Issues

• Speed report: DeepSeek EXL3 2.25 bpw on 1x RX 7900 XTX + Radeon 8060S, 1M context: 45.6 tok/s decode, 954 tok/s prefill at 160K (v1.1.56 + #904)

  speed-report
  mw00/project-maya#908 · Filter by author pepuscz opened on Nov 7, 2026

• Illegal memory access when running modified DeepSeek V4.1 Flash 1017

  bug
  mw00/project-maya#907 · Filter by author threefcata opened on Nov 7, 2026

• Speed report: DeepSeek EXL3 2.25 bpw on 2x RX 7900 XTX + Radeon 8060S, 1M context: 59.6 tok/s decode, 1,656 tok/s prefill at 160K (v1.1.56 + #867 + #904)

  speed-report
  mw00/project-maya#905 · Filter by author pepuscz opened on Nov 7, 2026

• Possible bug experienced while trying to get Maya-S24 run on dual RX 9070 16GB for the first time

  mw00/project-maya#899 · Filter by author bob-u opened on Nov 6, 2026

• Qwen 3.8 Flash Next long-context performance is much slower than Strata on the same hardware

  bug
  mw00/project-maya#898 · Filter by author talshebek opened on Nov 6, 2026

• RDNA2 (gfx1030) data point: 2× RX 6800, GLM-5.3 Maya-L, 1M context — ~13.5 tok/s decode / ~290 tok/s prefill (v1.1.55)

  mw00/project-maya#895 · Filter by author digitalhomesteader opened on Nov 6, 2026

• DeepSeek: loop-guard edit re-prefills the whole conversation (103k tokens); GLM-style checkpoint recovery would bound it

  mw00/project-maya#892 · Filter by author pjgmobile opened on Nov 6, 2026

• Simple --bench test of Maya-XS running on a Xeon e5-2690v4 with 128GB RAM and 2 x 32GB R9700s

  mw00/project-maya#890 · Filter by author imaginitivename opened on Nov 6, 2026

• 2x RTX 4090 (PCIe, no NVLink): DS4_TP_GGUF_UNEVEN=1 engages AR2 on a GGUF, but prompt reading drops 2-5x

  mw00/project-maya#888 · Filter by author manzonif opened on Nov 6, 2026

• Simple --bench test of Maya-S running on a Xeon e5-2690v4 with 128GB RAM and 2 x 32GB R9700s

  mw00/project-maya#887 · Filter by author imaginitivename opened on Nov 6, 2026

• DeepSeek EXL3 (turboderp 2.68bpw): vision tower stays OFF because the quant ships it in F16 and its tensors win over DeepSeek’s BF16 shard 1

  mw00/project-maya#886 · Filter by author anbow1 opened on Nov 6, 2026

• Windows, 20 GB + 255 GB RAM, v1.1.55: DS4_LENT_RAM_GB=64 fails every request with HTTP 400 (0 s short prompt, 77 s long one, no reason in the log)

  mw00/project-maya#884 · Filter by author pietroMastro92 opened on Nov 6, 2026

• Fair slices with --parallel N: one waiting request cuts all N running answers, not one

  mw00/project-maya#882 · Filter by author HTsulfuric opened on Nov 6, 2026

• DeepSeek --ablate: the L22 refusal direction works with thinking, not with reasoning_effort none (PRC framing stays); KL ~ base

  mw00/project-maya#877 · Filter by author olobarius opened on Nov 5, 2026

• Two-rig layer split for GLM-5.3-Flash (two X79 rigs with 4× RTX 3060 each, over 1 GbE): interest in upstreaming?

  mw00/project-maya#874 · Filter by author 0xluk opened on Nov 5, 2026
  1

• Speed report: DeepSeek EXL3 2.25 bpw on 2x RX 7900 XTX + Radeon 8060S, v1.1.54 + seven PRs

  mw00/project-maya#871 · Filter by author pepuscz opened on Nov 5, 2026
  1

• [regression] v1.1.53: GLM-5.3 UD-IQ4_XS fails with "invalid expert ID or non-finite score" (works in v1.1.51)

  mw00/project-maya#860 · Filter by author githug2024 opened on Nov 5, 2026
  8

• Metal: maya-metal and maya-metal-mm3 produce garbage tokens on an M4 Max (Apple9) - the same packs answer correctly under llama.cpp

  mw00/project-maya#843 · Filter by author mikog8tor opened on Nov 4, 2026

• DeepSeek --tp 4: all GPUs stuck at 100% busy while idle since v1.1.52 (still in v1.1.53), DS4_TP_SHARES=off clears it

  mw00/project-maya#841 · Filter by author Cass67 opened on Nov 4, 2026
  4

• Uneven VRAM allocation with 8x RTX309 using DepSeek-v4.1-flash (EXL3- 2.89 BPW)

  mw00/project-maya#833 · Filter by author vikcious opened on Nov 4, 2026
  5

• DS4 on 8× MI50 (gfx906): solo-8 deadlock at first prefill chunk + decode regression at all layouts — v1.1.51/52

  mw00/project-maya#831 · Filter by author abluesheep opened on Nov 4, 2026
  3

• Windows, RTX 4090 + 64 GB RAM, v1.1.52: GLM Maya-XS decode plateaus at ~24 tok/s with DFlash2 (93% accepted) - expert cache ~75% hit, misses fall through to NVMe

  mw00/project-maya#828 · Filter by author RAFOLIE opened on Nov 4, 2026
  2

• DeepSeek, one GPU, Windows: illegal memory access in cudaMemcpyAsync(c.emb, c.emb_host[slot] ...) on the request after one that ended with length

  mw00/project-maya#827 · Filter by author anbow1 opened on Nov 4, 2026
  1

• 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
  3

• 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
  1