mw00/project-maya

View on GitHub

Project Maya - run GLM-5.3-Flash (321B MoE) and DeepSeek-V4.1-Flash-1017 (748B MoE) privately on your own GPU(s), from one card to 16, NVIDIA or AMD, Linux or Windows. Engine, server, dashboard, OpenAI- and Anthropic-compatible API, quant tools.

Is mw00/project-maya beginner-friendly?

In the last 30 days, mw00/project-maya merged 139 of 378 pull requests from outside contributors, and its maintainers usually reply within 1 day. 1 beginner-friendly issue is open now.

Stars: 1.9k
Forks: 241
Open beginner issues: 1
Indexed issues: 171
Avg merge: 23h 7m
Merged PRs (30d): 178

Dominant language: C++
License: MIT
Last GitHub push: Oct 25, 2026
Latest indexed: Oct 26, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

How to contribute to mw00/project-maya

There is no contributing guide, so start from the README and a few recently merged pull requests to see how changes are proposed.
Your contributions will be published under the project's MIT license.
Pick the open beginner-friendly issue below and comment that you want to work on it before you start.

Beginner friendly Everything

Open All

Issues somebody may already be working on are listed last. List everything by date

DS4_AR_P2P never engages: bar1_bytes() reads the sysfs resource1 file size (256 MiB) instead of BAR1 (32 GiB)
Open
- Is something wrong?
Difficulty 2/5 1-3 hours Newbie friendliness 72/100

mw00/project-maya#798 · Opened Nov 2, 2026

Maintainers usually reply within 1 day
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
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 15/100

mw00/project-maya#797 · 1 comment · Opened Nov 2, 2026

Maintainers usually reply within 1 day
DS4_INDEX_TC on RDNA 3: the indexer's integer scores through WMMA (gfx11)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 10/100

mw00/project-maya#796 · 1 comment · Opened Nov 2, 2026

Maintainers usually reply within 1 day
The request failed cudaMalloc 0 MiB on CUDA0 failed
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 22/100

mw00/project-maya#794 · 9 comments · Opened Nov 2, 2026

Maintainers usually reply within 1 day
Maya text dashboard GPU stats jumping back-forth (cosmetic)
Open
- Is something wrong?
Difficulty 2/5 1-3 hours Newbie friendliness 68/100

mw00/project-maya#793 · Opened Nov 2, 2026

Maintainers usually reply within 1 day
Support refusal vectors
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 18/100

mw00/project-maya#792 · 2 comments · Opened Nov 2, 2026

Maintainers usually reply within 1 day
Hip and Cuda do not work together
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 15/100

mw00/project-maya#789 · 2 comments · Opened Nov 2, 2026

Maintainers usually reply within 1 day
Windows + HIP: the dashboard shows the wrong GPUs when HIP doesn't see every AMD adapter (DXGI order vs HIP numbering)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 22/100

mw00/project-maya#780 · 1 comment · Opened Nov 2, 2026

Maintainers usually reply within 1 day
Windows, one 20 GB GPU: --parallel 2 gives two real sessions (+14% aggregate); --parallel 3 collapses
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 12/100

mw00/project-maya#779 · 2 comments · Opened Nov 2, 2026

Maintainers usually reply within 1 day
add high context for glm
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 25/100

mw00/project-maya#769 · Opened Nov 1, 2026

Maintainers usually reply within 1 day
hardcoded logit_bias specified in server command line
Open
- Is something wrong?
Difficulty 3/5 1-2 days Newbie friendliness 35/100

mw00/project-maya#768 · 6 comments · Opened Nov 1, 2026

Maintainers usually reply within 1 day
[Speed Report + Unfilled VRAM issue] DS4.1 1017 flash UD_Q8_K_XL (4x rtx 3090 + Threadripper + 504GB DDR4)
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 14/100

mw00/project-maya#766 · 3 comments · Opened Nov 1, 2026

Maintainers usually reply within 1 day
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)
Open
- Is something wrong?
Difficulty 1/5 Under an hour Newbie friendliness 22/100

mw00/project-maya#757 · Opened Nov 1, 2026

Maintainers usually reply within 1 day
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
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 22/100

mw00/project-maya#755 · 1 comment · Opened Nov 1, 2026

Maintainers usually reply within 1 day
Speed report: AC922 (2x POWER9, 6x V100-16GB) - 69.6 tok/s decode, 1M context, Maya-S24 built from source on ppc64le
Open
- Is something wrong?
Difficulty 4/5 Half a day Newbie friendliness 12/100

mw00/project-maya#754 · 2 comments · Opened Nov 1, 2026

Maintainers usually reply within 1 day
[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
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 15/100

