Search results

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

Status: Open.

#1082 In linkedin/Liger-Kernel;

· yukiu00 opened on Feb 7, 2026

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

Status: Open.

#1067 In linkedin/Liger-Kernel;

· yukiu00 opened on Feb 4, 2026

Add mHC (Manifold-Constrained Hyper-Connections) fused kernels to Liger-Kernel

Status: Open.

#1066 In linkedin/Liger-Kernel;

· yukiu00 opened on Feb 4, 2026

[RFC] Standardized Model Config for Benchmarks

Status: Open.

#1051 In linkedin/Liger-Kernel;

· Tcc0403 opened on Jan 29, 2026

Compatibility issue between liger-kernel and zero3: The size of tensor a (0) must match the size of tensor b (4096) at non-singleton dimension 1

Status: Open.

#1048 In linkedin/Liger-Kernel;

· Snowdar opened on Jan 29, 2026

[NPU] Improve the performance of the embedding for NPU

Status: Open.

#1036 In linkedin/Liger-Kernel;

· TianHao324 opened on Jan 21, 2026

[NPU]support fused_neighborhood_attention for npu

Status: Open.

#1029 In linkedin/Liger-Kernel;

· Hailey-Zh opened on Jan 19, 2026

[NPU]RFC: Ascend CI Integration

Status: Open.

#1022 In linkedin/Liger-Kernel;

· noemotiovon opened on Jan 16, 2026

Transformers v5 compatibility

good first issueGood for newcomersGood for newcomers

help wantedExtra attention is neededExtra attention is needed

huggingface

Status: Open.

#978 In linkedin/Liger-Kernel;

· Tcc0403 opened on Dec 15, 2025

Fatal Python error when using Liger Kernel with Qwen-VL fine-tuning: <code>none_dealloc: deallocating None</code> at step ~875

Status: Open.

#975 In linkedin/Liger-Kernel;

· lxy21214 opened on Dec 13, 2025

[NPU Roadmap] NPU support for Liger-Kernel

Status: Open.

#969 In linkedin/Liger-Kernel;

· zheliuyu opened on Dec 8, 2025

Moe kernel for latest transformers v5

Status: Open.

#958 In linkedin/Liger-Kernel;

· Tcc0403 opened on Nov 28, 2025