Issues

• [Bug] 0.00 acc at the start of finetuning pretrained draft model from huggingface

  buglow accuracy
  sgl-project/SpecForge#300 · Filter by author AxeCameron777 opened on Nov 14, 2025
  5

• [RFC]: Integrate USP (Ulysses + Ring Attention) for Context Parallelism in SpecForge

  enhancementfeature
  sgl-project/SpecForge#299 · Filter by author uygnef opened on Nov 13, 2025
  16

• [Feature] Does SpecForge support using quantized models as the target model?

  sgl-project/SpecForge#294 · Filter by author abc123000111 opened on Nov 12, 2025
  3

• [Bug] Missing specforge.modeling.target.sgl_model_wrapper import SglangTargetModel

  sgl-project/SpecForge#292 · Filter by author jianc99 opened on Nov 12, 2025

• [Bug] prepare_hidden_states.py needs sglang==0.5.3 or lower

  sgl-project/SpecForge#274 · Filter by author xhdidi opened on Nov 4, 2025
  3

• [Question] Why d2t = [target_token_ids] - torch.arange(len)?

  sgl-project/SpecForge#272 · Filter by author Tomorrowdawn opened on Oct 31, 2025
  3

• [Bug] RuntimeError: can't start new thread

  sgl-project/SpecForge#264 · Filter by author Sayandip170900 opened on Oct 28, 2025
  2

• [Bug] Training Eagle3 for gpt-oss-120b fails with AttributeError: 'NoneType' object has no attribute 'evictable_size'

  sgl-project/SpecForge#263 · Filter by author gopalsarda opened on Oct 26, 2025
  1

• [Feature] Support for Deepseek Arch (including Kimi-K2)

  sgl-project/SpecForge#257 · Filter by author JensenFire opened on Oct 17, 2025
  2

• [Question] Why Qwen3-30B-A3B eagle3 draft model accept rate is lower than Qwen3-8B?

  sgl-project/SpecForge#252 · Filter by author wenqf11 opened on Oct 13, 2025

• Why is the TPS of eagle3-qwen in the sglang inference of single-card H20 not as high as that of the original QWEN3 when the decoding algorithm is added

  sgl-project/SpecForge#236 · Filter by author positive666 opened on Sep 17, 2025
  1

• SGLang Launch error with Qwen2.5-14B draft model, need help

  sgl-project/SpecForge#231 · Filter by author y-d-y opened on Sep 13, 2025
  8

• [Bug] prepare_data.py: error: unrecognized arguments: --output_path /home/luor/.cache/huggingface/xxxx/dataset/ --test-size 0.01

  sgl-project/SpecForge#229 · Filter by author MaltoseFlower opened on Sep 11, 2025
  1

• [Feature] TTT implementation is not the original paper’s concept, but merely data augmentation

  sgl-project/SpecForge#227 · Filter by author haiduo opened on Sep 10, 2025
  9

• [Bug] TP=16 or TP=32 Failed

  sgl-project/SpecForge#225 · Filter by author fan-niu opened on Sep 8, 2025
  3

• [Feature] Using Sglang For Online Training

  sgl-project/SpecForge#195 · Filter by author zyksir opened on Aug 30, 2025
  3

  

• [Bug] train error

  sgl-project/SpecForge#194 · Filter by author ggg-s opened on Aug 30, 2025
  8

  

• [Feature] VLM model support tp

  sgl-project/SpecForge#166 · Filter by author sleepcoo opened on Aug 22, 2025

• [Feature] VLM model support flex attention

  sgl-project/SpecForge#165 · Filter by author sleepcoo opened on Aug 22, 2025

• [Feature] Unify OnlineEagle3Model and OfflineEagle3Model

  sgl-project/SpecForge#156 · Filter by author fzyzcjy opened on Aug 19, 2025
  1

  

• [Bug] invalid device ordinal while using multi node in data prepare

  sgl-project/SpecForge#154 · Filter by author Ch3ngY1 opened on Aug 19, 2025
  1

• [Bug] Model weights saved incompletely under multi-TP training

  sgl-project/SpecForge#127 · Filter by author Qin10 opened on Aug 10, 2025
  4

• [Bug] No speedup observed when accelerating Qwen3-235B-MoE model

  sgl-project/SpecForge#121 · Filter by author zhangchushu opened on Aug 8, 2025
  2

• Any plan to support training draft model with target model's kv cache?

  sgl-project/SpecForge#119 · Filter by author smart-lty opened on Aug 7, 2025

• Naive brainstorm: accept length simulator

  enhancementhelp wanted
  sgl-project/SpecForge#63 · Filter by author fzyzcjy opened on Jul 24, 2025
  6