Issues

• [Megatron] apply_liger_kernel_to_megatron(cross_entropy=True) does not reach Megatron's CE call sites

  linkedin/Liger-Kernel#1485 · Filter by author Rapisurazurite opened on Sep 24, 2026

• [NPU] Track reverting Ascend Atlas 950 workarounds after triton-ascend upstream fixes (#1464)

  linkedin/Liger-Kernel#1481 · Filter by author zheliuyu opened on Sep 21, 2026

• JSD kernel returns NaN for interior beta when a vocabulary entry underflows under both distributions

  linkedin/Liger-Kernel#1453 · Filter by author KyleMylonakisProtopia opened on Sep 10, 2026

• [RFC] Migrate Nvidia CI from Modal to HF Jobs

  linkedin/Liger-Kernel#1393 · Filter by author ParagEkbote opened on Aug 17, 2026

• [RFC] Testing Methodology for FP32 Convergence Tests on Fused SwiGLU/dSwiGLU Kernels

  linkedin/Liger-Kernel#1383 · Filter by author Pearblossom-M opened on Aug 15, 2026
  1

• Add Liger kernels + FLCE token_accuracy plumbing for Muse Glimmer (muse_glimmer)

  linkedin/Liger-Kernel#1378 · Filter by author MegaGimen opened on Aug 13, 2026
  2

  

• [RFC] Full-pipeline SwiGLU MLP with forward/backward co-optimization

  linkedin/Liger-Kernel#1347 · Filter by author Pearblossom-M opened on Aug 6, 2026
  6

• [B300/SM103] Compiled Liger SwiGLU is ~3.1x slower than torch.compile for a Qwen3-shaped BF16 tensor

  linkedin/Liger-Kernel#1330 · Filter by author jessiewei7 opened on Aug 2, 2026
  2

• torch.compile fails with FLCE due to aten.addmm.dtype_out overload

  linkedin/Liger-Kernel#1327 · Filter by author hiwuhgds-pixel opened on Jul 30, 2026

• Reduce FLCE weight-gradient allocation overhead across FP16, BF16, and FP32

  linkedin/Liger-Kernel#1323 · Filter by author Anubhav-2003 opened on Jul 25, 2026

• Support GMPO Chunked loss

  linkedin/Liger-Kernel#1315 · Filter by author chinoll opened on Jul 23, 2026
  1

• Support Gemma 4 Unified (gemma4_unified) — kernel patching + fused linear cross-entropy

  linkedin/Liger-Kernel#1308 · Filter by author lefft opened on Jul 21, 2026

• Triton kernels can launch on the wrong GPU with multi-GPU inputs

  linkedin/Liger-Kernel#1303 · Filter by author kashif opened on Jul 16, 2026
  1

• Fused MoE kernels crash on Qwen3.5-397B (int32 offset overflow)

  linkedin/Liger-Kernel#1296 · Filter by author likejazz opened on Jul 10, 2026
  1

• Add Qwen2.5-Omni support

  linkedin/Liger-Kernel#1295 · Filter by author VietDunghacker opened on Jul 9, 2026
  1

• Feature Request: Add THD / Packed-Sequence Training Support

  linkedin/Liger-Kernel#1281 · Filter by author niyunsheng opened on Jul 1, 2026
  2

• Suggestion: per-kernel roofline/bandwidth-efficiency benchmarks in docs

  linkedin/Liger-Kernel#1272 · Filter by author ragulk143 opened on Jun 27, 2026
  1

• LigerORPOTrainer doesn't work when not using FSDP

  linkedin/Liger-Kernel#1229 · Filter by author Ola-Vish opened on May 19, 2026
  1

• Add fused Modulated RMSNorm for DiT-style scale/shift conditioning

  linkedin/Liger-Kernel#1224 · Filter by author yukiu00 opened on May 13, 2026
  5

• [RFC] Add Monitoring Interface for External Logging Backends

  linkedin/Liger-Kernel#1213 · Filter by author ParagEkbote opened on May 1, 2026
  3

• [New backend] cuTile backend for Liger kernels — performance data on B200 + integration interest

  linkedin/Liger-Kernel#1205 · Filter by author xjmxyt opened on Apr 27, 2026
  1

• Gemma4 text model support

  linkedin/Liger-Kernel#1186 · Filter by author ruilin-gif opened on Apr 6, 2026
  5

• Glm4vConfig object has no attribute 'hidden_size'

  linkedin/Liger-Kernel#1152 · Filter by author akowalsk opened on Mar 18, 2026
  1

• LigerFusedLinearGRPOLoss produces ~100x larger grad_norm than TRL's non-Liger-Kernel path due to missing vLLM IS correction and other differences

  linkedin/Liger-Kernel#1082 · Filter by author yukiu00 opened on Feb 7, 2026
  11

• Skip dW/dB computation in norm ops when weight/bias is frozen (LoRA/PEFT optimization)

  linkedin/Liger-Kernel#1067 · Filter by author yukiu00 opened on Feb 4, 2026
  1