Pull requests / #1931
#1931 fix(hip): add opt-in stability profile for reset reports
open · @CC-David-CC · 0 comentários · No GitHub
Descrição
Long HIP inference can stress host-memory transfers, prompt cache refills, and PCIe traffic. Add an opt-in conservative profile to reduce overlapping host-memory and GPU operations while preserving the requested context and model settings. Set `STRATA_HIP_STABLE=1` in the server environment to use owned 2K prompt buffers, pageable resident experts, CPU expert misses, checked prefill stream drains, and conservative verify synchronization. Set `STRATA_HIP_STABLE=2` to additionally reserve the full KV cache in VRAM and disable elastic KV remapping and adaptive expert swaps. Remove the variable and restart the engine to return to normal behavior. Related to #872: https://github.com/Niko1221/Strata/issues/872
No site
Links install, modelos, releases.