Issues

• EAGLE for GPT-OSS-20B

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

• [Bug] flex_attention BlockMask incorrectly created with H=num_heads

  sgl-project/SpecForge#452 · Filter by author Ximingwang-09 opened on Jan 27, 2026

• [Feature] [RFC] Add a step-disaggregated draft model for speculative decoding, PRISM

  sgl-project/SpecForge#444 · Filter by author Akemiiii opened on Jan 21, 2026

• [Regression] requirements.txt is empty

  sgl-project/SpecForge#432 · Filter by author Dogacel opened on Jan 18, 2026
  5

• [Feature] Optimization of Storage Efficiency for Hidden States in Large-Scale Models

  sgl-project/SpecForge#426 · Filter by author shelterwff-byte opened on Jan 13, 2026
  1

• [Feature] Improve multi-layer decoder of custom backends

  sgl-project/SpecForge#418 · Filter by author ghost opened on Jan 11, 2026

• [Bug] apply_rotary_pos_emb posiiton_ids should add lck

  sgl-project/SpecForge#414 · Filter by author jameswu2014 opened on Jan 9, 2026
  3

• Why not use input_ids to compute_target_p?

  sgl-project/SpecForge#413 · Filter by author jameswu2014 opened on Jan 8, 2026

• training context length

  sgl-project/SpecForge#405 · Filter by author ggg-s opened on Jan 4, 2026
  1

• [Feature] Better remote dev experience

  sgl-project/SpecForge#402 · Filter by author ghost opened on Jan 3, 2026
  1

• Poor Chinese Performance in EAGLE3 Model Due to Lack of Chinese Training Data

  sgl-project/SpecForge#388 · Filter by author lmyybh opened on Dec 24, 2025
  4

• [Bug] Question Regarding Memory Consumption

  sgl-project/SpecForge#380 · Filter by author ggg-s opened on Dec 22, 2025
  3

• Development Roadmap 2026 Q1

  sgl-project/SpecForge#374 · Filter by author FrankLeeeee opened on Dec 19, 2025
  16

• Supports multi-machine training

  sgl-project/SpecForge#373 · Filter by author dongyibo opened on Dec 17, 2025
  4

• [Bug] EAGLE3 Deepseek-r1-distill-llama8b traning Index errors

  sgl-project/SpecForge#368 · Filter by author Sylvan820 opened on Dec 16, 2025

• During training, the accuracy is below 1 and exhibits significant fluctuations.

  sgl-project/SpecForge#365 · Filter by author LOVE-YOURSELF-1 opened on Dec 13, 2025

• [Question] How to train the Vicuna model series?

  sgl-project/SpecForge#361 · Filter by author Sylvan820 opened on Dec 10, 2025
  1

• [Bug] TypeError: 'SGLangEagle3TargetModel' object is not callable

  sgl-project/SpecForge#360 · Filter by author xs-alt opened on Dec 9, 2025

• [Question]What factors contribute to Eagle3's reduced acceleration performance on VLM architectures compared to traditional LLMs?

  sgl-project/SpecForge#352 · Filter by author 330205812 opened on Dec 5, 2025
  5

• [Bug] args.build-dataset-num-proc doesn't work

  sgl-project/SpecForge#349 · Filter by author C3236455482 opened on Dec 3, 2025
  2

• [Bug] latest benchmark run llama3.1-8b mat1 RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != c10::Half

  sgl-project/SpecForge#348 · Filter by author Sylvan820 opened on Dec 3, 2025
  1

• vlm target model don't have set_aux_hidden_states_layers

  sgl-project/SpecForge#345 · Filter by author C3236455482 opened on Dec 2, 2025
  1

• draft train

  sgl-project/SpecForge#338 · Filter by author Abigbigbig opened on Nov 29, 2025

• @Abigbigbig This looks like a different issue from this PR. Let's move to a different issue. I can point you the fix

  sgl-project/SpecForge#336 · Filter by author Abigbigbig opened on Nov 28, 2025
  6

• EAGLE3 on Qwen2.5-VL / Qwen3-VL shows extremely low accept length (accept_len ≈ 1)

  buglow accuracy
  sgl-project/SpecForge#310 · Filter by author C3236455482 opened on Nov 19, 2025
  10