mw00/project-maya v1.1.20

Project Maya v1.1.20

on GitHub

Stop now stops at once, DeepSeek-V4.1-Flash decodes ~6.5% faster on a Strix Halo and fits its GPU memory by itself, Windows starts Maya again from START-MAYA.bat, the port can be changed from the dashboard, and a round of community work.

Update: the dashboard's About > Update, or git pull, then ./maya.sh (Windows: START-MAYA.bat). An install of v1.1.19 keeps its files and works as it is.

What's new

• Stop stops, even while a long prompt is read. Pressing Stop while DeepSeek-V4.1-Flash read a long prompt let it read to the end first (a minute or more on a 28K prompt). It now stops within seconds, and the conversation before it stays loaded, so your next message does not read it again.
• DeepSeek: faster drafting, ~6.5% more tokens a second with the DSpark drafter on a Strix Halo, the same answers: a draft whose earlier drafts would likely be rejected is no longer verified (MAYA_DS4_DSPARK_CUM, 0.5; 0 = off).
• DeepSeek on a Strix Halo / Windows APU: the GPU's memory, measured. Windows' HIP counts an APU's GPU memory plus shared system memory (172 GB on a 160 GB Gorgon Halo); DeepSeek now keeps the model within the dedicated part, as GLM already did - no --vram-experts needed, the full speed by itself.
• Windows: START-MAYA.bat works again (#315, #322): since v1.1.18 its screen showed the setup and then did nothing; thanks to @kelheor for pinning down the cause.
• Dashboard: Settings > Port moves Maya to another port; it starts again there and the page follows.
• Setup: you choose the drafter, or skip the speed test. Before the 10-30 minute measurement the setup asks: measure, use the drafter, or no drafter (--drafting measure|on|off); the choice is kept.
• Server: a request whose client went away leaves the queue at once (#314, thanks @jettoblack); with --parallel each request reports its own timings and cached tokens (#310, thanks @pepuscz).
• DeepSeek's EXL3 quants on the Tesla V100. The setup now offers the EXL3 quants from the V100 on (text; pictures from the RTX 30 / A100 on): the prompt path runs portable kernels there, and the experts take the newer kernels where the older ones did not cover DeepSeek's width.
• EXL3 on AMD GPUs: the prompt path's expert kernels (#298 by @LostVector) - the groundwork for running EXL3 quants on a Strix Halo.
• GLM: fused prompt experts on the RX 7900 XT / XTX, +5-13% prompt speed (#304 by @Cass67); experimental AMD Instinct MI100 support (#245 by @nsaxploit); batched decode settings for several conversations, opt-in (#323, #324, #325 by @somnergy); prompt-path tools and the MMQ scratch's room counted (#284, #286, #289 by @gopinath87607); and .gitignore (#308 by @rkarp).

Checked

• DeepSeek-V4.1-Flash: Stop on an 18K prompt on a Strix Halo and on a V100 (the same tokens after it); the DSpark cut on and off on a Strix Halo; the GGUF and EXL3 kernel tests on a Tesla V100 and a Radeon 8065S.
• GLM-5.3-Flash: greedy answers unchanged on a V100 beside the merged PRs.
• Builds: Linux CUDA (Volta), Windows CUDA, Windows HIP (Radeon 8065S); the setup's, the server's and the dashboard's tests.