Issues

• Release freeze: 0.3.8

  release-freeze
  midagedev/bloomery#648 · Filter by author midagedev opened on Oct 22, 2026

• gate-gpu-qwen3-serve is red on the 3090: four clauses assume a whole-card load the 30B default does not fit

  bug
  midagedev/bloomery#642 · Filter by author midagedev opened on Oct 22, 2026

• gate-gpu-clef-serve image_refusals: the decode-budget refusal now fires before the encoder-length refusal

  bug
  midagedev/bloomery#641 · Filter by author midagedev opened on Oct 22, 2026

• Problems found while working on bloomery, by severity

  buggpuhostplacementtracking
  midagedev/bloomery#637 · Filter by author juliankang4 opened on Oct 22, 2026

• serve/hermes: a complete second call opener becomes content at EOF

  review-bot
  midagedev/bloomery#636 · Filter by author github-actions opened on Oct 21, 2026

• model: Flint (flint, 256 experts, sliding-window attention with sinks)

  model
  midagedev/bloomery#632 · Filter by author midagedev opened on Oct 21, 2026

• gate-gpu-lib opens the card without the gate lock; gated items on the same card fail to allocate

  tools
  midagedev/bloomery#630 · Filter by author midagedev opened on Oct 21, 2026

• Raon-1-mini: a verify window of 2 rows costs 10.9 ms more than one row

  enhancement
  midagedev/bloomery#617 · Filter by author midagedev opened on Oct 21, 2026
  1

• gpu: head-128 safe append APIs permit duplicate cache writers

  review-bot
  midagedev/bloomery#614 · Filter by author github-actions opened on Oct 21, 2026

• Compose a model's body from common blocks read off its file; refuse a missing block by name

  enhancementtracking
  midagedev/bloomery#609 · Filter by author midagedev opened on Oct 21, 2026
  3

• Raon-1-mini: DFlash2 speculative draft (0.3.8)

  tracking
  midagedev/bloomery#601 · Filter by author midagedev opened on Oct 21, 2026

• Raon-1-mini-preview support (in progress)

  new-modeltracking
  midagedev/bloomery#598 · Filter by author midagedev opened on Oct 21, 2026

• gate-refset: the hf-qwen4exp-nvfp4 sets are not on the maintainer's box

  midagedev/bloomery#596 · Filter by author midagedev opened on Oct 21, 2026
  1

• mac: run bloomery on Apple Silicon (native host build, CPU path, Metal backend)

  enhancementhardwareno-boxtracking
  midagedev/bloomery#593 · Filter by author midagedev opened on Oct 21, 2026
  7

• serve: a long prompt yields to the busy streams on card-resident seats (wire Engine::prompt_quantum)

  enhancementserve
  midagedev/bloomery#583 · Filter by author midagedev opened on Oct 21, 2026

• serve: the prompt cache keeps a state at every message boundary, its states sharing positional rows

  enhancementserve
  midagedev/bloomery#582 · Filter by author midagedev opened on Oct 21, 2026

• Draft walk: lift the remaining per-model draft types into one owner

  midagedev/bloomery#581 · Filter by author midagedev opened on Oct 21, 2026
  1

• gpu/rope_neox: safe head-256 launcher permits duplicate cache writers

  review-bot
  midagedev/bloomery#580 · Filter by author github-actions opened on Oct 21, 2026
  3

• #565 follow-ups on the maintainer side: QROWS card columns, BF16 as shipped, V4.1 gates for round 4

  gpuplacementtracking
  midagedev/bloomery#570 · Filter by author midagedev opened on Oct 20, 2026

• residency: a warm profile that keeps its decode gain without the first-prompt loss, on by default

  enhancementgputracking
  midagedev/bloomery#569 · Filter by author midagedev opened on Oct 20, 2026

• NVIDIA FP4-Local checkpoints on DGX Spark (TP=1, TP=2): format, model and speed plan in rounds

  dgx-sparkgpuhardwaremodelplacementtracking
  midagedev/bloomery#565 · Filter by author juliankang4 opened on Oct 20, 2026
  4

• Schemajump: fewer engine calls per constrained answer, and the Clef-Flash comparison

  servetracking
  midagedev/bloomery#532 · Filter by author midagedev opened on Oct 20, 2026
  1

  

• tools: gate keys per recipe from what its bins reach, not whole crates

  toolstracking
  midagedev/bloomery#529 · Filter by author midagedev opened on Oct 20, 2026
  2

• To tidy (not this release): run-time size flags match llama-server and never shrink a set value

  servetracking
  midagedev/bloomery#380 · Filter by author midagedev opened on Oct 16, 2026

• serve: the Qwen3 seats never cut a prompt at the message start, so a shared system prompt is never reused

  serve
  midagedev/bloomery#378 · Filter by author juliankang4 opened on Oct 16, 2026