Issues

• Test hygiene pass

  HawkBearPig/dgpp#134 · Filter by author WhollyMindless opened on Oct 18, 2026
  1

• Default HTTP binding to localhost with explicit remote access

  HawkBearPig/dgpp#131 · Filter by author WhollyMindless opened on Oct 17, 2026

• Backlog: skip neutral penalty preparation in the host sampler

  HawkBearPig/dgpp#123 · Filter by author WhollyMindless opened on Oct 17, 2026

• Two findings from MiMo-V2.6-Flash-MOPD on two Sparks: grow admission sheds live requests while idle prefix-cache entries hold the pool; default prefill controller slows a long prompt 4.5x beside short ones

  HawkBearPig/dgpp#116 · Filter by author Pelmog opened on Oct 15, 2026
  5

  

• Interest in supporting audio-input models (Qwen3-ASR series)?

  HawkBearPig/dgpp#115 · Filter by author AhmmedSamier opened on Oct 14, 2026

• MiMo: shared_prefill_mimo's depth-3 graph verdict differs from the plain oracle when the case runs alone

  HawkBearPig/dgpp#114 · Filter by author HawkBearPig opened on Oct 14, 2026

• Evaluate native FP8 attention on GB10: performance and model quality

  HawkBearPig/dgpp#98 · Filter by author HawkBearPig opened on Oct 7, 2026
  1

• Investigate full GLM-5.3 long-context decode performance

  HawkBearPig/dgpp#96 · Filter by author HawkBearPig opened on Oct 7, 2026

• Prompt-token logprobs on /v1/completions (echo + logprobs, max_tokens 0/1) for perplexity and loglikelihood evals on GLM-5.3-Flash

  HawkBearPig/dgpp#83 · Filter by author jetnet opened on Oct 3, 2026
  1

• Add audio support for MiMo-V2.6-Flash

  enhancement
  HawkBearPig/dgpp#82 · Filter by author HawkBearPig opened on Oct 3, 2026

• Feature request: Qwen3-Next-80B-A3B (qwen3_next) on the Qwen path

  HawkBearPig/dgpp#77 · Filter by author MARKYMARK55 opened on Oct 1, 2026
  2

• Stopping a long streaming response causes prefix cache misses and slow subsequent turns

  bug
  HawkBearPig/dgpp#72 · Filter by author HawkBearPig opened on Sep 30, 2026

• Add support for Anthropic-compatible API endpoints

  enhancement
  HawkBearPig/dgpp#71 · Filter by author HawkBearPig opened on Sep 30, 2026

• Head bit-plane exact argmax: read 25–33 % fewer head bytes on every draft/greedy head read, all families

  enhancement
  HawkBearPig/dgpp#69 · Filter by author HawkBearPig opened on Sep 29, 2026
  2

• W4A4 NVFP4 expert prefill must stay opt-in until a model-level gate passes

  HawkBearPig/dgpp#68 · Filter by author HawkBearPig opened on Sep 29, 2026

• Add support for Qwen3.8-27B PrismaAQUA 5.5-bit mixed-precision quantization

  enhancement
  HawkBearPig/dgpp#52 · Filter by author HawkBearPig opened on Sep 25, 2026

• Batching-isolation check fails on GLM Flash hybrid with default dense kernels

  HawkBearPig/dgpp#32 · Filter by author HawkBearPig opened on Sep 22, 2026

• Add NVMe-backed cold storage for reusable KV and prefix cache

  enhancement
  HawkBearPig/dgpp#26 · Filter by author HawkBearPig opened on Sep 21, 2026
  3