang opened Sep 11, 2026Collaborator

Pull requests