All pull requests

Pull requests

• Add gated_residual_segmented: per-segment gated residual for reference-image tokens (triton, eager)

  #310·IvenHsu01 opened Nov 6, 2026·

  ·
  1

• Add temporal_interleave_upsample_nearest2x: Wan VAE frame interleave + nearest x2 upsample in one pass (triton, eager)

  #309·IvenHsu01 opened Nov 6, 2026·

  ·
  1

• Add depth_to_space3d_residual_add: VAE up-block depth-to-space shortcut + residual add in one pass (triton, eager)

  #308·IvenHsu01 opened Nov 6, 2026·

  ·
  1

• Add adaln_segmented: per-segment adaLN modulation for reference-image tokens (triton, eager)

  #307·IvenHsu01 opened Nov 6, 2026·

  ·
  1

• cuda: initialize the cuBLAS scalar constants statically (capture-safe first scaled_mm_nvfp4)

  #306·PaulFidika opened Nov 6, 2026·

  ·
  1

  2

• Add w3 and w5 quantization formats

  #303·kijai opened Nov 3, 2026Contributor·

  ·
  1

• fix(eager): accumulate int4_linear MMA in fp32 to avoid fp16 overflow on hosts without the fused kernel

  #300·Stanley-MD opened Nov 2, 2026·

  ·
  1

  6

• Add ck.input_act as a dispatched standalone input activation

  #298·LXD-8 opened Nov 2, 2026·

  ·
  18

• Add Metal RoPE kernels on the mps backend

  #297·LXD-8 opened Nov 2, 2026·

  ·
  9

• Add l2_norm_silu: Wan VAE channel L2 norm + SiLU

  #296·klin2024 opened Nov 2, 2026·

  ·
  2

• rms_rope: norm_over_heads (cross-head RMSNorm + RoPE) for Wan

  #295·klin2024 opened Nov 2, 2026·

  ·
  4

• fix: update type hints to typing.List for PyTorch 2.4+ infer_schema c…

  #294·rodrigomateus98 opened Nov 2, 2026·

  ·
  2

• Add linear_w8a16: W8A16 linear for unscaled FP8 weights (triton, eager)

  #293·klin2024 opened Nov 1, 2026·

  ·
  2

• Add act_cat_quantize_fp8: fused activation + concat + per-tensor FP8 quantize (triton, eager)

  #291·klin2024 opened Nov 1, 2026·

  ·
  2

• hip: port connected_components and edge_collapse_checks

  #290·klin2024 opened Oct 31, 2026·

  ·
  2

• Add linear_geglu / linear_swiglu: gated FFN up projections in one GEMM (triton, eager)

  #288·klin2024 opened Oct 30, 2026·

  ·
  12

• Add linear_gelu: linear + tanh-GELU epilogue (triton, eager)

  #287·klin2024 opened Oct 30, 2026·

  ·
  7

• [Weight adapter] Add fused weight-adapter ops (LoRA, LoHa, LoKr, OFT, DoRA, IA3)

  #284·KohakuBlueleaf opened Oct 28, 2026·

  ·
  16

• sol_attn: write the ragged last block length on the device

  #283·PaulFidika opened Oct 28, 2026·

  ·
  3

• Accept DLPack-normalized size-1 strides in flash decode GQA

  #279·potoior opened Oct 26, 2026·

  ·
  1

  4

• Add FP16 tensor-core fallback for int8_linear on devices without INT8 tensor cores

  #277·seningpd opened Oct 24, 2026·

  ·
  15

• Issue HIP INT8 attention K/V staging loads together

  #276·qweqweewqe7-create opened Oct 23, 2026·

  ·
  3

• Add the gated input_act pairs to the mps backend

  #275·LXD-8 opened Oct 23, 2026·

  ·
  25

• Decline fp16-accumulate CUTLASS kernels on sub-sm_80 GPUs

  #272·potoior opened Oct 23, 2026·

  ·
  1

  5

• sol_attn: caller-supplied block lists (block-sparse attention)

  #271·kijai opened Oct 22, 2026Contributor·

  ·
  1