Repository Issue Activity (beta)

#1453 JSD kernel returns NaN for interior beta when a vocabulary entry underflows under both distributions

#969 [NPU Roadmap, Updated to 2026-Q2] NPU support for Liger-Kernel

#1432 fused_linear_jsd rounds logits to the input dtype before the documented FP32 cast, costing up to 23% gradient error in bf16

#623 [transformers] support DeepSeek V3

#1082 LigerFusedLinearGRPOLoss produces ~100x larger grad_norm than TRL's non-Liger-Kernel path due to missing vLLM IS correction and other differences

#1393 [RFC] Migrate Nvidia CI from Modal to HF Jobs

#1378 Add Liger kernels + FLCE `token_accuracy` plumbing for Muse Glimmer (`muse_glimmer`)

#1383 [RFC] Testing Methodology for FP32 Convergence Tests on Fused SwiGLU/dSwiGLU Kernels

#1375 LigerGroupNorm backward writes out of bounds and then raises for any input with more than 3 dimensions

#1373 LigerTVDLoss returns a non-zero gradient where p == q

#1347 [RFC] Full-pipeline SwiGLU MLP with forward/backward co-optimization

#131 [feat] Add support for encoder-only transformers (e.g. BERT)

#1295 Add Qwen2.5-Omni support

#1335 llama4_rope / qwen2vl_mrope kernels missing the int64 program_id cast that #804 added to rope.py/rms_norm.py

#1330 [B300/SM103] Compiled Liger SwiGLU is ~3.1x slower than torch.compile for a Qwen3-shaped BF16 tensor

#1302 pyproject.toml has no requires-python — pip installs on any interpreter; publish workflows disagree (3.8 vs 3.10)

#1303 Triton kernels can launch on the wrong GPU with multi-GPU inputs

#1296 Fused MoE kernels crash on Qwen3.5-397B (int32 offset overflow)

#1327 `torch.compile` fails with FLCE due to `aten.addmm.dtype_out` overload

#1323 Reduce FLCE weight-gradient allocation overhead across FP16, BF16, and FP32

#1315 Support GMPO Chunked loss

#1281 Feature Request: Add THD / Packed-Sequence Training Support

#1308 Support Gemma 4 Unified (`gemma4_unified`) — kernel patching + fused linear cross-entropy

#920 Check if cross entropy is applied correctly in all monkey patch functions

#1229 LigerORPOTrainer doesn't work when not using FSDP