v1.4.1

[v1.4.1] 2026-10-28

Bug Fixes

• Fixed direct GPTQ-to-GGUF export for Qwen3.8 dense and Qwen3.6 MoE models by applying the required linear-attention V-head permutation and supporting both native and nested text_config layouts. Previously, affected GGUF models could produce corrupted text and extremely high perplexity.
• Fixed generated Hugging Face Hub model cards to reference the published quantized repository in Usage examples while preserving the original base_model metadata and avoiding broken Hub links for local model paths.
• Fixed Gemma 4 MoE GGUF export by preserving the model.language_model.* namespace during dense reconstruction.

Tests

• Fixed tests/onecomp/utils/test_unfuse_moe.py leaving reloaded GPTQ modules and a
  gemlite stub in sys.modules. Tests that ran after it saw a second GPTQLinear
  class, so a full pytest tests/ run failed the slow JointQ + PostProcessLoraSFT
  test with No GPTQLinear layers were found. The test now imports GPTQ,
  GPTQLinear and unfuse_moe from the onecomp package.
• Added a regression test that runs test_unfuse_moe.py in a child process and fails
  if the run replaces an imported onecomp module or the GPTQLinear class.