All pull requests

Pull requests

• feat(attention): FlyWheel self-attention backend for Wan inference on TPU

  #523·Deepolar opened Nov 15, 2026·

  ·
  3

• [SDY][MaxDiffusion][cleanup] Remove obsolete jax_use_shardy_partitioner=False updates in attention_flax.py

  #522·mbohlool opened Nov 12, 2026Collaborator·

  ·
  1

• Fall back to sharding the height in the WAN VAE.

  #521·ninatu opened Nov 11, 2026Collaborator·

  ·
  4

• perf(wan): batch-sharded cross-attn KV cache and batch-aware VAE mesh

  #519·Perseus14 opened Oct 25, 2026Collaborator·

  ·
  4

  

• Sol Attention Implementation

  #518·jitendra-jalwaniya opened Oct 23, 2026Collaborator·

  ·
  1

• Cast the Wan VAE weights to vae_weights_dtype instead of weights_dtype

  #517·boshjerns opened Oct 22, 2026·

  ·
  2

• Read rope_theta from rope_parameters when loading the Z-Image text encoder

  #516·boshjerns opened Oct 22, 2026·

  ·
  2

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

  #514·boshjerns opened Oct 18, 2026·

  ·
  1

• Keep the full LoRA delta when merging into bfloat16 weights

  #513·boshjerns opened Oct 18, 2026·

  ·
  7

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

  pull ready
  #511·Perseus14 opened Oct 15, 2026Collaborator·

  ·
  21

  

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

  pull ready
  #510·Perseus14 opened Oct 15, 2026Collaborator·

  ·
  7

  

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

  #509·Perseus14 opened Oct 15, 2026Collaborator·

  ·
  17

  

• 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")

  pull ready
  #491·Perseus14 opened Sep 25, 2026Collaborator·

  ·
  18

  

• SVG implementation for LTX 2

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

  ·
  14

• Run FLUX text encoders under Torchax instead of Flax

  #483·olehtika opened Sep 16, 2026·

  ·
  6

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

  #482·olehtika opened Sep 15, 2026·

  ·
  6

• 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