Issues

• Speed report: RX 7900 XTX 24 GB (gfx1100, ROCm 10.2), DeepSeek-V4.1-Flash Maya-XS 64K + DSpark - 18.4 tok/s decode, 395 tok/s prefill (v1.1.46)

  mw00/project-maya#757 · Filter by author Cayaf89 opened on Nov 1, 2026

• Speed report: RTX 5070 12 GB + EPYC 7302 (4-channel DDR4), Maya-S24 - 16.6 -> 24.2 tok/s decode, 323 -> 428 prefill; DFlash2 ~5 tok/s on one GPU + one socket

  mw00/project-maya#755 · Filter by author hugou74130 opened on Nov 1, 2026

• Speed report: AC922 (2x POWER9, 6x V100-16GB) - 69.6 tok/s decode, 1M context, Maya-S24 built from source on ppc64le

  mw00/project-maya#754 · Filter by author jefflaichunyin opened on Nov 1, 2026
  1

• DeepSeek: with MAYA_DS4_READ_ORDER=arrival, a running answer gets one step per long read, so back-to-back reads starve it

  mw00/project-maya#751 · Filter by author mgabor3141 opened on Nov 1, 2026

• [bug] Maya-XS: every request fails with "cudaEventRecord(d.hc_done, d.st2): an illegal memory access" on 4x RTX 2080 Ti (sm_75) — v1.1.39 and v1.1.46

  mw00/project-maya#750 · Filter by author Fox-B1ack opened on Nov 1, 2026

• DeepSeek: after the thinking budget's soft close the engine reads the whole thinking again (59 s for 32K tokens)

  mw00/project-maya#746 · Filter by author tobitege opened on Nov 1, 2026

• Windows: DeepSeek vision works on v1.1.45 with a pack rebuilt by tools/ds4/vision_pack.py (v1.1.44 crashed the CUDA context on every picture)

  mw00/project-maya#744 · Filter by author pietroMastro92 opened on Nov 1, 2026

• tools/ds4/slot_keep_test.py does not run on Windows: select() on a pipe fails with WinError 10093 (PeekNamedPipe fallback included)

  mw00/project-maya#743 · Filter by author pietroMastro92 opened on Nov 1, 2026

• Request new Exl3 quant: DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw

  mw00/project-maya#740 · Filter by author BahamutRU opened on Oct 31, 2026

• DeepSeek: make the 8 GiB system-RAM budgeting margin configurable

  mw00/project-maya#739 · Filter by author davedgd opened on Oct 31, 2026

• Windows 11 speed report: MiniMax-M3.1-Flash Maya-S on RTX PRO 6000 Blackwell 96 GB — 48.6 tok/s decode, 757 tok/s prefill (8K)

  mw00/project-maya#738 · Filter by author Andrey0710 opened on Oct 31, 2026

• Windows 11 speed report: DeepSeek-V4.1-Flash Maya-S on RTX PRO 6000 Blackwell 96 GB — 72.3 tok/s decode, 1757 tok/s prefill (8K)

  mw00/project-maya#737 · Filter by author Andrey0710 opened on Oct 31, 2026
  6

• [bug] DS4 HIP: reloading DeepSeek fails while copying CPU-lane experts from model files

  mw00/project-maya#735 · Filter by author Biotrioo opened on Oct 31, 2026
  2

• TP on two cards: the second card keeps ~1.5-2 GB free — is asymmetric expert residency possible, or is that expected?

  mw00/project-maya#734 · Filter by author tonykee opened on Oct 31, 2026

• DeepSeek: the automatic layer split balances streamed bytes, not time, so short prompts wait on the slow-link GPUs

  mw00/project-maya#733 · Filter by author mgabor3141 opened on Oct 31, 2026

• DeepSeek: v1.1.42's slow-link ring rule makes fresh prompt reads ~8% slower on 3090s on PCIe 3.0 x8 (#609)

  mw00/project-maya#731 · Filter by author mgabor3141 opened on Oct 31, 2026

• Windows 11 speed report: RTX PRO 6000 Blackwell 96 GB, Maya-L — 48.8 tok/s decode, 3810 tok/s prefill (8K)

  mw00/project-maya#730 · Filter by author Andrey0710 opened on Oct 31, 2026

• [bug] TUI stuck at "Loading the model" while engine is loaded and serving

  mw00/project-maya#729 · Filter by author sexavier opened on Oct 31, 2026

• [bug] loop-guard retry overflows context by (note - slack) -> "reply was cut short" stubs

  mw00/project-maya#727 · Filter by author sexavier opened on Oct 31, 2026
  3

• [bug]Codex CLI 0.160.0 fails against Maya's Responses API

  mw00/project-maya#726 · Filter by author zhouxiaoxing91 opened on Oct 31, 2026
  2

• Let the token-savings ledger live somewhere other than next to --config

  mw00/project-maya#724 · Filter by author gabacool opened on Oct 31, 2026

• Speed report: GLM-5.3-Flash UD-IQ3_XXS on 2x RTX 4060 Ti 16 GB (PCIe 3.0 x8), and a few side findings

  mw00/project-maya#723 · Filter by author oisee opened on Oct 31, 2026

• maya-mm3 GGUF: a changed GPU list silently invalidates the prefill calibration (tune key includes the GPUs) + 2x R9700 chunk measurements

  mw00/project-maya#721 · Filter by author eugenrepp opened on Oct 31, 2026

• Speed & benchmark report: Maya-S DeepSeek-V4.1-Flash-1017 on 2× RTX PRO 5000 72GB, engine v1.1.41 with --parallel 2 — 14h real agent-traffic concurrency soak

  mw00/project-maya#719 · Filter by author realsquall1980-art opened on Oct 31, 2026

• DeepSeek PROMPT_SKIP=2: the 128-row replay attends to window rows from before the replay start (left by earlier requests); masked, it is +37 % on 4K prompts here

  mw00/project-maya#718 · Filter by author SeriousRising opened on Oct 31, 2026