Repository navigation

All pull requests

fix loss micro_batch_scale
#360meichangsu1 opened Oct 30, 2026Collaborator1 of 4 tasks
Add Qwen3-TTS minimal adapter inference (5/5)
#359Wallbreazzz opened Oct 30, 2026
Add Qwen3-TTS manifest preparation and training CLI (4/5)
#358Wallbreazzz opened Oct 30, 2026
Add Qwen3-TTS gradient accumulation and checkpoint recovery (3/5)
#357Wallbreazzz opened Oct 30, 2026
Add Qwen3-TTS Twinkle LoRA integration (2/5)
#356Wallbreazzz opened Oct 30, 2026
feat(model): dp balance
#355lincq2000 opened Oct 30, 2026Collaborator2 of 4 tasks
4
[doc] Clarify Ascend hardware naming in README
#354Devin-DC-Huang opened Oct 29, 20261 of 4 tasks
Add Qwen3-TTS data and forward/loss contracts (1/5)
#350Wallbreazzz opened Oct 28, 2026
1
feat(fsdp): per-field mixed_precision config for native FSDP2
#347kevssim opened Oct 28, 2026Collaborator
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, 20261 of 4 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
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

Pull requests