v1.4.0

[v1.4.0] 2026-10-14

Breaking Changes

• Removed Runner's layer-wise multi-GPU quantization and its multi_gpu / gpu_ids
  arguments. Remove these arguments from existing Runner(...) calls and use the
  standard quantization workflow. Distributed post-processing via
  GlobalPTQDistributed remains available.

New Features

• MDBF Global PTQ: GlobalPTQ and GlobalPTQDistributed now optimize MDBF
  amplitude parameters through KL distillation. Set optimize_binary=True to also
  train binary factors using a straight-through estimator (STE). See the
  MDBF GlobalPTQ example.
• Global PTQ device placement: GlobalPTQ accepts student_device and
  teacher_device; GlobalPTQDistributed accepts teacher_device. The distributed
  implementation now defaults the teacher to CPU when using DeepSpeed or multiple
  processes, reducing GPU memory usage.
• MoE router fine-tuning: added RouterFineTuning, a post-process that trains
  only router parameters with next-token prediction loss while quantized experts
  and other weights remain frozen. Results use the normal quantized-model save/load
  workflow. See the router fine-tuning example.
• MDBF GGUF export: plain and rotated MDBF checkpoints can now be exported for
  stock llama.cpp through the fallback path. Dense weights are reconstructed and
  online Hadamard transforms are folded into them before conversion. The default
  Q4_K_M output is re-quantized and does not retain MDBF's low-bit compression;
  use Q8_0 or qtype=None (f16, Python API only) to reduce additional quantization
  error. See the CPU inference guide.

Environment

• Added an isolated, dependency-locked OpenVINO 2026.3.1 environment and a
  OneComp GPTQ 4-bit text-generation export example. Checkpoint metadata is
  normalized in a temporary copy without modifying the source checkpoint.
  See the OpenVINO export guide for supported-model
  requirements and inference setup.

Bug Fixes

• GGUF export now rejects forced direct and mixed modes for non-GPTQ layouts
  and rotated checkpoints. Use auto or fallback for these checkpoints.
• Hardened Global PTQ state restoration for GPTQ, DBF, and MDBF after training
  failures, and rollback to the initial state when distributed training does not
  improve the baseline loss.

Tests (CI)

• Added README example-link validation and timeout handling for pending SLURM
  test jobs.