mw00/project-maya v1.1.25

Project Maya v1.1.25

on GitHub

Several conversations at once on GLM get faster, DeepSeek keeps conversations across switches and restarts, and a
batch of fixes - among them DeepSeek's EXL3 setup and its thinking budget.

• DeepSeek EXL3 setup no longer stops at the pack step ("keeps dense weights in FP8", #376): since v1.1.21 every
  fresh EXL3 setup did. A download already made carries on from the pack step.
• DeepSeek keeps conversations across switches and restarts, so agents with sub-agents stop re-reading their
  prompts: opt-in, MAYA_DS4_SLOT_KEEP=1 (#356, @mrsirg97-rgb). A retried prompt or a re-rendered turn is taken back at
  the conversation's checkpoints (#394, @mgabor3141).
• DeepSeek's thinking budget (thinking_budget in the config) is kept: the reasoning ends there and the answer
  follows (#363). A 1..100 reasoning_effort shows as itself at the start.
• DeepSeek decode: the PCIe lane is now judged over whole decode steps too, and streams only where the token gets
  faster (#329). DS4_STREAM_TOKEN_AB=0 keeps the earlier check alone.
• GLM, several conversations at once (--parallel, @somnergy): MTP drafting in the batched steps
  (MAYA_GLM_MULTI_MTP, #385) and the rows' attention and small kernels in one launch (MAYA_GLM_ROWS_ATTN /
  _ROWS_SMALL, #391); running requests keep decoding while a new prompt is read (MAYA_GLM_PREFILL_INTERLEAVE /
  _GAPS, #386; MAYA_GLM_PREFILL_MIN_CHUNKS, #392); a conversation's system prompt is kept in a slot, so new
  conversations restore it instead of reading it again (#387). All opt-in except #387's, which is on with the slots.
• GLM EXL3 prompts: short chunks on the tensor cores, MAYA_GLM_EXL3_MMA_ROWS_SHORT - 2K-token prompts 28% faster
  on the contributor's PC (#388, @somnergy).
• Fixes: the server no longer answers "exceeds the context (0)" to every request after an engine that died could
  not start again - each request tries again, or says why it cannot (#351); --speed-calibration measures DeepSeek's
  DSpark drafter again (#393); the Monitor finds an engine started through a launcher (#398, @somnergy); the pack's MTP
  message no longer reads as drafting off (#327); docs on the CPU lane's threads (#390, @somnergy).