Issues

• Speed report: 2x RTX PRO 6000 Blackwell 96 GB, DeepSeek-V4.1-Flash UD-Q8_K_XL at 1M context (v1.1.51)

  mw00/project-maya#819 · Filter by author markrussi opened on Nov 3, 2026

• Windows, 2x A40: GLM Maya-L 2-GPU split stops at "the pinned RAM tier did not allocate" — part 2 sees 0.50 GB pin room after part 1 pins 11.78 GB (v1.1.26+; v1.1.25 starts)

  mw00/project-maya#817 · Filter by author yjyz1011 opened on Nov 3, 2026

• Speed report: MiniMax M3.1 Maya-S on 2x V100 16GB + dual E5-2696 v4, 128GB DDR3 (v1.1.51, no-pin-pool)

  mw00/project-maya#816 · Filter by author c200762620-max opened on Nov 3, 2026

• Speed report: DeepSeek V4.1 Maya-S on 2x V100 16GB + dual E5-2696 v4, 128GB DDR3 (v1.1.51)

  mw00/project-maya#815 · Filter by author c200762620-max opened on Nov 3, 2026

• DS4_AR_P2P on AMD: bar1_bytes() still reads 0 MiB (VRAM BAR is table line 0), and with detection fixed the engine hangs

  mw00/project-maya#809 · Filter by author Jlauti opened on Nov 3, 2026

• DeepSeek: "reasoning_effort": "4096" in the config stops Maya after the model loads; the token cap is "thinking_budget"

  mw00/project-maya#807 · Filter by author talisp opened on Nov 3, 2026
  1

• DeepSeek EXL3: four gaps when loading third-party V4.1-Flash-1017 quants (type-ordered shards, BF16 head, mcg MTP experts, INT4 Engram)

  mw00/project-maya#806 · Filter by author olobarius opened on Nov 3, 2026

• EXL3 quants of DeepSeek-V4.1-Flash on Maya: how close are they to the original?

  mw00/project-maya#803 · Filter by author pepuscz opened on Nov 3, 2026
  2

• Speed report: DeepSeek-V4.1-Flash on Strix Halo + 2x RX 7900 XTX: DS4_INDEX_TC on RDNA 3 (PR #796) reads a 480K prompt 21 % faster, a 130K prompt 6 % faster - Maya-XS v2, v1.1.50, 512K context

  mw00/project-maya#797 · Filter by author pepuscz opened on Nov 2, 2026
  1

• DS4_INDEX_TC on RDNA 3: the indexer's integer scores through WMMA (gfx11)

  mw00/project-maya#796 · Filter by author pepuscz opened on Nov 2, 2026
  1

• The request failed cudaMalloc 0 MiB on CUDA0 failed

  mw00/project-maya#794 · Filter by author wolfwoof88 opened on Nov 2, 2026
  10

• Support refusal vectors

  mw00/project-maya#792 · Filter by author fukc-gihtub opened on Nov 2, 2026
  2

• Hip and Cuda do not work together

  mw00/project-maya#789 · Filter by author Fringe210 opened on Nov 2, 2026
  3

• Windows + HIP: the dashboard shows the wrong GPUs when HIP doesn't see every AMD adapter (DXGI order vs HIP numbering)

  mw00/project-maya#780 · Filter by author Meistertech-JP opened on Nov 2, 2026
  2

• Windows, one 20 GB GPU: --parallel 2 gives two real sessions (+14% aggregate); --parallel 3 collapses

  mw00/project-maya#779 · Filter by author pietroMastro92 opened on Nov 2, 2026
  2

• add high context for glm

  mw00/project-maya#769 · Filter by author TingTingin opened on Nov 1, 2026

• hardcoded logit_bias specified in server command line

  mw00/project-maya#768 · Filter by author vbooka1 opened on Nov 1, 2026
  8

• 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
  3

• 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
  2

• 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

• 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

• 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