Commits

Commit history

1. Commits on Sep 22, 2026

   • [fix] residual step order

     jingyaogong

     committed
     f659b55View commit details

     Copy full SHA for f659b55

     Browse repository at this point

   • Merge pull request #868 from DaoyuanLi2816/fix/final-accumulation-checkpoint

     Show description for 6fff76f
     jingyaogong

     authored
     6fff76fView commit details

     Copy full SHA for 6fff76f

     Browse repository at this point

   • Merge pull request #867 from DaoyuanLi2816/fix/lora-inference-paths

     Show description for 7da0f92
     jingyaogong

     authored
     7da0f92View commit details

     Copy full SHA for 7da0f92

     Browse repository at this point

   • fix: update residual gradients before saving final checkpoints

     DaoyuanLi2816

     committed
     a6ffddcView commit details

     Copy full SHA for a6ffddc

     Browse repository at this point

   • fix: align LoRA inference paths with training output

     DaoyuanLi2816

     committed
     4b86522View commit details

     Copy full SHA for 4b86522

     Browse repository at this point
2. Commits on Sep 20, 2026

   • [update] trim redundant comments

     jingyaogong

     committed
     1e6e909View commit details

     Copy full SHA for 1e6e909

     Browse repository at this point

   • Merge pull request #769 from qizwiz/fix/llm-response-unguarded

     Show description for b00be3e
     jingyaogong

     authored
     b00be3eView commit details

     Copy full SHA for b00be3e

     Browse repository at this point

   • Merge pull request #821 from lycDragonSlayer/refactor/remove-redundant-update-policy

