All pull requests

Pull requests

• [CuTe DSL] Add native Blackwell fused scaled cross entropy

  #1597·yueyiming2009 opened Nov 4, 2026Collaborator·1 of 3 tasks·

  ·

• fix(monkey_patch): warn on kwargs dropped by the dispatcher

  #1596·ASHMEET555 opened Nov 3, 2026·3 tasks·

  ·
  1

• ops: validate targets in LigerVocabParallelCEFunction

  #1593·ASHMEET555 opened Nov 3, 2026·3 tasks·

  ·
  1

• test(rope): attach parametrize to the Qwen3.5 vision RoPE test

  #1592·0z5a opened Nov 3, 2026·1 of 3 tasks·

  ·

• fix(cutile): guard the JSD mixture log against fp32 underflow

  #1591·0z5a opened Nov 3, 2026·1 of 3 tasks·

  ·

• fix(utils): skip device_context guard under torch.compile

  #1590·0z5a opened Nov 3, 2026·1 of 3 tasks·

  ·

• fix(flce): preserve head gradients with frozen inputs

  #1588·taking-lying-flat opened Nov 2, 2026Contributor·

  ·

• Restore torch cross-entropy state in revert_liger_kernel helpers

  #1587·reginaldalfret opened Nov 1, 2026·

  ·
  1

• feat(chunked_loss): add fused linear IRPO (Iterative Reasoning Preference Optimization) loss

  #1586·Yulong-Cauli opened Nov 1, 2026Contributor·

  ·

• feat(chunked_loss): add fused linear BCO (Binary Classifier Optimization) loss

  #1585·Yulong-Cauli opened Nov 1, 2026Contributor·

  ·

• docs: per-kernel roofline/bandwidth benchmark (RMSNorm, SwiGLU)

  #1584·ragulk143 opened Nov 1, 2026·1 of 3 tasks·

  ·
  3

• fix(grpo): keep the Triton gradient when delta < 1 + epsilon_high

  #1583·arun-elr opened Nov 1, 2026·1 of 3 tasks·

  ·

• fix(grpo): mask the luspo loss elementwise to match TRL

  #1582·arun-elr opened Nov 1, 2026·1 of 3 tasks·

  ·

• Add apply_liger_kernel_to_glm5_next (GLM-5.3-Flash): fused linear cross-entropy and RMSNorm

  #1580·eilamc14 opened Oct 31, 2026·3 tasks done·

  ·
  1

• docs: add Simplified Chinese translation (README.zh.md)

  #1578·JasonYeYuhe opened Oct 31, 2026·1 of 3 tasks·

  ·
  1

• fix(cross_entropy): reduce per-token loss in fp32 and add an opt-in in-place backward for reduction="none"

  #1574·JobQiu opened Oct 31, 2026·2 of 3 tasks·

  ·
  2

• feat: add TileLang backend support with Ascend SwiGLU

  #1572·xvxuopop opened Oct 28, 2026Contributor·4 tasks·

  ·

• fix(fused_linear_kl_div): avoid common-offset loss cancellation

  #1569·Levius-Fubuki opened Oct 28, 2026·1 of 3 tasks·

  ·
  1

• test(convergence): make MoE node comparison flip-aware

  #1568·Yulong-Cauli opened Oct 27, 2026Contributor·5 of 6 tasks·

  ·

• misc(npu): pin triton-ascend 3.6.0 and cap cv workspace to one slot

  #1567·zheliuyu opened Oct 27, 2026Contributor·4 of 5 tasks·

  ·

• fix(group_norm): accept 2-D inputs and fix extra_repr

  #1563·anh-nguyenn opened Oct 26, 2026Contributor·

  ·
  1

• docs(README): use Huawei Ascend in the NPU section

  #1561·lltiaor opened Oct 26, 2026·

  ·
  1

• fix(cute): use local scratch for SM90 split-K dH

  #1559·ReinforcedKnowledge opened Oct 25, 2026·3 tasks done·

  ·

• feat(grpo): support top_entropy_quantile in fused linear GRPO loss

  #1558·nvdm59 opened Oct 24, 2026Contributor·1 of 3 tasks·

  ·
  1

• feat(grpo): add off-policy sequence masking (OPSM) to fused linear GRPO loss

  #1557·nvdm59 opened Oct 24, 2026Contributor·1 of 3 tasks·

  ·
  1