# Changelog

## Score

- CAI 50 → 47 (-2.5)

## Lenses

- Code Health 56 → 56 (-0.3)
- Architecture 95 → 95 (-0.0)
- Maturity 54 → 59 (+4.7)
- Readiness 40 → 43 (+2.8)
- Security 51 → 38 (-12.7)
- Performance 100 → 100 (+0.0)

## Resolved (276)

- Base-context workflow trigger runs with an unscoped token
- BertPipelineForwards.bert_for_multiple_choice_forward (cognitive 20) (colossalai/shardformer/modeling/bert.py)
- BertPipelineForwards.bert_for_question_answering_forward (cognitive 19) (colossalai/shardformer/modeling/bert.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cognitive 30) (colossalai/shardformer/modeling/bert.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cyclomatic 17) (colossalai/shardformer/modeling/bert.py)
- BertPipelineForwards.bert_model_forward (cognitive 84) (colossalai/shardformer/modeling/bert.py)
- BertPipelineForwards.bert_model_forward (cyclomatic 56) (colossalai/shardformer/modeling/bert.py)
- BertPolicy.module_policy (cognitive 17) (colossalai/shardformer/policies/bert.py)
- BertPolicy.module_policy (cyclomatic 16) (colossalai/shardformer/policies/bert.py)
- DeepseekV3Policy.get_held_layers (cognitive 17) (colossalai/shardformer/policies/deepseek_v3.py)
- DistGaloreAwamW.setup_distributed (cognitive 16) (colossalai/nn/optimizer/distributed_galore.py)
- DistGaloreAwamW.step (cognitive 85) (colossalai/nn/optimizer/distributed_galore.py)
- DistGaloreAwamW.step (cyclomatic 27) (colossalai/nn/optimizer/distributed_galore.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/modeling/bert.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/modeling/deepseek.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/modeling/opt.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/policies/bert.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/policies/deepseek.py)
- Duplicated block (10 lines × 2) (colossalai/shardformer/policies/vit.py)
- Duplicated block (10 lines × 3) (colossalai/shardformer/policies/gptj.py)
- …and 256 more

## New (1068)

- Base-context workflow trigger runs with an unscoped token (.github/workflows/translate_comment.yml)
- BertPipelineForwards.bert_for_multiple_choice_forward (cognitive 20) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPipelineForwards.bert_for_multiple_choice_forward (cognitive 20) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_for_question_answering_forward (cognitive 19) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPipelineForwards.bert_for_question_answering_forward (cognitive 19) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cognitive 30) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cognitive 30) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cyclomatic 17) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPipelineForwards.bert_for_sequence_classification_forward (cyclomatic 17) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_model_forward (cognitive 70) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_model_forward (cognitive 84) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPipelineForwards.bert_model_forward (cyclomatic 50) (colossalai/shardformer/modeling/_bert_tf5.py)
- BertPipelineForwards.bert_model_forward (cyclomatic 56) (colossalai/shardformer/modeling/_bert_tf4.py)
- BertPolicy.module_policy (cognitive 17) (colossalai/shardformer/policies/_bert_tf4.py)
- BertPolicy.module_policy (cognitive 17) (colossalai/shardformer/policies/_bert_tf5.py)
- BertPolicy.module_policy (cyclomatic 16) (colossalai/shardformer/policies/_bert_tf4.py)
- BertPolicy.module_policy (cyclomatic 16) (colossalai/shardformer/policies/_bert_tf5.py)
- BlipPolicy.module_policy (cyclomatic 18) (colossalai/shardformer/policies/blip2.py)
- Change coupling: bert.py ↔ bloom.py (colossalai/shardformer/policies/bert.py)
- Change-coupling hub: mixtral.py → _operation.py, deepseek.py, deepseek.py, mixtral.py (colossalai/shardformer/modeling/mixtral.py)
- …and 1048 more

## Changes since last survey

- 29 commits — 18 feature/other, 11 fixes

## By area

- (repo) — 9 commits
- colossalai/shardformer — 6 commits
- .github/workflows — 4 commits
- .github/scripts — 3 commits
- docker/torch213 — 3 commits
- colossalai/booster — 2 commits
- extensions/pybind — 1 commit
- tests/test_checkpoint_io — 1 commit

## Notable commits

- fix: Merge pull request #6459 from b1ack-j/fix/cuda-wheel-build-and-loading
- fix: Merge pull request #6467 from hpcaitech/fix/moe-dp-grad-scale
- fix: Merge pull request #6468 from hpcaitech/fix/moe-sp-expert-grad-sync
- fix: Merge pull request #6469 from hpcaitech/fix/zero-ep-grad-norm
- fix: [compatibility]: fix ChatGLM and PEFT for Transformers 5.17
- fix: [pre-commit.ci] auto fixes from pre-commit.com hooks
- fix: [pre-commit.ci] auto fixes from pre-commit.com hooks
- fix: [pre-commit.ci] auto fixes from pre-commit.com hooks
- fix: fix(ci): run compatibility tests through Python
- fix: fix(tests): restore optional Apex capability probes
- fix: fix: correct CUDA wheel builds and prebuilt extension loading
- change: Add version-dispatched Transformers 5.17 ShardFormer compatibility
- change: Apply repository import sorting and Python formatting to HF5 changes
- change: Gather Qwen3 vocabulary logits when parallel output is disabled
- change: Merge pull request #6462 from hpcaitech/ci/three-layer-workflow-validation
- change: Merge pull request #6463 from hpcaitech/codex/torch213-ci-image
- change: Merge pull request #6472 from hpcaitech/ci/remove-obsolete-submodule-workflow
- change: Merge pull request #6474 from Icarus1411/kai-dev
- change: Merge pull request #6476 from hpcaitech/feat/upgrade_transformer_5.17
- change: [ci] Remove obsolete submodule synchronization workflow
- …and 9 more