deepseek-ai/DeepSpec

View on GitHub

DeepSpec: a full-stack codebase for training and evaluating speculative decoding algorithms

Stars: 7.1k
Forks: 667
Open beginner issues: 0
Indexed issues: 32

Dominant language: Python
License: MIT
Last GitHub push: Jul 9, 2026
Latest indexed: Sep 15, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed
PR merge metrics: No merged PRs in 30d

Open All

32 open issues indexed

DSpark as a Confidence-Scheduled Speculative-Decoding System: Adaptive Verification, State-Exposed Semi-Autoregression, and Scalable Target-Feature Caches Open

deepseek-ai/DeepSpec#4 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
我看到sglang和vllm都已提供该特性，是否对minimax等其他模型也具有同样推理提升？ Open

deepseek-ai/DeepSpec#5 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[RFC] Add Ascend NPU Support for DeepSpec Open

deepseek-ai/DeepSpec#6 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Support online training without precomputing the target cache Open

deepseek-ai/DeepSpec#21 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
DSpark trained models for qwen3.5-27B Open

deepseek-ai/DeepSpec#22 · 1 comment · 10 reactions · 0 assignees · Indexed Sep 15, 2026
Turn update my email Open

deepseek-ai/DeepSpec#24 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat: add support for FP8 target cache quantization Open

deepseek-ai/DeepSpec#26 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Draft sampling uses a different distribution than speculative rejection Open

deepseek-ai/DeepSpec#31 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Warmup to cosine scheduler handoff can overshoot peak LR Open

deepseek-ai/DeepSpec#32 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Eagle3 fallback attention mask crashes with KV cache Open

deepseek-ai/DeepSpec#33 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
DSpark train/loss metric is rank-weighted instead of globally weighted Open

deepseek-ai/DeepSpec#34 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Concurrent training data generation can corrupt resume order and buffer unbounded results Open

deepseek-ai/DeepSpec#35 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
launch_sglang_server.sh ignores CUDA_VISIBLE_DEVICES and always launches GPUs 0-7 Open

deepseek-ai/DeepSpec#36 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
RFC: Support Domino speculative decoding in DeepSpec? Open

deepseek-ai/DeepSpec#46 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
一个关于生成范式的远期思考（非Bug报告） Open

deepseek-ai/DeepSpec#49 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
reading my chats not funny Open

deepseek-ai/DeepSpec#51 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[RFC] White-Box Architectural Reference for DSpark: Spherical Normalization, 3-Tier Semantic Cache & Unified Multimodal Token Open

deepseek-ai/DeepSpec#52 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Release speculators used for ablations Open

deepseek-ai/DeepSpec#57 · 1 comment · 2 reactions · 0 assignees · Indexed Sep 15, 2026
Feat: Compatible with Gamme4 MoE structure Open

deepseek-ai/DeepSpec#60 · 0 comments · 3 reactions · 0 assignees · Indexed Sep 15, 2026
请问deepseek官方是否有针对新版本模型写作文笔能力修正？ Open

deepseek-ai/DeepSpec#61 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Proposal] PathOracle: Hidden-State Prediction Layer Skipping for Prefill Acceleration, Complementary to DSpark Open

deepseek-ai/DeepSpec#64 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
feat: support Multi-Token Prediction (MTP) training for LLM and speech modalities Open

deepseek-ai/DeepSpec#70 · 0 comments · 1 reaction · 0 assignees · Indexed Sep 15, 2026
DSpark Qwen3 draft modeling silently ignores `partial_rotary_factor`, producing checkpoints whose config mismatches the trained weights (breaks vLLM serving AL) Open

deepseek-ai/DeepSpec#71 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Question] Why DFlash is worse than EAGLE3 for Gemma4-12B in paper's experiment Open

deepseek-ai/DeepSpec#76 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Off-topic] Question for Damai Dai / Chengqi Deng: has profile-based MoE routing been explored? Open

deepseek-ai/DeepSpec#77 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Target feature selection for DSpark Open

deepseek-ai/DeepSpec#81 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
DSpark在多模态场景适用吗，例如qwen3-vl模型 Open

deepseek-ai/DeepSpec#83 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Eagle3 peaks at ~4.6 epochs then degrades; also 6% short of the released checkpoint (DSpark/DFlash reproduce) Open

deepseek-ai/DeepSpec#84 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
can I train dspark for qwen3.8-flash-next? Open

deepseek-ai/DeepSpec#86 · 1 comment · 1 reaction · 0 assignees · Indexed Sep 15, 2026
Does this framework currently support only the three Qwen3 models listed in the documentation? Does it support training for the Qwen3.5 and Qwen3.6 model series? If not, are there any plans to add support for them in the future? Open

deepseek-ai/DeepSpec#88 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Can I train dspark for deepseek-v4-flash-0731? Open

deepseek-ai/DeepSpec#89 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[架构提案] 针对递归自我改进（RSI）的硬件门控遏制框架（Génesis 协议 V5.0） Open

deepseek-ai/DeepSpec#90 · 0 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

deepseek-ai/DeepSpec

Get new issues in your inbox