MiniMax answers faster, DeepSeek reads copied text faster and keeps its pictures off the GPU, orcarouter's
abliterated DeepSeek in the setup, and a new web UI.
- MiniMax-M3.1-Flash much faster: decoding with the DSpark drafter is about 1.5x as fast as in v1.1.30 on
three V100s and on a Strix Halo, plain decoding about 1.2x. A bug since v1.1.29 left the drafter guessing
without the prompt, so it rarely helped; it now drafts from the whole conversation. Faster dense kernels too,
with the same answers as before. - DeepSeek: copied text up to ~1.6x faster (#540, #543, @needmorevram): when the answer repeats earlier text -
a file sent back with small edits, quoted text - the engine checks the tokens that followed that text instead of
only the drafter's guesses, with and without DSpark. On by default on NVIDIA and AMD (about +27% on an edit on
a Strix Halo);--suffix-draft 0turns it off. - DeepSeek pictures take no VRAM between pictures: the vision tower borrows its working memory only while a
picture is read and gives it back, so about 320 MB more of the model stays on the GPU - as GLM and MiniMax. - orcarouter's abliterated DeepSeek-V4.1-Flash (#520, @needmorevram):
--model deepseekoffers
"orcarouter Abliterated Mix-128G" and "Mix-256G" below Maya's quants, the one that fits the PC recommended. They
use Maya's own n-gram tables (their table is DeepSeek's own, so the higher-precision copy is shared with Maya's
DeepSeek quants and their 110 GB table file is not downloaded). A gated repo: the setup uses your Hugging Face
login. Also from #520: a tensor-parallel group reads long prompts much faster (~7x at--tp 8in his runs). - A new web UI (#538, @needmorevram): a new chat, Settings, Monitor, a Models page to switch models and edit
each one's engine settings, and "Saved" - what the model's own API would have charged. The new look is the
default; the violet colours are a choice in Settings › General. Works offline: every library and font is in the
repo.
From the community, thank you:
- DeepSeek decode latency (#495, @needmorevram): the CPU lane works in one batch per node, kernels launch
dependent on the one before (newest GPUs), and the CPU answers its own requests - the same answers. - DeepSeek on more PCs and setups, all opt-in: a series for tensor-parallel groups and a helper GPU - fused
decode stages, exact trims checked at load, a pipelined all-reduce for two GPUs, uneven groups and the helper GPU in
the prompt path, long prompts' exact skip and borrowed expert slots, a VRAM cap for the drafter's experts (#511-#519,
@pepuscz); tensor parallel with uneven heads (#411, @NenadSteric); full zero-copy routes (#530) and a
cache-budgeted Engram read-ahead (#524, @handmade0octopus); re-quantized attention matrices at load (#473) and RAM
copies that follow the routes (#449, @cryptedy); direct I/O for the RAM copies (#421, @tctiger); packed cache
transfers for EXL3 tensor parallel 2 (#493) and a refresh of the startup expert profiles (#399, @LostVector); a
direction projection before each residual write,--ablate(#469, @JoshBolding). - Loading DeepSeek: the experts' pages are dropped as they are copied, with the load's threads and file order set
(#434, @Ross-Yang-94). - AMD: RDNA 4 attention acceleration (#529, @nicholas631), MI50 / gfx906 as a build target (#505, @0c33), and
opt-in prompt-path options: transfer pruning (#297, @LostVector) and int8 MMQ for routed experts (#354,
@boxwrench). - NVIDIA DGX Spark (GB10) and Grace on CUDA, and the prompt path recovers from a failed allocation (#471,
@xcpatr9231975).
Fixes:
- DeepSeek GGUF experts uploaded from NUMA-split RAM copies are right again (#533, @tctiger).
- Calibration respects the effective tensor parallel (#435, @Whamp).
--parallelis refused for DeepSeek and MiniMax before it is written into their configs (#525, @DailyDisco).- MiniMax quants on the Hub: the dense layers' indexers file is now
<name>-maya-indexers.bin(the same bytes;
the Hub read the lone.ggufas the model). The setup renames a v1.1.30 download in place.