     Show description for 7f5ba40
     jingyaogong

     authored
     7f5ba40View commit details

     Copy full SHA for 7f5ba40

     Browse repository at this point

   • Merge pull request #829 from DaoyuanLi2816/fix/dpo-paired-think-format

     Show description for 949488f
     jingyaogong

     authored
     949488fView commit details

     Copy full SHA for 949488f

     Browse repository at this point

   • Merge pull request #863 from Linxiushen/fix-fp16-attention-mask-nan

     Show description for e41bf7c
     jingyaogong

     authored
     e41bf7cView commit details

     Copy full SHA for e41bf7c

     Browse repository at this point

   • Merge pull request #860 from Linxiushen/fix-tokenizer-pretrain-format

     Show description for b330a4d
     jingyaogong

     authored
     b330a4dView commit details

     Copy full SHA for b330a4d

     Browse repository at this point

   • Merge pull request #865 from Linxiushen/fix-requirements-missing-fastapi

     Show description for b956bad
     jingyaogong

     authored
     b956badView commit details

     Copy full SHA for b956bad

     Browse repository at this point

   • Merge pull request #864 from Linxiushen/fix-nonstream-max-tokens

     Show description for d698c06
     jingyaogong

     authored
     d698c06View commit details

     Copy full SHA for d698c06

     Browse repository at this point

   • [fix] requirements.txt 缺 fastapi/uvicorn，按 README 跑 serve_openai_api.py 直接崩

     Show description for 984a2c3
     Linxiushen

     committed
     984a2c3View commit details

     Copy full SHA for 984a2c3

     Browse repository at this point

   • [fix] serve_openai_api 非流式端点的 max_tokens 不生效，与流式端点行为不一致

     Show description for d4b67f1
     Linxiushen

     committed
     d4b67f1View commit details

     Copy full SHA for d4b67f1

     Browse repository at this point

   • [fix] float16 下 attention mask 的 -1e9 溢出成 -inf，导致整条序列输出 NaN

     Show description for ba10c4d
     Linxiushen

     committed
     ba10c4dView commit details

     Copy full SHA for ba10c4d

     Browse repository at this point

   • [fix] train_tokenizer 读预训练语料时产出 0 条文本，静默训出空词表

     Show description for cb08b93
     Linxiushen

     committed
     cb08b93View commit details

     Copy full SHA for cb08b93

     Browse repository at this point
3. Commits on Sep 18, 2026

   • [fix] unify Ctrl+C handling

     jingyaogong

     committed
     cc312c1View commit details

     Copy full SHA for cc312c1

     Browse repository at this point

   • Merge pull request #857 from wbbeyourself/fix-ctrl-c-exception

     Show description for e2ee878
     jingyaogong

     authored
     e2ee878View commit details

     Copy full SHA for e2ee878

     Browse repository at this point

   • [fix] restore top-1 router gradient #858

     jingyaogong

     committed
     b63a956View commit details

     Copy full SHA for b63a956

     Browse repository at this point

   • Merge pull request #859 from basil-k-aji-dev/fix/moe-router-gradient-topk1

     Show description for 1aae259
     jingyaogong

     authored
     1aae259View commit details

     Copy full SHA for 1aae259

     Browse repository at this point
4. Commits on Sep 17, 2026

   • [fix] keep the MoE router trainable at num_experts_per_tok=1

     Show description for 1c87643
     basil-k-aji-dev

     committed
     1c87643View commit details

     Copy full SHA for 1c87643

     Browse repository at this point
5. Commits on Sep 16, 2026

   • fix: eval_llm.py Ctrl+C 退出异常修复

     wbbeyourself

     committed
     9fa620eView commit details

     Copy full SHA for 9fa620e

     Browse repository at this point
6. Commits on Sep 15, 2026

   • [update] trim verbose ddp comments

     jingyaogong

     committed
     7a9137dView commit details

     Copy full SHA for 7a9137d

     Browse repository at this point

   • Merge pull request #855 from basil-k-aji-dev/perf/ddp-followups

     Show description for 7988bc8
     jingyaogong

     authored
     7988bc8View commit details

     Copy full SHA for 7988bc8

     Browse repository at this point
7. Commits on Sep 14, 2026

   • [perf] skip the per-forward RoPE buffer broadcast in the RL trainers

     Show description for 37ac206
     basil-k-aji-dev

     committed
     37ac206View commit details

     Copy full SHA for 37ac206

     Browse repository at this point

   • [update] simplify safe math eval

     jingyaogong

     committed
     21ec325View commit details

     Copy full SHA for 21ec325

     Browse repository at this point

   • Merge pull request #847 from a0917-cell/fix/safe-math-eval

     Show description for cdf0a93
     jingyaogong

     authored
     cdf0a93View commit details

     Copy full SHA for cdf0a93

     Browse repository at this point

   • Merge pull request #854 from basil-k-aji-dev/fix/rl-ddp-grad-sync

     Show description for c367572
     jingyaogong

     authored
     c367572View commit details

     Copy full SHA for c367572

     Browse repository at this point

   • [fix] route RL trainer forward through DDP so gradients all-reduce

     Show description for 9c088dd
     basil-k-aji-dev

     committed
     9c088ddView commit details

     Copy full SHA for 9c088dd

     Browse repository at this point
8. Commits on Sep 10, 2026

   • [fix] keep agent rollout token-aligned #850

     jingyaogong

     committed
     a3c7b01View commit details

     Copy full SHA for a3c7b01

     Browse repository at this point
9. Commits on Sep 6, 2026

   • [fix] replace tool eval() with an AST math evaluator

     Show description for 0f6803e
     

     a0917-cell

     and

     claude

     committed
     0f6803eView commit details

     Copy full SHA for 0f6803e

     Browse repository at this point
10. Commits on Sep 4, 2026

    • [update] remove expert args

      jingyaogong

      committed
      6fc918bView commit details

      Copy full SHA for 6fc918b

      Browse repository at this point

    • Merge pull request #836 from JustinGuese/moe-cli-flags

      Show description for 4bd6062
      jingyaogong

      authored
      4bd6062View commit details

      Copy full SHA for 4bd6062

      Browse repository at this point

    • Merge pull request #840 from JustinGuese/readme-benchmark-resolution

      Show description for 4d7a74a
      jingyaogong

      authored
      4d7a74aView commit details

      Copy full SHA for 4d7a74a

      Browse repository at this point