DeepSpec
Copyright (c) 2026 The DeepSpec Authors

This product includes software adapted from third-party open-source projects.
The original copyright notices and license terms of those projects are retained
below. Files that incorporate adapted code carry an in-file comment pointing to
the upstream source.

================================================================================
SpecForge — https://github.com/sgl-project/SpecForge
License: Apache License 2.0
--------------------------------------------------------------------------------
Portions of the Eagle3 modeling, loss, optimizer, attention, and evaluation code
are adapted from SpecForge. Affected files include (non-exhaustive):

  deepspec/modeling/eagle3/loss.py
  deepspec/modeling/eagle3/common.py
  deepspec/modeling/eagle3/qwen3/modeling.py
  deepspec/modeling/eagle3/gemma4/modeling.py
  deepspec/eval/eagle3/evaluator.py
  deepspec/utils/optim.py
  scripts/data/download_and_split.py

SpecForge's fused log-softmax loss kernel itself incorporates ideas from the
Unsloth (https://github.com/unslothai/unsloth) and Liger-Kernel
(https://github.com/linkedin/Liger-Kernel) projects.

A copy of the Apache License 2.0 is available at:
  https://www.apache.org/licenses/LICENSE-2.0

================================================================================
DFlash — https://github.com/z-lab/dflash
License: MIT
--------------------------------------------------------------------------------
The DFlash draft-model design and training recipe inform the DSpark/DFlash
configurations and modeling in this repository.

================================================================================