mw00/project-maya#750 · 2 comments · Opened Nov 1, 2026

Maintainers usually reply within 1 day
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)
Open
- Is something wrong?
Difficulty 2/5 Under an hour Newbie friendliness 8/100

mw00/project-maya#744 · Opened Nov 1, 2026

Maintainers usually reply within 1 day
Request new Exl3 quant: DSV4.1-Flash-1017-SAGE-EXL3-1.62bpw
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 18/100

mw00/project-maya#740 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
DeepSeek: make the 8 GiB system-RAM budgeting margin configurable
Open
- Is something wrong?
Difficulty 3/5 1-2 days Newbie friendliness 45/100

mw00/project-maya#739 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
[bug] DS4 HIP: reloading DeepSeek fails while copying CPU-lane experts from model files
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 18/100

mw00/project-maya#735 · 2 comments · Opened Oct 31, 2026

Maintainers usually reply within 1 day
TP on two cards: the second card keeps ~1.5-2 GB free — is asymmetric expert residency possible, or is that expected?
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 15/100

mw00/project-maya#734 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
DeepSeek: the automatic layer split balances streamed bytes, not time, so short prompts wait on the slow-link GPUs
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 25/100

mw00/project-maya#733 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
DeepSeek: v1.1.42's slow-link ring rule makes fresh prompt reads ~8% slower on 3090s on PCIe 3.0 x8 (#609)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 34/100

mw00/project-maya#731 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
[bug] loop-guard retry overflows context by (note - slack) -> "reply was cut short" stubs
Open
- Is something wrong?
Difficulty 3/5 Half a day Newbie friendliness 66/100

mw00/project-maya#727 · 3 comments · Opened Oct 31, 2026

Maintainers usually reply within 1 day
[bug]Codex CLI 0.160.0 fails against Maya's Responses API
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 25/100

mw00/project-maya#726 · 2 comments · Opened Oct 31, 2026

Maintainers usually reply within 1 day
Let the token-savings ledger live somewhere other than next to --config
Open
- Is something wrong?
Difficulty 3/5 Half a day Newbie friendliness 50/100

mw00/project-maya#724 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
maya-mm3 GGUF: a changed GPU list silently invalidates the prefill calibration (tune key includes the GPUs) + 2x R9700 chunk measurements
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 35/100

mw00/project-maya#721 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
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
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 12/100

mw00/project-maya#718 · Opened Oct 31, 2026

Maintainers usually reply within 1 day
DeepSeek on Linux: Engram reader threads default to 8 on a 16-thread host, and cold prompts read ~21 % faster with 32
Open
- Is something wrong?
Difficulty 2/5 1-3 hours Newbie friendliness 45/100

mw00/project-maya#716 · 2 comments · Opened Oct 31, 2026

Maintainers usually reply within 1 day
DS4_COPY_GB doesn't appear to work with EXL3
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 48/100

mw00/project-maya#708 · Opened Oct 30, 2026

Maintainers usually reply within 1 day
Qwen engine: after a client abandons a prompt read, every later request fails with "cancelled" until a restart (one-line fix: clear err per request)
Open
- Is something wrong?
Difficulty 2/5 1-3 hours Newbie friendliness 86/100

mw00/project-maya#703 · 2 comments · Opened Oct 30, 2026

Maintainers usually reply within 1 day
Sampled decode (temperature > 0) is ~3.5x slower in v1.1.37+ while greedy decode is unaffected
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 45/100

mw00/project-maya#699 · 3 comments · Opened Oct 30, 2026

Maintainers usually reply within 1 day
DeepSeek: the PCIe lane's decode check turned streaming off on an RTX 3090 where it is ~20-40% faster, and the times file keeps it for every later start
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

mw00/project-maya#697 · Opened Oct 30, 2026

Maintainers usually reply within 1 day
[AMD / v1.1.41 / #592] Dual R9700 exact peer transport works; faster TP2 prefill, layer split still faster for Maya-S abliterated
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

mw00/project-maya#694 · 1 comment · Opened Oct 30, 2026

Maintainers usually reply within 1 day
ds4 v1.1.40 TP3 with most experts outside VRAM: the last-user-turn restore point adds 1-4 s to most agent turn reads (only the tail point is checked)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 45/100

mw00/project-maya#693 · 1 comment · Opened Oct 30, 2026

Maintainers usually reply within 1 day
DeepSeek at 1M: queued concurrent requests end with a `window ring` fault; loading takes 20-25 min so `TimeoutStartSec=1200` loops forever
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 48/100

