All pull requests

Pull requests

• Fix the Wan VAE attention block, which never attends over spatial positions

  #514·boshjerns opened Oct 18, 2026·

  ·

• Keep the full LoRA delta when merging into bfloat16 weights

  #513·boshjerns opened Oct 18, 2026·

  ·
  3

• perf(attention): shard-major Ulysses output all-to-all (wan_ulysses_out_a2a="chunked"/"shard_major")

  #511·Perseus14 opened Oct 15, 2026Collaborator·9/10·

  ·
  10

• feat(wan): token-sharded patch embedding (wan_patch_embed_mode="tokens")

  #510·Perseus14 opened Oct 15, 2026Collaborator·8/10·

  ·
  2

• feat(wan): short-KV text cross-attention Pallas kernel (wan_cross_attn_kernel="pallas")

  #509·Perseus14 opened Oct 15, 2026Collaborator·7/10·

  ·
  4

• perf(wan): XLA-fused RMSNorm + RoPE producer for self-attention Q/K

  #508·Perseus14 opened Oct 15, 2026Collaborator·4/10·

  ·
  6

• fix(attention): dot-product fallback layout for 4-D inputs; tile_size_grid_search iters/bkv_compute guards

  #507·Perseus14 opened Oct 15, 2026Collaborator·2/10·

  ·
  6

• chore: migrate gsutil usage to gcloud storage

  #506·gurusai-voleti opened Oct 15, 2026Contributor·

  ·

• style: fix pyink formatting on main (2 files)

  #503·Perseus14 opened Oct 13, 2026Collaborator·

  ·
  1

  

• docs: add Cluster Toolkit (gcluster) guide and mark XPK as deprecated

  #501·Perseus14 opened Oct 3, 2026Collaborator·

  ·
  6

• chore: repo hygiene, CI consistency and documentation fixes

  #500·Perseus14 opened Oct 3, 2026Collaborator·

  ·
  6

• perf(wan): pad sequence to 128-lane multiple for self-attention (wan_seq_pad="lane")

  #491·Perseus14 opened Sep 25, 2026Collaborator·10/10·

  ·
  5

• SVG implementation for LTX 2

  #489·jitendra-jalwaniya opened Sep 24, 2026Collaborator·

  ·
  14

• feat(wan): fused RMSNorm+RoPE Pallas producer; Wan switches in config, not env

  #488·Perseus14 opened Sep 22, 2026Collaborator·6/10·

  ·
  12

• Run FLUX text encoders under Torchax instead of Flax

  #483·olehtika opened Sep 16, 2026·

  ·
  3

• Fix three FLUX regressions from the Flux.2-Klein onboarding

  #482·olehtika opened Sep 15, 2026·

  ·
  5

• feat(wan): fast serving with persistent AOT caching and tuned inference recipe

  pull ready
  #479·Perseus14 opened Sep 13, 2026Collaborator·5/10·

  ·
  23

  

• feat(attention): 2D Ulysses+Ring custom attention with fixed-m (pre-a2a norm reduction, unpadded ring K/V)

  #478·Perseus14 opened Sep 13, 2026Collaborator·3/10·

  ·
  20

  

• feat(attention): fixed-m splash attention kernel with dynamic bounds and safety fallbacks

  pull ready
  #477·Perseus14 opened Sep 13, 2026Collaborator·1/10·

  ·
  39

  

• fix(deps): skip nvidia-cuda-cccl on macOS (no wheels available)

  pull ready
  #476·harkanwalbedi opened Sep 11, 2026·

  ·
  1

  3

• fix: prevent silent double-application of LoRA weights

  #473·WangXukang-cypher opened Sep 4, 2026·2 of 3 tasks·

  ·
  12

• Feat: internal 1d and 2d ring attention kernel

  #472·eltsai opened Sep 3, 2026Collaborator·

  ·
  4

  

• Minor improvement to Flux models

  #457·Perseus14 opened Aug 6, 2026Collaborator·

  ·
  3

• Add heads_per_tile (multi-head-per-tile) support to Ulysses ring attention

  pull ready
  #444·edgexyz opened Jul 12, 2026Contributor·

  ·
  12

• Tf records from gcs

  #213·entrpn opened Jul 28, 2025Collaborator·

  ·
  1