All pull requests

Pull requests

• HIP: build and enable the vision encoder on AMD GPUs

  #345·Ehco opened Oct 21, 2026·

  ·

• DeepSeek-V4.1 tensor parallel on 9 GPUs: prompts ~1.9x and answers ~1.55x faster at --tp 8 than main, the same perplexity (all-reduces off the SMs, the prompt skip on several GPUs, the expert cache's recent slots); ./maya.sh --calibrate tunes --tp

  #344·needmorevram opened Oct 21, 2026Collaborator·

  ·

• ds4: prompt skip across a 2-GPU prefix layer split

  #343·Gazelle221B opened Oct 21, 2026·

  ·

• GLM prompt attention on Volta in one pass (mma.m8n8k4): long prompts +8-9% on a V100

  #335·merbanan opened Oct 21, 2026Contributor·

  ·

• Metal: GLM-5.3-Flash on Apple Silicon (maya-metal), experimental

  #334·hotocoo opened Oct 21, 2026·

  ·
  1

• glm slots: set aside conversations that have pictures too

  #331·Cass67 opened Oct 21, 2026Contributor·

  ·

• Server: fix a blocked FIFO — requests with wrong/duplicate rids never cleared, queue blocked forever

  #330·jettoblack opened Oct 21, 2026·

  ·

• setup: --tp 1 on one GPU so a DeepSeek prompt takes the chunked path

  #320·pietroMastro92 opened Oct 20, 2026·

  ·

• GLM: allow CPU lane retiming to select an all-PCIe plan

  #305·JumpEggKiller opened Oct 20, 2026·

  ·

• perf(ds4): opt-in HIP prefill transfer pruning

  #297·LostVector opened Oct 20, 2026Contributor·

  ·
  2

• glm prefill: each GPU can choose its own chunk size

  #288·gopinath87607 opened Oct 20, 2026Contributor·

  ·
  2

• glm prefill: more memory for a long prompt while it reads

  #283·gopinath87607 opened Oct 20, 2026Contributor·

  ·
  2

• feat(glm): DFMULTI shared multi-token verification, ragged windows, 32-row arena

  #257·btothec opened Oct 17, 2026·4 tasks done·

  ·
  4

• EXL3-sc_4.10bpw in the setup: turboderp's self-calibrated quant at 4.10 bits (8-bit attention and shared experts, the MTP block in its shards)

  #256·needmorevram opened Oct 17, 2026Collaborator·

  ·

• Add native whole video input support

  #248·kabachuha opened Oct 17, 2026·

  ·
  1

  2

• Add Simplified Chinese translations for README and five docs

  #244·chenyb999-zhcn opened Oct 17, 2026Contributor·

  ·

• OPTIMIZATION + NEW DECODE VALUES FOR THE 3XV100, +~25-50%; CPU lane: AVX2 1-token decode kernels, CUDA sm_70 direct signs, and multi-GPU pipeline tuning

  #234·saatvik-viniak opened Oct 16, 2026Contributor·

  ·
  7

• End a silent engine when its GPUs stay idle

  #181·briepala opened Oct 14, 2026·2 tasks done·

  ·
  1

• Calibration: measure more CPU-lane threads than the engine's own (STRATA_GLM_CPU_LANE_MAX)

  #178·Jw88888 opened Oct 14, 2026Contributor·

  ·
  3

• EXL3: serve experts-only community quants (Q8_0 twins, float rows) + native MTP draft block

  #170·waxinz opened Oct 13, 2026Contributor·

  ·
  2

• GLM: balance pinned RAM across expert classes for full Windows RAM residency

  #153·PhysicistLeon opened Oct 13, 2026·

  ·
  4

• Pause a long prompt read for a waiting request and keep what was read (STRATA_FAIR_READ_S)

  #144·DailyDisco opened Oct 12, 2026·

  ·
  6

• Pipeline a short prompt across a split's cards instead of one chunk

  #138·DailyDisco opened Oct 12, 2026·

  ·
  5

• Keep and share cached conversation state

  #126·ksanislo opened Oct 12, 2026Contributor·

  ·
  2

• Several conversations on a layer split: prompts read beside the decode, elastic contexts

  #125·ksanislo opened Oct 12, 2026Contributor·

  ·
  3