mw00/project-maya v1.1.24

Project Maya v1.1.24

on GitHub

A fix for DeepSeek on two NVIDIA GPUs under Windows, faster DeepSeek prompts on NVIDIA RTX 30 and newer and on RDNA 4,
and new options for large-RAM servers.

• DeepSeek on two or more NVIDIA GPUs on Windows no longer hangs on the first prompt (#384, #303): one GPU waited
  for a hand-off the other GPU's queued work had not yet sent. DS4_WIN_SUBMIT=0 restores the old behaviour for
  comparison. Linux and one-GPU setups are unchanged.
• Monitor: DeepSeek's "Where the experts live" card shows VRAM, RAM and SSD apart again, each with its experts and
  GB, and the experts a token from RAM, from the SSD and promoted to VRAM - as GLM's - with the n-gram memory beside
  them (from the SSD: measured on Linux).
• Prompt attention on NVIDIA tensor cores (RTX 30 and newer, sm_80+; #369, @jinsoojinsoo) and on RDNA 4's
  matrix cores (#370, @Gazelle221B), like RDNA 3's since v1.1.21. DS4_ATTN_MMA=0 / DS4_ATTN_WMMA=0 turn them off.
• Tensor parallel: a GPU group member reads only its own expert rows from the RAM copy (#377, @NenadSteric).
• Large-RAM Linux servers: opt-in locked Engram tables (DS4_ENGRAM_RAM=1) and every CPU-lane expert held in RAM
  with no fallback to the SSD (DS4_REQUIRE_ALL_RAM=1) (#368, @tctiger). docs/SETTINGS.md.
• Options (@pepuscz): MAYA_DS4_Q8_BATCH=1 rounds a GEMV call's inputs in one launch, with the same bits (#382);
  MAYA_DS4_LT_STORE keeps AMD's measured hipBLASLt picks outside the pack (#381); DS4_PROF prints its line for each
  served request (#380); the kernel test fails on a NaN or Inf (#379).