All pull requests

Pull requests

• fix(wan): replace deprecated v2.ToTensor() with ToImage() + ToDtype()

  #1648·pseudodream opened Oct 22, 2026·

  ·
  1

• feat: bind Qwen Image 2.1 INT8 Metal runtime

  #1647·q6y6y6 opened Oct 22, 2026Contributor·

  ·

• feat(ar): unify native runners and optimize KV-cached inference

  #1646·gushiqiao opened Oct 22, 2026Contributor·

  ·

• fix: route SekoTalk RS2V HTTP inference through Shot pipeline

  #1645·Murasame-done opened Oct 22, 2026·

  ·

• EdenLive AR: optional decode_to_cpu so GPU memory does not grow with the video length

  #1642·hiroki-abe-58 opened Oct 20, 2026·

  ·

• Wan MXFP8: drop the stored import error's traceback (it keeps the first runner alive)

  #1640·hiroki-abe-58 opened Oct 20, 2026·

  ·

• Fix the SM120 Sage and ConvRot GELU kernels on Windows

  #1639·hiroki-abe-58 opened Oct 20, 2026·

  ·

• Add native BF16 TRELLIS.2 inference

  #1636·for-the-juan opened Oct 20, 2026·

  ·

• Add native BF16 VGGT reconstruction

  #1635·for-the-juan opened Oct 20, 2026·

  ·

• feat(seko): support compile, warmup & metax，musa platform。

  #1617·Watebear opened Oct 15, 2026Collaborator·

  ·
  3

• [ROCm] Run Sol-Attn on gfx1201 (RDNA4) through aiter

  #1599·Shan2L opened Oct 11, 2026·3 tasks·

  ·

• [ROCm] Add aiter_sage_attn attention backend for gfx1201 (RDNA4)

  #1598·Shan2L opened Oct 11, 2026·3 tasks·

  ·

• perf(dreamzero): accelerate NVIDIA Thor inference with FA4 and native kernels

  #1597·Lunemere opened Oct 11, 2026Contributor·

  ·

• [ROCm] Opt-in deterministic FP32 GEMM/conv via rocBLAS

  #1596·Shan2L opened Oct 10, 2026·

  ·
  1

  1

• Set PYTORCH_CUDA_ALLOC_CONF before platform init

  #1595·Shan2L opened Oct 10, 2026·

  ·
  1

  1

• [AMD ROCm] 45:5 INT4/INT8 mixed precision sage attention kernels for MiniMax-H3 on gfx1201

  #1583·jundali77 opened Oct 9, 2026·

  ·

• [MiniMax-H3] Radeon SP: Sol sparse attention (RADEON_CORESW_SOL, _D/_L/_T)

  #1582·Shan2L opened Oct 8, 2026·

  ·

• docs: fix quantization tutorial links in quickstart

  #1578·pratikgx opened Oct 8, 2026·

  ·

• fix(config): stop model bundles from choosing runner and remote-code trust

  #1576·wanjiang263 opened Oct 3, 2026·2 tasks done·

  ·

• fix(minimax_h3): cache ref2av audio-reference timesteps in the fl2av AdaLN cache

  #1574·yinli-systems opened Oct 1, 2026·

  ·

• fix(seedvr): fall back to FlashAttention-3 when flash_attn is missing

  #1573·yinli-systems opened Oct 1, 2026·

  ·

• feat(minimax-h3): Radeon gfx1201 multi-GPU path (Ulysses SP2/4/8 + TP2)

  #1572·Shan2L opened Oct 1, 2026·

  ·
  1

• fix(cosmos3): default trust_remote_code to False in tokenizer load

  #1571·AmirF194 opened Sep 29, 2026·

  ·

• Add reusable contiguous CPU block offload across models and accelerators

  #1565·Chernobyllight opened Sep 28, 2026Contributor·

  ·

• feat(minimax-h3): add Radeon core-sw attention path

  #1563·Shan2L opened Sep 28, 2026·5 tasks·

  ·
  1