Issues

• [Bug] loss_mask_filter never runs on the raw-conversation data path

  sgl-project/SpecForge#913 · Filter by author awesome-pro opened on Sep 30, 2026
  1

• [Bug] qwen3.5 / qwen3-next-thinking loss mask only supervises the last assistant turn

  sgl-project/SpecForge#912 · Filter by author awesome-pro opened on Sep 30, 2026

• Wondering if there are documents for human being not for agents

  sgl-project/SpecForge#768 · Filter by author donggrame opened on Aug 16, 2026
  1

• [Question] Disaggregated training (producer:consumer 1:7) ~3.2x slower than DP for Qwen3.6-35B-A3B — expected?

  sgl-project/SpecForge#718 · Filter by author junzhang-zj opened on Jul 24, 2026
  2

• [Feature] [RFC] Add EDSD Support to SpecForge

  sgl-project/SpecForge#695 · Filter by author KerwinKai opened on Jul 16, 2026
  2

• [Feature] DSpark support

  sgl-project/SpecForge#626 · Filter by author Ofir408 opened on Jun 30, 2026
  1

• [Feature] Support image-text multimodal input for DFlash training, similar to Eagle3

  sgl-project/SpecForge#583 · Filter by author curnane-lab opened on Jun 17, 2026
  5

• Qwen3-8B EAGLE3 example: learning_rate=1e-4 shows large grad-norm outliers in early training

  sgl-project/SpecForge#577 · Filter by author heiheiha798 opened on Jun 10, 2026

• [Feature] EAGLE3.1 support

  sgl-project/SpecForge#567 · Filter by author bluecoffee8 opened on May 27, 2026

• [Feature] Reproducing Qwen3.6 35B-A3B DFlash Training

  sgl-project/SpecForge#560 · Filter by author Sylvan820 opened on May 25, 2026
  10

• [Bug] Found an error when running the example.

  sgl-project/SpecForge#557 · Filter by author menogrey opened on May 21, 2026

• [Bug] Chat Template Not Applied Correctly to Recent Models

  sgl-project/SpecForge#551 · Filter by author Dogacel opened on Apr 30, 2026
  2

• High results fluctuation in repeating experiments with same setup

  sgl-project/SpecForge#550 · Filter by author Rachelcoll opened on Apr 28, 2026

• Question about train dflash for a large target model

  sgl-project/SpecForge#546 · Filter by author dongyibo opened on Apr 27, 2026

• [Bug] There is a gap between the acceptance rates of training and inference

  sgl-project/SpecForge#533 · Filter by author xinanjiao opened on Apr 14, 2026
  3

• [Feature] Need offline training support for Dflash

  sgl-project/SpecForge#521 · Filter by author UltramanKuz opened on Apr 1, 2026
  1

• [Bug] qwen3.5-35b-a3b dflash

  sgl-project/SpecForge#520 · Filter by author Jim2016713 opened on Mar 31, 2026
  5

• [Bug] Dflash

  sgl-project/SpecForge#501 · Filter by author Jim2016713 opened on Mar 16, 2026
  3

• [Feature] Plots using benchmarking result files

  sgl-project/SpecForge#499 · Filter by author MLKoz2 opened on Mar 12, 2026
  1

• [Feature] will support qwen3.5-mtp training and qwen3.5-eagle3 training?

  sgl-project/SpecForge#497 · Filter by author aidenchn opened on Mar 9, 2026
  3

• Support for target and draft models on different GPUs

  sgl-project/SpecForge#487 · Filter by author chenxiaodong2002 opened on Mar 3, 2026
  2

• [Feature] [RFC] Add Qwen3.5 Support for Speculative Decoding Training (EAGLE3 / DFlash)

  sgl-project/SpecForge#486 · Filter by author EanWang211123 opened on Mar 3, 2026
  3

• [Feature] What is the upper limit to the performance improvement of Dflash training as the step increases?

  sgl-project/SpecForge#469 · Filter by author 5SSjw opened on Feb 13, 2026
  13

• [Feature] [RFC] DFlash Training Adaptation for Qwen VL Models

  sgl-project/SpecForge#461 · Filter by author ggg-s opened on Feb 5, 2026

• EAGLE for GPT-OSS-20B

  sgl-project/SpecForge#459 · Filter by author yanzi-elastix opened on Feb 3, 2026
  4