Liger Kernel Files

Summary

Liger-Kernel v0.8.4 expands model and kernel coverage while improving correctness, performance, and deployment reliability across CUDA, Triton, CuTeDSL, and Ascend NPU backends. The release adds DeepSeek-V3 integration, a fused cross-entropy + TV distance kernel, configurable NVSHMEM support, and a shared LCK configuration API for FLSCE and MoE. It also includes PPO optimizations, improved numerical stability for JSD, LayerNorm, TVD, and cross-entropy, fixes for large tensor indexing and multidimensional group norm, and several model-specific correctness fixes. Packaging and release infrastructure are strengthened with PyPI Trusted Publishing and clearer CUDA/Python compatibility handling, alongside documentation updates and contributions from 12 new community members.

What's Changed

• Use a dedicated PyPI credential for LCK by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1473
• Publish LCK with PyPI Trusted Publishing by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1474
• Fix CUTEDSL FLCE functional wrapper arguments by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1475
• fix(npu): Ascend A3/950 UB tile cap and 950 workarounds for group norm, GRPO, fused MoE by @zheliuyu in https://github.com/linkedin/Liger-Kernel/pull/1464
• fix(npu): Replace Ascend A5 comments with 950 by @zheliuyu in https://github.com/linkedin/Liger-Kernel/pull/1484
• perf(ppo): skip dW/dB calculation when lm_head parameters are frozen by @djl11 in https://github.com/linkedin/Liger-Kernel/pull/1488
• refactor(kl_div): route fused-linear KL inner kernel through the dispatcher by @Yulong-Cauli in https://github.com/linkedin/Liger-Kernel/pull/1476
• Fix ignored-token masking in cosine distillation by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1478
• fix(triton): widen blocked RMSNorm row index by @arcusbuilds in https://github.com/linkedin/Liger-Kernel/pull/1482
• [Feature] Support apply_liger_kernel_to_deepseek_v3 by @Yulong-Cauli in https://github.com/linkedin/Liger-Kernel/pull/1424
• Add LCK CUDA-selectable NVSHMEM extras; retain CUDA 12.9 releases by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1492
• fix(kto): sum chosen/rejected rewards per sample when chunk_size > 1 by @kwy404 in https://github.com/linkedin/Liger-Kernel/pull/1495
• fix(triton): widen softmax/sparsemax row index to int64 by @znzryb in https://github.com/linkedin/Liger-Kernel/pull/1497
• Fix softcap padding in cross entropy by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1490
• fix(llama4): honour skip_logits in lce_forward by @yupengtang in https://github.com/linkedin/Liger-Kernel/pull/1445
• Load unsupported model types instead of raising KeyError in AutoLigerKernelForCausalLM by @nileshpatil6 in https://github.com/linkedin/Liger-Kernel/pull/1404
• Add fused CE + TV distance kernel by @kashif in https://github.com/linkedin/Liger-Kernel/pull/1384
• Preserve FP32 LayerNorm statistics for backward by @taking-lying-flat in https://github.com/linkedin/Liger-Kernel/pull/1498
• docs: sync README triton floor with setup.py and fix contributing install command by @kobihikri in https://github.com/linkedin/Liger-Kernel/pull/1301
• Add requires-python >=3.10 to pyproject.toml by @yashb98 in https://github.com/linkedin/Liger-Kernel/pull/1332
• docs: fix input/target shapes in DPO and ORPO forward docstrings by @Aryan95614 in https://github.com/linkedin/Liger-Kernel/pull/1406
• Fix AttnRes silently dropping blocks past 32 by @adityasingh2400 in https://github.com/linkedin/Liger-Kernel/pull/1344
• Fix TVD gradient at p == q by @truong-v in https://github.com/linkedin/Liger-Kernel/pull/1374
• Fix group norm backward for inputs with more than three dimensions by @truong-v in https://github.com/linkedin/Liger-Kernel/pull/1376
• Fix JSD underflow for low-probability vocabulary entries by @mdeff in https://github.com/linkedin/Liger-Kernel/pull/1397
• fix: request float32 logits from the fused JSD projection GEMM by @KyleMylonakisProtopia in https://github.com/linkedin/Liger-Kernel/pull/1433
• Cast tl.program_id to int64 in llama4_rope and qwen2vl_mrope by @adityasingh2400 in https://github.com/linkedin/Liger-Kernel/pull/1342
• Add shared LCK configure API for FLSCE and MoE by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1499
• Warn and skip native configuration on unsupported devices by @kolehma8 in https://github.com/linkedin/Liger-Kernel/pull/1502

New Contributors

• @djl11 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1488
• @taking-lying-flat made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1478
• @arcusbuilds made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1482
• @kwy404 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1495
• @znzryb made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1497
• @nileshpatil6 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1404
• @kobihikri made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1301
• @yashb98 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1332
• @Aryan95614 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1406
• @adityasingh2400 made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1344
• @truong-v made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1374
• @mdeff made their first contribution in https://github.com/linkedin/Liger-Kernel/pull/1397

Full Changelog: https://github.com/linkedin/Liger-Kernel/compare/v0.8.3...v0.8.4