Repository Issue Activity (beta)

#81 Target feature selection for DSpark

#80 Target feature selection for DSpark on hybrid GDN models

#21 Support online training without precomputing the target cache

#77 [Off-topic] Question for Damai Dai / Chengqi Deng: has profile-based MoE routing been explored?

#76 [Question] Why DFlash is worse than EAGLE3 for Gemma4-12B in paper's experiment

#73 推理效率异常

#71 DSpark Qwen3 draft modeling silently ignores `partial_rotary_factor`, producing checkpoints whose config mismatches the trained weights (breaks vLLM serving AL)

#70 feat: support Multi-Token Prediction (MTP) training for LLM and speech modalities

#49 一个关于生成范式的远期思考（非Bug报告）

#57 Release speculators used for ablations

#61 请问deepseek官方是否有针对新版本模型写作文笔能力修正？

#52 [RFC] White-Box Architectural Reference for DSpark: Spherical Normalization, 3-Tier Semantic Cache & Unified Multimodal Token

#64 [Proposal] PathOracle: Hidden-State Prediction Layer Skipping for Prefill Acceleration, Complementary to DSpark

#60 Feat: Compatible with Gamme4 MoE structure

#53 Any plan for supporting DS v4 Flash/Pro DSpark Training?

#56 Question about prefill mtp forward

#51 reading my chats not funny

#43 Where is hardware-aware prefix scheduler implemented?

#46 RFC: Support Domino speculative decoding in DeepSpec?

#5 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？

#44 [Draft Checkpoint] eagle3_qwen3_4b_ttt7 hidden_size mismatch (config: 2560, weights: 5120)

#22 DSpark trained models for qwen3.5-27B

#36 launch_sglang_server.sh ignores CUDA_VISIBLE_DEVICES and always launches GPUs 0-7

#35 Concurrent training data generation can corrupt resume order and buffer unbounded results

#34 DSpark train/loss metric is rank-weighted instead of globally weighted