THIRD_PARTY_NOTICES

Third-party notices

This repository contains a small amount of code vendored from, or adapted from, the projects below. Each such file
keeps its upstream notice and says what it took and what it changed. None of the projects' source trees are
included. The training stack (Megatron-LM, slime, PyTorch, Transformer Engine) and the Python packages the scripts
import (numpy, PyYAML, pytest, pyarrow, transformers) are runtime dependencies installed
separately; they are not distributed here.

------------------------------------------------------------------------------------------------------------------------
NVIDIA Megatron-Bridge
  Source:  https://github.com/NVIDIA-NeMo/Megatron-Bridge, commit c3c974d4f2e315db8db01945acab6555049c40d7
  License: Apache License 2.0 (https://www.apache.org/licenses/LICENSE-2.0)
  Used in: skill/scripts/correctness/mbridge_convergence.py
  What:    validate_convergence() and its numpy import, vendored verbatim from scripts/performance/utils/evaluate.py,
           with NVIDIA's Apache-2.0 header (lines 1-13 of the file). The correctness check's loss-curve review calls it
           unmodified. Lines 15-19 are a provenance note added here.

------------------------------------------------------------------------------------------------------------------------
slime
  Source:  https://github.com/THUDM/slime
  License: Apache License 2.0 (https://www.apache.org/licenses/LICENSE-2.0)
  Used in: skill/scripts/profile_analyzer/taxonomy.py
  What:    The kernel-name classification rules in v1_category() are adapted from slime's tools/analyze_profile.py.
           The rest of the profile analyzer is original.
  Changes: Rules moved into a taxonomy module alongside new training-specific classifications (compute classes,
           parallelism-aware communication, host-side calls).
  Used in: skill/scripts/data/patch_glm5_mask.py
  What:    gen_multi_turn_loss_mask_glm5() is adapted from gen_multi_turn_loss_mask_qwen3_5() in slime's
           slime/utils/mask_utils.py (the rendered-text, offset-mapping and character-mask approach).
  Changes: GLM-5 chat-template markers (<|assistant|>, <|user|>, <|observation|>, <|endoftext|>), think-prefix
           handling, and an in-place installer that adds the method and its get_loss_mask() branch.
  Used in: skill/scripts/data/patch_gptoss_mask.py
  What:    gen_multi_turn_loss_mask_gpt_oss() is adapted from gen_multi_turn_loss_mask_qwen3_5() in slime's
           slime/utils/mask_utils.py (the rendered-text, offset-mapping and character-mask approach).
  Changes: harmony chat-template markers (<|start|>assistant, <|end|>, <|return|>, <|call|>), a frozen template date,
           checks that refuse conversations the template would not render faithfully, and an in-place installer that
           adds the method and its get_loss_mask() branch.