GLM keeps every expert in VRAM on NVIDIA where they fit, tuned speed settings are measured again after a driver
change, GLM conversations with pictures keep their slots, faster long prompts for Qwen-3.8-Flash-Next (CYBER-FROST) on
three or more GPUs, and a MiniMax --parallel fix.
- GLM: every expert stays in VRAM on NVIDIA where it fits (#474, @walker2kk): where the GPUs' memory holds every
expert, NVIDIA now gets the same complete expert pool as AMD, so a long prompt no longer pushes experts out of VRAM
and the decode after it stays at full speed (the reporter's 4x A100 split: 2.5x decode). Where the experts do not all
fit, nothing changes.MAYA_GLM_COMPLETE_POOL=0keeps the old pool. - A tuned speed setting is measured again after a driver or P2P change (#419, @catonooka): the CPU-lane tuning
(MAYA_GLM_PCIE_SHARE,MAYA_GLM_CPU_LANE,MAYA_GLM_READ_CHUNKS) now remembers the NVIDIA driver, the GPUs' slots
and which GPUs reach each other directly (P2P). When one of them changes, Maya says what changed, drops the old values
and the engine measures again at its start (the reporter's 2x RTX 3090 after a P2P driver: 1.6x decode back);
./maya.sh --calibratetunes it again. A setup no longer re-applies a stale stored tuning. - GLM: a conversation with pictures keeps its slot (#331, @Cass67; also reported in #418 by @CoolJohn-lab): a chat
with pictures is set aside in a conversation slot like a text chat, so a chat with one screenshot no longer reads its
whole history again each time another chat pushed it out. Kept slots from an older version are removed once at start
(the slot format changed). - GLM: an edited last message after switching chats (#400, @Zcg2021): A -> B -> back to A with the last message
edited no longer reads the whole conversation again (that step's prompt read about 10x faster on a 2K-token chat). - GLM calibration with more CPU-lane threads (#178, @Jw88888): on one GPU whose CPUs show no SMT siblings (a VM's
vCPUs)--calibratealso measures more CPU-lane threads than the engine's own count, and lower counts are no longer
measured slower than they run; new settingMAYA_GLM_CPU_LANE_MAX. - MiniMax
--parallel: a short question no longer evicts a long picture conversation: the picture was read again
at that conversation's next turn. A new request now takes an empty conversation (or the least recently used one);
continued conversations are reused as before (that next turn 2.9x faster on 3x V100, its answer the same). - Qwen-3.8-Flash-Next (CYBER-FROST) on three or more GPUs reads long prompts with every GPU at once: with the model
split across GPUs, the second and third GPU used to take turns on each piece of the prompt; each now works on a
different piece at the same time. On three V100s with most experts in VRAM, prompts of 4K, 8K and 16K tokens read
about 1.14x, 1.35x and 1.5x faster (a 2K prompt is one piece: no change); with experts streamed from RAM the gain is
small. Answers are the same; one GPU and two GPUs are unchanged. - Chat: the attach menu's item reads "Select file".