All pull requests

Pull requests

• RDNA4, 9060XT, GFX1200 Support including: tuning table, Context steps and tev-320b model download

  #909·Neresco opened Nov 7, 2026·

  ·

• DeepSeek: a kept shared start left in the model is not a conversation (a STOP's read made the next turn re-read its whole prompt)

  #906·markrussi opened Nov 7, 2026Contributor·

  ·

• DeepSeek EXL3 with a helper: the DSpark blocks' experts in the helper's pool too (DS4_HELPER_DRAFT)

  #904·pepuscz opened Nov 7, 2026Contributor·

  ·

• DeepSeek: read a chat's next turn ahead after a thinking answer (MAYA_DS4_PREREAD, off by default, the follow-up's output changes slightly)

  #902·danexpo opened Nov 7, 2026Contributor·

  ·

• GLM: an optional residual-direction ablation (carved out of #642)

  #901·ClintMoody opened Nov 7, 2026·

  ·

• ds4: pool layout-sized GGUF expert RAM cache backing

  #897·cgi-yf opened Nov 6, 2026·

  ·

• maya-mm3: a helper GPU for the CPU lane's experts (MAYA_MM3_HELPER), per-GPU staging copy timing

  #896·GoKu-Gear6 opened Nov 6, 2026Contributor·

  ·

• Windows: fix build log garbled text (mojibake) and suppress MSVC C4819 warnings

  #894·Cayahuanca opened Nov 6, 2026·

  ·

• Dashboard: show CPU temperature beside GPU temperatures

  #893·org-tekeli-borisp opened Nov 6, 2026Contributor·

  ·

• A Chat Completions request's own thinking_budget: that request alone gets its cap, the config's stays every other request's

  #891·linus-aegu opened Nov 6, 2026·

  ·

• feat(hip/gfx906): activate 4-helper EP DMA pipelines on MiniMax-M3.1 and calibrate DSpark K=8 horizon on MI50

  #889·nicholas631 opened Nov 6, 2026Contributor·6 tasks done·

  ·

• Fix quantized Engram rows in GGUF reference

  #885·midhatn opened Nov 6, 2026·

  ·

• Fair slices under --parallel: one waiting request takes one place, not every runner's

  #883·HTsulfuric opened Nov 6, 2026·

  ·
  1

• serve: MAYA_PAR_MAX caps --parallel before the engine starts (20 GB cards: 3 conversations leave the prompt path no room)

  #880·pietroMastro92 opened Nov 6, 2026Contributor·

  ·
  1

• GLM: opt-in CUDA+HIP expert helper for mixed-vendor GPUs

  #879·RenWuJun opened Nov 6, 2026·

  ·
  1

• Allow peer calibration with the implicit CUDA AR2 default

  #876·midhatn opened Nov 5, 2026·

  ·
  1

• WIP: fail closed on GGUF prompt-row errors

  #872·midhatn opened Nov 5, 2026·

  ·

• DeepSeek EXL3 in place: DS4_EXL3_LEND lends the coldest VRAM slots

  #867·pepuscz opened Nov 5, 2026Contributor·

  ·
  1

• Support mmap Q8_0 Qwen n-gram tables

  #863·midhatn opened Nov 5, 2026·

  ·

• DeepSeek: add native GBNF constrained decoding

  #856·NektoDron opened Nov 4, 2026·

  ·
  2

• serve: add opt-in CUDA checkpoint sleep

  #855·NektoDron opened Nov 4, 2026·

  ·
  2

• DeepSeek: load exactly representable F16 embeddings

  #854·NektoDron opened Nov 4, 2026·

  ·
  2

• web: add English and Simplified Chinese interface localization

  #847·oscarlius opened Nov 4, 2026Contributor·

  ·
  2

• serve: close Linux telemetry files explicitly

  #829·midhatn opened Nov 4, 2026·

  ·

• ds4: add opt-in cache-aware expert substitution for decode

  #821·LostVector opened Nov 3, 2026Contributor·

  ·
  1