mw00/project-maya v1.1.22

Project Maya v1.1.22

on GitHub

DeepSeek-V4.1-Flash gets Maya-XS - the whole model in a 128 GB PC's memory - tensor parallel on several GPUs, and much
faster drafting on one GPU.

• Maya-XS (experimental): DeepSeek-V4.1-Flash in 120.6 GB of memory, its own 2-bit DSpark drafter included - every
  one of its 384 experts a layer kept, in ~1.6 bits (IQ1_S), rounded with error feedback layer by layer on the
  quantized model itself, from three times Maya-S's calibration text (code, three.js and front-end, tool calls and
  agent traces, long reasoning). The Engram tables (161 GB, Q6_K) stay on the SSD as with Maya-S. On a 128 GB Ryzen AI
  Max every expert sits in GPU memory. ./setup.sh --setup --model DeepSeek-V4.1-Flash-1017 --ds4-quant maya-xs.
  The engine now runs IQ1_S and IQ1_M experts on NVIDIA, on AMD and in the CPU lane.
• DeepSeek tensor parallel on several GPUs (#344, @needmorevram): prompts ~1.9x and answers ~1.55x faster at
  --tp 8 than v1.1.21 on 9 GPUs, the same perplexity; ./maya.sh --calibrate tunes --tp. On one GPU, drafting
  decodes ~1.85x faster (Maya-S with DSpark on a Ryzen AI Max, the same drafts kept: each verify round about half
  the time).
• AMD: a 2-GPU DeepSeek hang on the first long prompt with hipBLASLt fixed (#367, @Gazelle221B; #366); the packed
  sparse attention's LDS swizzle, opt-in DS4_HIP_ATTN_SWIZZLE=1 (#348, @LostVector); --bench takes the DeepSeek
  engine on HIP (#350, @GravLabs).
• DeepSeek on Windows (@OscarSheng): a 1 ms timer while the engine runs (#358; MAYA_DS4_TIMER_MS=0 leaves it),
  64 Engram readers (#361) and unbuffered Engram reads (#362; DS4_EG_DIRECT=0 = before) - faster prompts.
• DeepSeek prompt report: which experts a prompt found resident and the bytes it streamed, DS4_PREFILL_STATS=1
  (#364, @LostVector).
• GLM: a prompt's lending keeps the more recently used experts in RAM - 40% fewer disk reads after a fresh prompt
  (#352, @merbanan).
• Fixes: maya-ds4 builds with CUDA 12.0 / 12.1 (#336); the HIP build's test programs link (#355); VERSION holds
  the version again (v1.1.21's was empty).