-
donor-v0.3.4 — donated-capacity auto-scale
Pre-releaseAll checks were successfuldeploy / deploy (push) Successful in 4sreleased this
2026-06-28 22:48:43 +00:00 | 513 commits to main since this releaseNative GPU donor v0.3.4 — auto-scale now counts DONATED capacity, not full card VRAM.
Builds on v0.3.3 (flap-resistant keepalive + platform telemetry) and adds:
- The donor now reports its donation in gpu_register: utilizationPct (duty-cycle) + donatedMB (explicit --memory VRAM cap). The brain weights community-compute auto-scale by EFFECTIVE donated capacity, so e.g. two 15GB cards set to 60% count as 18GB, not 30GB — no false tier-up.
- Honors the data-parallel caveat: utilization is a throughput duty-cycle, not VRAM held; the admin /autoscale + dashboard now expose minDonorMB (the smallest donor's committed VRAM = the real brain-SIZE bound).
Downloads: Windows (.exe) + Linux (x86-64). NVIDIA = CUDA; auto-falls-back to WebGPU on AMD/Intel/Apple. WIP.
Downloads
-
Source code (ZIP)
0 downloads
-
Source code (TAR.GZ)
0 downloads