Skip to content

Pull requests: sgl-project/SpecForge

New pull request New
71 Open 432 Closed
71 Open 432 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

refactor: consolidate training on the unified runtime
#678 opened Jul 14, 2026 by maocheng23 Collaborator Loading…
2
feat(training): single-GPU online-mode memory options
#671 opened Jul 12, 2026 by notwitcheer Loading…
6
feat(training): --vocab-mapping-path to pin the vocab mapping across resumes
#670 opened Jul 12, 2026 by notwitcheer Loading…
3
feat(training): gradient checkpointing for the TTT unroll (--draft-gradient-checkpointing)
#669 opened Jul 12, 2026 by notwitcheer Loading…
fix(data): pass per-sample tools list for preformatted batches
#668 opened Jul 10, 2026 by curnane-lab Contributor Loading…
1 of 6 tasks
1
feat: add mtp support
#667 opened Jul 9, 2026 by curnane-lab Contributor Loading…
4 of 6 tasks
1
feat(target_utils): auto-detect embed/lm_head keys for LLM and VLM checkpoints
#666 opened Jul 9, 2026 by curnane-lab Contributor Loading…
6 tasks
3
feat: optional Liger fused RMSNorm/SwiGLU kernels for DFlash training
#665 opened Jul 8, 2026 by jessiewei7 Loading…
3
Fix DFlash sglang target backend for sglang 0.5.13+ scheduler/batch API
#664 opened Jul 8, 2026 by violet73 Loading…
6 tasks
2
fix(optimizer): clip by global grad norm across FSDP ranks in BF16Optimizer
#663 opened Jul 8, 2026 by jessiewei7 Loading…
2
fix(train_dflash): atomic per-file checkpoint writes + _SUCCESS-gated resume
#659 opened Jul 7, 2026 by jessiewei7 Loading…
1
fix(dflash): add sglang 0.5.14 post-load init to the DFlash target model
#658 opened Jul 7, 2026 by jessiewei7 Loading…
2
[Modeling] DFlash MLA (DeepSeek) draft + rebase MLA eagle3 onto the #645 stack
#651 opened Jul 5, 2026 by maocheng23 Collaborator • Draft
8
[Online track] O1.3 — live SGLang-server capture (reforward transport)
#641 opened Jul 2, 2026 by maocheng23 Collaborator • Draft
2
[Modeling] MLA (DeepSeek) Eagle3 draft architecture
#640 opened Jul 2, 2026 by maocheng23 Collaborator • Draft
1
[Eval track] E1 — acceptance-length eval harness: EvalConfig + EvalCache + gates
#639 opened Jul 1, 2026 by maocheng23 Collaborator • Draft
1
fix: resolve NPU OOM with default training config
#620 opened Jun 29, 2026 by curnane-lab Contributor Loading…
1 of 6 tasks
2
feat: add Ascend NPU support for VP-Drafter
#619 opened Jun 29, 2026 by curnane-lab Contributor Loading…
6 tasks
4
feat(dspark): add Ascend NPU support for Qwen3.5-4B DSpark training
#617 opened Jun 29, 2026 by curnane-lab Contributor Loading…
5 of 6 tasks
1
feat: DSpark trainer (DFlash + Markov/confidence heads + L1 distillation)
#613 opened Jun 28, 2026 by maocheng23 Collaborator Loading…
13
feat: support training DFlash for Kimi K2.7 Code and Qwen 3.6 27B
#593 opened Jun 24, 2026 by Boreas618 Loading…
6 tasks
1
feat: add --save-total-limit to auto-cleanup old checkpoints
#590 opened Jun 23, 2026 by jxiaof Loading…
3 of 6 tasks
17
Fix per-rank training state resume
#587 opened Jun 20, 2026 by gq112 Loading…
6 tasks
4
[Feature] VLM DFlash Training: Multi-Model Support for Qwen3-VL / Qwen3.5 / Qwen3.6
#585 opened Jun 18, 2026 by zyk42 Loading…
4
feat: Add Ascend npu support.
#576 opened Jun 9, 2026 by menogrey Loading…
6 tasks
1
10
ProTip! Mix and match filters to narrow down what you’re looking for.