Issues

• [Feature] [RFC] Add LongSpark: fixed-cost parallel drafting for long-context speculative decoding

  sgl-project/SpecForge#976 · Filter by author PFLiu1 opened on Oct 14, 2026

• [Experiment] Compare KV-based and hidden-state-based training and evaluate data scaling

  discussionlow accuracy
  sgl-project/SpecForge#954 · Filter by author jiapingW opened on Oct 13, 2026
  

• [Feature] Expand GPU platform support

  enhancementfeaturehelp wanted
  Feature

  sgl-project/SpecForge#953 · Filter by author jiapingW opened on Oct 13, 2026
  1

  

• [Feature] Integrate Mooncake through the SGLang plugin mechanism

  enhancementhigh priority
  Feature

  sgl-project/SpecForge#952 · Filter by author jiapingW opened on Oct 13, 2026
  

• [Feature] Improve draft model training

  enhancementhelp wantedhigh priority
  Feature

  sgl-project/SpecForge#951 · Filter by author jiapingW opened on Oct 13, 2026
  

• [RFC] Align draft model configuration formats across vLLM, and SGLang

  discussionenhancement
  Feature

  sgl-project/SpecForge#950 · Filter by author jiapingW opened on Oct 13, 2026
  

• Development Roadmap (2026 Q4)

  enhancementhigh priority
  sgl-project/SpecForge#949 · Filter by author jiapingW opened on Oct 12, 2026
  

• [Bug] TargetHead fails for targets with tied word embeddings (e.g. Qwen2.5-0.5B) because lm_head.weight is missing

  sgl-project/SpecForge#927 · Filter by author lru49 opened on Oct 4, 2026

• [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
  3

• [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
  1

• 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