Issues

• [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）

  deepseek-ai/DeepSpec#90 · Filter by author guienjavar-a11y opened on Sep 13, 2026
  1

• Can I train dspark for deepseek-v4-flash-0731?

  deepseek-ai/DeepSpec#89 · Filter by author Ximingwang-09 opened on Sep 9, 2026
  1

• 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?

  deepseek-ai/DeepSpec#88 · Filter by author Flowerskill opened on Sep 9, 2026
  1

• can I train dspark for qwen3.8-flash-next?

  deepseek-ai/DeepSpec#86 · Filter by author CHNtentes opened on Sep 2, 2026
  1

• Eagle3 peaks at ~4.6 epochs then degrades; also 6% short of the released checkpoint (DSpark/DFlash reproduce)

  deepseek-ai/DeepSpec#84 · Filter by author liayan opened on Aug 24, 2026

• DSpark在多模态场景适用吗，例如qwen3-vl模型

  deepseek-ai/DeepSpec#83 · Filter by author wenyangchou opened on Aug 18, 2026

• Target feature selection for DSpark

  deepseek-ai/DeepSpec#81 · Filter by author sleepyshep opened on Aug 10, 2026

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

  deepseek-ai/DeepSpec#77 · Filter by author washingtoneimae-dot opened on Jul 31, 2026

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

  deepseek-ai/DeepSpec#76 · Filter by author DeclK opened on Jul 29, 2026

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

  deepseek-ai/DeepSpec#71 · Filter by author arthurgao2003 opened on Jul 22, 2026
  1

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

  deepseek-ai/DeepSpec#70 · Filter by author pengyanai opened on Jul 21, 2026

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

  deepseek-ai/DeepSpec#64 · Filter by author xhy-h opened on Jul 9, 2026

• 请问deepseek官方是否有针对新版本模型写作文笔能力修正？

  deepseek-ai/DeepSpec#61 · Filter by author 9DHans opened on Jul 8, 2026
  2

• Feat: Compatible with Gamme4 MoE structure

  deepseek-ai/DeepSpec#60 · Filter by author waiting-xia opened on Jul 8, 2026

• Release speculators used for ablations

  deepseek-ai/DeepSpec#57 · Filter by author sheelfshah opened on Jul 7, 2026
  1

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

  deepseek-ai/DeepSpec#52 · Filter by author Xuan-yi-yan opened on Jul 6, 2026
  2

• reading my chats not funny

  deepseek-ai/DeepSpec#51 · Filter by author ezydubs opened on Jul 4, 2026

• 一个关于生成范式的远期思考（非Bug报告）

  deepseek-ai/DeepSpec#49 · Filter by author madheu opened on Jul 3, 2026
  11

• RFC: Support Domino speculative decoding in DeepSpec?

  deepseek-ai/DeepSpec#46 · Filter by author jianuo-huang opened on Jul 2, 2026
  1

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

  deepseek-ai/DeepSpec#36 · Filter by author Oxygen56 opened on Jul 1, 2026

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

  deepseek-ai/DeepSpec#35 · Filter by author morluto opened on Jun 30, 2026

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

  deepseek-ai/DeepSpec#34 · Filter by author morluto opened on Jun 30, 2026

• Eagle3 fallback attention mask crashes with KV cache

  deepseek-ai/DeepSpec#33 · Filter by author morluto opened on Jun 30, 2026

• Warmup to cosine scheduler handoff can overshoot peak LR

  deepseek-ai/DeepSpec#32 · Filter by author morluto opened on Jun 30, 2026

• Draft sampling uses a different distribution than speculative rejection

  deepseek-ai/DeepSpec#31 · Filter by author morluto opened on Jun 30, 2026