sgl-project/SpecForge

View on GitHub

Train speculative decoding models effortlessly and port them smoothly to SGLang serving.

Stars: 1.2k
Forks: 346
Open beginner issues: 0
Indexed issues: 73
Avg merge: 4d 1h
Merged PRs (30d): 41

Dominant language: Python
License: MIT
Last GitHub push: Sep 16, 2026
Latest indexed: Sep 19, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: help wanted

Beginner friendly Everything

Open All

73 open issues indexed

[Bug] Issue with the loss mask of eagle3 in data/parse.py Open

Difficulty 2/5 1-3 hours Newbie friendliness 82/100

sgl-project/SpecForge#885 · 1 comment · Opened Sep 16, 2026
Wondering if there are documents for human being not for agents Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

sgl-project/SpecForge#768 · 1 comment · Opened Aug 16, 2026
[Question] Disaggregated training (producer:consumer 1:7) ~3.2x slower than DP for Qwen3.6-35B-A3B — expected? Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

sgl-project/SpecForge#718 · 2 comments · Opened Jul 24, 2026
[Feature] [RFC] Add EDSD Support to SpecForge Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

sgl-project/SpecForge#695 · 2 comments · 2 reactions · Opened Jul 16, 2026
[Feature] DSpark support Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#626 · 1 comment · Opened Jun 30, 2026
[Feature] Support image-text multimodal input for DFlash training, similar to Eagle3 Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

sgl-project/SpecForge#583 · 5 comments · Opened Jun 17, 2026
Qwen3-8B EAGLE3 example: learning_rate=1e-4 shows large grad-norm outliers in early training Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

sgl-project/SpecForge#577 · Opened Jun 10, 2026
[Feature] EAGLE3.1 support Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

sgl-project/SpecForge#567 · 1 reaction · Opened May 27, 2026
[Feature] Reproducing Qwen3.6 35B-A3B DFlash Training Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#560 · 10 comments · 1 reaction · Opened May 25, 2026
[Bug] Found an error when running the example. Open

Difficulty 3/5 1-2 days Newbie friendliness 52/100

sgl-project/SpecForge#557 · Opened May 21, 2026
[Bug] Chat Template Not Applied Correctly to Recent Models Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

sgl-project/SpecForge#551 · 2 comments · Opened Apr 30, 2026
High results fluctuation in repeating experiments with same setup Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#550 · Opened Apr 28, 2026
Question about train dflash for a large target model Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

sgl-project/SpecForge#546 · Opened Apr 27, 2026
[Bug] There is a gap between the acceptance rates of training and inference Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

sgl-project/SpecForge#533 · 3 comments · 1 reaction · Opened Apr 14, 2026
[Feature] Need offline training support for Dflash Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#521 · 1 comment · Opened Apr 1, 2026
[Bug] qwen3.5-35b-a3b dflash Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

sgl-project/SpecForge#520 · 5 comments · Opened Mar 31, 2026
[Bug] Dflash Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#501 · 3 comments · Opened Mar 16, 2026
[Feature] Plots using benchmarking result files Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

sgl-project/SpecForge#499 · 1 comment · Opened Mar 12, 2026
[Feature] will support qwen3.5-mtp training and qwen3.5-eagle3 training? Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

sgl-project/SpecForge#497 · 3 comments · Opened Mar 9, 2026
Support for target and draft models on different GPUs Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

sgl-project/SpecForge#487 · 2 comments · Opened Mar 3, 2026
[Feature] [RFC] Add Qwen3.5 Support for Speculative Decoding Training (EAGLE3 / DFlash) Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#486 · 3 comments · 8 reactions · Opened Mar 3, 2026
[Feature] What is the upper limit to the performance improvement of Dflash training as the step increases? Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#469 · 13 comments · 2 reactions · Opened Feb 13, 2026
[Feature] [RFC] DFlash Training Adaptation for Qwen VL Models Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#461 · 1 reaction · Opened Feb 5, 2026
EAGLE for GPT-OSS-20B Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#459 · 4 comments · Opened Feb 3, 2026
[Bug] flex_attention BlockMask incorrectly created with H=num_heads Open

Difficulty 2/5 1-3 hours Newbie friendliness 58/100

sgl-project/SpecForge#452 · Opened Jan 27, 2026
[Feature] [RFC] Add a step-disaggregated draft model for speculative decoding, PRISM Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

sgl-project/SpecForge#444 · Opened Jan 21, 2026
[Regression] requirements.txt is empty Open

