All pull requests

Pull requests

• fix(triton): widen blocked RMSNorm row index

  #1482·arcusbuilds opened Sep 22, 2026·

  ·
  1

• docs: add Halo training cookbook

  #1480·advpropsys opened Sep 21, 2026·

  ·

• bench(flce): add configurable chunk-memory sweep

  #1479·fuxiangdu opened Sep 21, 2026·

  ·

• Fix ignored-token masking in cosine distillation

  #1478·taking-lying-flat opened Sep 21, 2026·

  ·

• Fix Falcon-H1 output scaling in LCE forward

  #1477·taking-lying-flat opened Sep 21, 2026·

  ·

• refactor(kl_div): route fused-linear KL inner kernel through the dispatcher

  #1476·Yulong-Cauli opened Sep 19, 2026·

  ·

• fix(monkey_patch): patch Q/K norms on existing Qwen models

  #1470·luca-888 opened Sep 16, 2026·

  ·

• fix(cutedsl): fuse Gemma RMSNorm weight offset in FP32

  #1462·luca-888 opened Sep 11, 2026·

  ·

• fix(grpo): respect independent token and mask strides

  #1461·LOGO127 opened Sep 11, 2026·

  ·
  1

• fix(mlp): preserve packed activations across backward replay

  #1460·arde171 opened Sep 11, 2026·

  ·

• fix(cutile): respect native shape and backward contracts

  #1459·arde171 opened Sep 11, 2026·

  ·

• test(cutedsl): gap-op dispatcher tests + fused_add_rms_norm 3-D backward + shape fallbacks

  #1457·arde171 opened Sep 10, 2026·

  ·
  2

• feat(cutile): register six backends and harden CE/SwiGLU contracts

  #1456·arde171 opened Sep 10, 2026·

  ·
  5

• feat(cutile): chunk FLCE with Triton sizing and cuBLAS GEMMs

  #1455·vaibhavjindal opened Sep 10, 2026·

  ·

• feat(ops): support partial RoPE in Triton kernel

  #1451·piyushumate opened Sep 9, 2026·

  ·
  1

• feat(cutedsl): add CuTe-DSL Newton-Schulz operator and LigerMuon optimizer

  #1449·piyushumate opened Sep 9, 2026·

  ·

• [WIP] Add Qwen4Exp text model support

  #1447·alifurkanstahl opened Sep 9, 2026·

  ·

• fix(llama4): honour skip_logits in lce_forward

  #1445·yupengtang opened Sep 7, 2026·

  ·

• fix(monkey_patch): let the norm flags cover the whole vision tower

  #1443·yupengtang opened Sep 7, 2026·

  ·

• [Draft] Apple Silicon (MPS) support via triton-ext

  #1442·zheliuyu opened Sep 4, 2026·

  ·
  9

• perf(norm): skip dW/dB calculation when parameters are frozen (#1067)

  #1440·piyushumate opened Sep 4, 2026·

  ·
  1

• feat(flce): expose direct chunk-size control via LIGER_FLCE_CHUNK_SIZE

  #1439·jin-five opened Sep 4, 2026·

  ·
  5

• fix: request float32 logits from the fused JSD projection GEMM

  #1433·KyleMylonakisProtopia opened Sep 1, 2026·

  ·
  1

  1

• perf(flce): write the weight-gradient product directly into grad_weight

  #1429·ananthv26 opened Sep 1, 2026·

  ·

• fix(peft): patch the LayerNorm copies, not the ModulesToSaveWrapper itself

  #1425·yurekami opened Aug 30, 2026·

  ·
  1