matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-11 00:07:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

review comments and coverage

Browse Source
This commit is contained in:
Daniel Hiltgen
2024-06-05 12:07:20 -07:00
parent ff4f0cbd1d
commit 6f351bf586
18 changed files with 375 additions and 456 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
server/sched.go
View File

@@ -182,7 +182,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
// We want to avoid loading on any GPUs that have other
// models still loading on them to avoid potential races
// with VRAM consumption ramping up during load
availGpus := s.filterGPUsWithLoadingModels(gpus)
availGpus := s.filterGPUsWithoutLoadingModels(gpus)
// Update free memory from currently loaded models
s.updateFreeSpace(availGpus)
@@ -414,9 +414,7 @@ func (s *Scheduler) updateFreeSpace(allGpus gpu.GpuInfoList) {
r.refMu.Lock()
if r.llama != nil {
for _, gpu := range allGpus {
// if slices.Contains(gpuIDs, gpu.ID) {
predMap[predKey{gpu.Library, gpu.ID}] += r.llama.EstimagedVRAMByGPU(gpu.ID)
// }
predMap[predKey{gpu.Library, gpu.ID}] += r.llama.EstimatedVRAMByGPU(gpu.ID)
}
} else {
slog.Warn("unexpected nil runner reference, memory prediction may be incorrect")
@@ -448,7 +446,7 @@ func (s *Scheduler) updateFreeSpace(allGpus gpu.GpuInfoList) {
// to avoid scheduling another model on the same GPU(s) that haven't stabilized.
// This routine returns the set of GPUs that do not have an active loading model.
// If all GPUs have loading models, an empty list will be returned (not a single CPU entry)
func (s *Scheduler) filterGPUsWithLoadingModels(allGpus gpu.GpuInfoList) gpu.GpuInfoList {
func (s *Scheduler) filterGPUsWithoutLoadingModels(allGpus gpu.GpuInfoList) gpu.GpuInfoList {
ret := append(gpu.GpuInfoList{}, allGpus...)
s.loadedMu.Lock()
defer s.loadedMu.Unlock()
@@ -702,5 +700,4 @@ func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, ggml *llm.GGML,
// TODO - optimization: try to find CPU only runners first, or partial offloads with enough in system memory to make room
return s.findRunnerToUnload()
}