Difficulty 1/5 Under an hour Newbie friendliness 45/100

sgl-project/SpecForge#432 · 5 comments · 1 reaction · Opened Jan 18, 2026
[Feature] Optimization of Storage Efficiency for Hidden States in Large-Scale Models Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

sgl-project/SpecForge#426 · 1 comment · 2 reactions · Opened Jan 13, 2026
[Feature] Improve multi-layer decoder of custom backends Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#418 · Opened Jan 11, 2026
[Bug] apply_rotary_pos_emb posiiton_ids should add lck Open

Difficulty 3/5 1-2 days Newbie friendliness 25/100

sgl-project/SpecForge#414 · 3 comments · Opened Jan 9, 2026
Why not use input_ids to compute_target_p? Open

Difficulty 3/5 1-2 days Newbie friendliness 35/100

sgl-project/SpecForge#413 · Opened Jan 8, 2026
training context length Open

Difficulty 5/5 Over a week Newbie friendliness 10/100

sgl-project/SpecForge#405 · 1 comment · Opened Jan 4, 2026
[Feature] Better remote dev experience Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

sgl-project/SpecForge#402 · 1 comment · 1 reaction · Opened Jan 3, 2026
Poor Chinese Performance in EAGLE3 Model Due to Lack of Chinese Training Data Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#388 · 4 comments · Opened Dec 24, 2025
[Bug] Question Regarding Memory Consumption Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#380 · 3 comments · Opened Dec 22, 2025
Development Roadmap 2026 Q1 Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#374 · 16 comments · 17 reactions · Opened Dec 19, 2025
Supports multi-machine training Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#373 · 4 comments · Opened Dec 17, 2025
[Bug] EAGLE3 Deepseek-r1-distill-llama8b traning Index errors Open

Difficulty 4/5 3-5 days Newbie friendliness 38/100

sgl-project/SpecForge#368 · Opened Dec 16, 2025
During training, the accuracy is below 1 and exhibits significant fluctuations. Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

sgl-project/SpecForge#365 · 1 reaction · Opened Dec 13, 2025
[Question] How to train the Vicuna model series? Open

Difficulty 3/5 1-2 days Newbie friendliness 25/100

sgl-project/SpecForge#361 · 1 comment · Opened Dec 10, 2025
[Bug] TypeError: 'SGLangEagle3TargetModel' object is not callable Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

sgl-project/SpecForge#360 · Opened Dec 9, 2025
[Question]What factors contribute to Eagle3's reduced acceleration performance on VLM architectures compared to traditional LLMs? Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

sgl-project/SpecForge#352 · 5 comments · Opened Dec 5, 2025
[Bug] args.build-dataset-num-proc doesn't work Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#349 · 2 comments · Opened Dec 3, 2025
[Bug] latest benchmark run llama3.1-8b mat1 RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != c10::Half Open

Difficulty 4/5 3-5 days Newbie friendliness 30/100

sgl-project/SpecForge#348 · 1 comment · Opened Dec 3, 2025
vlm target model don't have set_aux_hidden_states_layers Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#345 · 1 comment · 2 reactions · Opened Dec 2, 2025
draft train Open

Difficulty 5/5 Over a week Newbie friendliness 15/100

sgl-project/SpecForge#338 · Opened 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 Open

Difficulty 4/5 3-5 days Newbie friendliness 15/100

sgl-project/SpecForge#336 · 6 comments · Opened Nov 28, 2025
EAGLE3 on Qwen2.5-VL / Qwen3-VL shows extremely low accept length (accept_len ≈ 1) Open

bug low accuracy

Difficulty 4/5 3-5 days Newbie friendliness 25/100

sgl-project/SpecForge#310 · 10 comments · Opened Nov 19, 2025
[Bug] 0.00 acc at the start of finetuning pretrained draft model from huggingface Open

bug low accuracy

Difficulty 4/5 3-5 days Newbie friendliness 35/100

sgl-project/SpecForge#300 · 5 comments · Opened Nov 14, 2025
[RFC]: Integrate USP (Ulysses + Ring Attention) for Context Parallelism in SpecForge Open

enhancement feature

Difficulty 5/5 Over a week Newbie friendliness 30/100

sgl-project/SpecForge#299 · 16 comments · Opened Nov 13, 2025

sgl-project/SpecForge

Get new issues in your inbox