mw00/project-maya#690 · 5 comments · Opened Oct 30, 2026

Maintainers usually reply within 1 day
maya-mm3 (MiniMax M3.1 · Maya-XS) crashes with 0xC0000005 on the first request on Windows (RTX 4070 Ti SUPER + RTX 5070 Ti)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 43/100

mw00/project-maya#672 · 3 comments · Opened Oct 29, 2026

Maintainers usually reply within 1 day
Full GLM-5.3 (glm-dsa) on the GLM engine
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 30/100

mw00/project-maya#661 · 1 reaction · Opened Oct 29, 2026

Maintainers usually reply within 1 day
Feature request: cross-model KV-cache handoff (SwaT, arXiv 2610.15095) - start with KV export/import
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 28/100

mw00/project-maya#655 · 1 comment · 1 reaction · Opened Oct 29, 2026

Maintainers usually reply within 1 day
Maya-M on 2x RTX 2080 Ti (sm_75): non-finite router score at layer 17/18 on every request
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 45/100

mw00/project-maya#651 · 4 comments · Opened Oct 29, 2026

Maintainers usually reply within 1 day
v1.1.35/36: GLM prefill -43/-53% and DeepSeek prefill -47% with a real (large) max_tokens on a 20 GB card
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 35/100

mw00/project-maya#648 · 4 comments · Opened Oct 29, 2026

Maintainers usually reply within 1 day
Qwen3.8-Flash-Next (CYBER-FROST) decodes ~40% slower on Maya than on Strata 0.2.7 with the same weights: the Qwen engine is still 0.1.27 (last synced from Strata 0.1.41, v1.0.13)
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

mw00/project-maya#604 · Opened Oct 28, 2026

Maintainers usually reply within 1 day
MiniMax-M3.1-Flash at its original precision on maya-mm3: Unsloth's "UD-Q8_K_XL" GGUF holds the original NVFP4 experts (no 8-bit tensors at all), and AMD's MXFP4 checkpoint uses the format maya-ds4 already runs
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 25/100

mw00/project-maya#559 · Opened Oct 27, 2026

Maintainers usually reply within 1 day
Not fill the full VRAM with DSpark on 4x rtx 3090 (Minimax 3.1 flash + DeepSeek V4.1 flash 1017)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 35/100

mw00/project-maya#556 · 4 comments · Opened Oct 27, 2026

Maintainers usually reply within 1 day
ds4: the drafter's routes are never counted in the expert profile, so its experts are always placed "as room goes" (TP 2 + DSpark leaves ~4.5 GiB idle on the second GPU)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 45/100

mw00/project-maya#501 · 2 comments · Opened Oct 26, 2026

Maintainers usually reply within 1 day
Speed report: 10x RTX 3090 (TP2) + 64 GB — DeepSeek-V4.1-Flash-1017 EXL3 2.9 bpw: 97 tok/s decode, 1,900–2,390 tok/s prefill, 1M context (v1.1.28)
Open
- Is something wrong?
Difficulty 1/5 Under an hour Newbie friendliness 35/100

mw00/project-maya#494 · 5 comments · Opened Oct 25, 2026

Maintainers usually reply within 1 day
DeepSeek-V4.1-Flash Maya-XS on Strix Halo (gfx1151, Linux): ~0.1 tok/s decode, q8_mv_kernel and iq1 expert kernels ~100x too slow
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 35/100

mw00/project-maya#491 · 1 comment · Opened Oct 25, 2026

Maintainers usually reply within 1 day
Experimental: compress-then-correct (lossy expert routing + exact re-decode) for DeepSeek on V100, feedback wanted
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 25/100

mw00/project-maya#443 · Opened Oct 24, 2026

Maintainers usually reply within 1 day
DeepSeek on a 20 GB Ada (Windows): v1.1.26 prefill 3.7-5.4x slower than v1.1.25 (decode unchanged)
Open
- Is something wrong?
Difficulty 4/5 3-5 days Newbie friendliness 44/100

mw00/project-maya#438 · 9 comments · Opened Oct 24, 2026

Maintainers usually reply within 1 day
A real multi-user lab for DeepSeek on Maya: admin staff, programmers and research data work all sharing one server — volunteers ready to stress-test
Open
- Is something wrong?
Difficulty 5/5 Over a week Newbie friendliness 35/100

mw00/project-maya#437 · 4 comments · 2 reactions · Opened Oct 23, 2026

Maintainers usually reply within 1 day

mw00/project-maya

Is mw00/project-maya beginner-friendly?

How to contribute to mw00/project-maya

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Is something wrong?

Get new issues in your inbox