Liger Kernel Files

Summary

Liger-Kernel v0.8.5 adds partial RoPE, selected-token log probabilities in fused scaled cross-entropy, and DeepSeek V2 support. It improves RL-loss correctness, model compatibility, and  torch.compile  support, with fixes for CUDA 13/Blackwell MoE and CuTe DSL. Performance optimizations eliminate unnecessary distillation loss and frozen-parameter normalization gradient computations, alongside improvements to CI and release/nightly publishing.

Compatibility update: The minimum supported Transformers version is now 5.0.0.

What's Changed

• Fix CUDA 13 Blackwell MoE with uniform 1SM UMMA by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1505
• feat(ops): support partial RoPE in Triton kernel by @piyushumate in https://github.com/linkedin/Liger-Kernel/pull/1451
• Allow manual publishing of Liger Kernel releases by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1504
• fix(ppo): warn instead of raising when ref_per_token_logps and ref_input are both passed by @kwy404 in https://github.com/linkedin/Liger-Kernel/pull/1506
• fix(grpo): correct sequence-level KL and masked loss gradients by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1491
• [Feature] Support apply_liger_kernel_to_deepseek_v2 by @Yulong-Cauli in https://github.com/linkedin/Liger-Kernel/pull/1501
• Fix sequence-level vLLM correction in VESPO loss by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1493
• fix(glm4v): version-aware hidden_size/vocab_size for Glm4vConfig by @1fanwang in https://github.com/linkedin/Liger-Kernel/pull/1214
• Fix qwen2_vl / qwen2_5_vl class-level RMSNorm patch on transformers >= 5.2 by @Zhuoxi2000 in https://github.com/linkedin/Liger-Kernel/pull/1511
• Fix llama4 class-level SwiGLU patch for configs with dense layers by @Zhuoxi2000 in https://github.com/linkedin/Liger-Kernel/pull/1507
• [Megatron] Rebind CE call sites so apply_liger_kernel_to_megatron(cross_entropy=True) takes effect by @Rapisurazurite in https://github.com/linkedin/Liger-Kernel/pull/1486
• fix(kl_div, tvd): make torch.compile work (NameError, then fp64 loop-carried scalar) by @roy6n23 in https://github.com/linkedin/Liger-Kernel/pull/1509
• fix: respect gradient column strides in fused linear CE by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1500
• bench(flce): add configurable chunk-memory sweep by @fuxiangdu in https://github.com/linkedin/Liger-Kernel/pull/1479
• fix(monkey_patch): patch Q/K norms on existing Qwen models by @luca-888 in https://github.com/linkedin/Liger-Kernel/pull/1470
• fix(cutedsl): fuse Gemma RMSNorm weight offset in FP32 by @luca-888 in https://github.com/linkedin/Liger-Kernel/pull/1462
• fix(grpo): respect independent token and mask strides by @LOGO127 in https://github.com/linkedin/Liger-Kernel/pull/1461
• Fix Falcon-H1 output scaling in LCE forward by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1477
• test(cutedsl): gap-op dispatcher tests + fused_add_rms_norm 3-D backward + shape fallbacks by @arde171 in https://github.com/linkedin/Liger-Kernel/pull/1457
• fix(test): use even head_dim in rope functional-correctness weird shape by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1513
• perf(distillation): skip the hard (CE) loss when it is neither weighted nor returned by @djl11 in https://github.com/linkedin/Liger-Kernel/pull/1489
• fix(monkey_patch): let the norm flags cover the whole vision tower by @yupengtang in https://github.com/linkedin/Liger-Kernel/pull/1443
• perf(norm): skip dW/dB calculation when parameters are frozen (#1067) by @piyushumate in https://github.com/linkedin/Liger-Kernel/pull/1440
• fix(test): drop unsupported device kwarg from config-based RoPE con… by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1514
• ci: run only Checkstyle in the merge queue; make AMD/Intel manual-only by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1515
• fix(test): stop passing removed device kwarg to Llama4 rotary embed… by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1516
• Update README.md by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1522
• Add selected-token log probabilities to fused scaled cross entropy by @yueyiming2009 in https://github.com/linkedin/Liger-Kernel/pull/1533
• fix(rope): pass GPT-OSS's rotary width explicitly instead of inferring it by @jayzuccarelli in https://github.com/linkedin/Liger-Kernel/pull/1530
• Raise minimum transformers to 5.0.0 and stabilize mini-model tests by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1543
• Bump shared Liger release version to 0.8.5 by @yueyiming2009 in https://github.com/linkedin/Liger-Kernel/pull/1546
• Update publish-nightly.yml by @vaibhavjindal in https://github.com/linkedin/Liger-Kernel/pull/1552

New Contributors

• @1fanwang made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1214
• @Zhuoxi2000 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1511
• @Rapisurazurite made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1486
• @roy6n23 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1509
• @fuxiangdu made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1479
• @LOGO127 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1461
• @jayzuccarelli made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1530

Full Changelog: https://github.com/linkedin/Liger-Kernel/compare/v0.8.4...v0.8.5