baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

sched: don't error if paging to disk on Windows and macOS (#5523)

Browse source
This commit is contained in:
Jeffrey Morgan 2024-07-06 22:01:52 -04:00 committed by GitHub
parent f8241bfba3
commit 0ee87615c7
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 24 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
server/sched.go
View file

@ -197,9 +197,12 @@ func (s *Scheduler) processPending(ctx context.Context) {
break
}
// Block attempting to load a model larger than system memory + GPU memory
estimate := llm.EstimateGPULayers(gpus, ggml, pending.model.ProjectorPaths, pending.opts)
maxSize := systemMem.FreeMemory
// Add available GPU memory to the total pool
// macOS hardware has unified memory so don't double count
if runtime.GOOS != "darwin" {
for _, gpu := range gpus {
if gpu.Library == "cpu" {
continue
@ -212,11 +215,19 @@ func (s *Scheduler) processPending(ctx context.Context) {
maxSize += gpu.TotalMemory
}
}
}
// Block attempting to load a model larger than system memory + GPU memory
if estimate.TotalSize > maxSize {
slog.Warn("model request too large for system", "requested", format.HumanBytes2(estimate.TotalSize), "system", format.HumanBytes2(maxSize))
// Linux will crash if over-allocating memory - return an error to the user.
// TODO (jmorganca): add reasonable upper limits for darwin and windows as well
if runtime.GOOS == "linux" {
pending.errCh <- fmt.Errorf("requested model (%s) is too large for this system (%s)", format.HumanBytes2(estimate.TotalSize), format.HumanBytes2(maxSize))
break
}
}
// Evaluate if the model will fit in the available system memory, or if we should unload a model first
if len(gpus) == 1 && gpus[0].Library == "cpu" {