Search results

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

Status: Open.

#1330 In linkedin/Liger-Kernel;

· jessiewei7 opened on Aug 2, 2026

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

Status: Open.

#1327 In linkedin/Liger-Kernel;

· hiwuhgds-pixel opened on Jul 30, 2026

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

Status: Open.

#1323 In linkedin/Liger-Kernel;

· Anubhav-2003 opened on Jul 25, 2026

Support GMPO Chunked loss

Status: Open.

#1315 In linkedin/Liger-Kernel;

· chinoll opened on Jul 23, 2026

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

Status: Open.

#1308 In linkedin/Liger-Kernel;

· lefft opened on Jul 21, 2026

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

Status: Open.

#1303 In linkedin/Liger-Kernel;

· kashif opened on Jul 16, 2026

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

Status: Open.

#1302 In linkedin/Liger-Kernel;

· kobihikri opened on Jul 15, 2026

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

Status: Open.

#1296 In linkedin/Liger-Kernel;

· likejazz opened on Jul 10, 2026

Add Qwen2.5-Omni support

Status: Open.

#1295 In linkedin/Liger-Kernel;

· VietDunghacker opened on Jul 9, 2026

Feature Request: Add THD / Packed-Sequence Training Support

Status: Open.

#1281 In linkedin/Liger-Kernel;

· niyunsheng opened on Jul 1, 2026

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

Status: Open.

#1272 In linkedin/Liger-Kernel;

· ragulk143 opened on Jun 27, 2026

LigerORPOTrainer doesn't work when not using FSDP

Status: Open.

#1229 In linkedin/Liger-Kernel;

· Ola-Vish opened on May 19, 2026