Repository navigation

All pull requests

fix(trainer): release the device cache through the framework helper before the metrics gather
#343li-lizhe opened Oct 26, 2026Contributor
RL续训
#340indigoRM opened Oct 26, 20264 tasks
fix(loss): compute the IPO loss as in Eq. 17 of the IPO paper
#327flyingdoubleG opened Oct 21, 2026Contributor2 of 4 tasks
1
fix(loss): truncate the CISPO IS weight at 1 + epsilon_high
#326flyingdoubleG opened Oct 21, 2026Contributor2 of 4 tasks
feat(sp): Ulysses 模块化
#321werwrewe opened Oct 16, 2026Contributor1 of 4 tasks
31
fix: restore local optimizer state without distributed broadcast
#317Liuchenbing-2026 opened Oct 14, 2026Contributor1 of 4 tasks
1
fix: update on complete gradient accumulation boundaries
#316Liuchenbing-2026 opened Oct 14, 2026Contributor1 of 4 tasks
1
feat: add native typed decision training components
#314Liuchenbing-2026 opened Oct 14, 2026Contributor3 of 4 tasks
feat: add StartLux decision LoRA training on Ascend
#305Liuchenbing-2026 opened Oct 12, 2026Contributor4 tasks done
feat: add Intern-Decision text training cookbook on Ascend
#295Liuchenbing-2026 opened Oct 8, 2026Contributor2 of 3 tasks
fix: weight PPO token-mean micro-batches by valid tokens
#292taking-lying-flat opened Sep 20, 20261 of 4 tasks
fix: detach probability weights in DFT loss
#291taking-lying-flat opened Sep 20, 20261 of 4 tasks
fix: avoid duplicate gradient norm reduction in pure DDP
#290taking-lying-flat opened Sep 20, 20261 of 4 tasks
feat: add Kunlunxin XPU support
#289loveleaves opened Sep 17, 2026
[feature]reward_loop
#284doctorMcy opened Sep 15, 20263 tasks
feat(rl): add SAO synchronous training baseline with skip-observation GAE, DIS policy loss and value critic ```
#269xxyyrr598 opened Aug 19, 2026Contributor
6
add java client
#268ljhandhya opened Aug 18, 20263 tasks
Fix/multi lora ep memory efficient init
#265meichangsu1 opened Aug 14, 2026Collaborator4 tasks
[New feature] 支持多教师在线蒸馏（MOPD）
#239doctorMcy opened Jun 30, 20263 tasks
10

Sunbelt Computer Software

PL/B Language Development and Support

Navigation Menu

Repository navigation

All pull requests

fix(trainer): release the device cache through the framework helper before the metrics gather

RL续训

fix(loss): compute the IPO loss as in Eq. 17 of the IPO paper

fix(loss): truncate the CISPO IS weight at 1 + epsilon_high

feat(sp): Ulysses 模块化

fix: restore local optimizer state without distributed broadcast

fix: update on complete gradient accumulation boundaries

feat: add native typed decision training components

feat: add StartLux decision LoRA training on Ascend

feat: add Intern-Decision text training cookbook on Ascend

fix: weight PPO token-mean micro-batches by valid tokens

fix: detach probability weights in DFT loss

fix: avoid duplicate gradient norm reduction in pure DDP

feat: add Kunlunxin XPU support

[feature]reward_loop

feat(rl): add SAO synchronous training baseline with skip-observation GAE, DIS policy loss and value critic ```

add java client

Fix/multi lora ep memory efficient init

[New feature] 支持多教师在线蒸馏（MOPD）

Sunbelt Computer Software

PL/B Language Development and Support

Pull requests