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

sched: only error when over-allocating system memory (#5626)

Browse source
This commit is contained in:
Jeffrey Morgan 2024-07-11 00:53:12 -07:00 committed by GitHub
parent efbf41ed81
commit 791650ddef
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 9 additions and 37 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
llm/server.go
View file

@ -122,6 +122,15 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
}
}
// On linux, over-allocating CPU memory will almost always result in an error
if runtime.GOOS == "linux" {
systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
if systemMemoryRequired > systemTotalMemory {
slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "system", format.HumanBytes2(systemTotalMemory))
return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(systemMemoryRequired), format.HumanBytes2(systemTotalMemory))
}
}
estimate.log()
// Loop through potential servers

37
server/sched.go
View file

@ -135,11 +135,6 @@ func (s *Scheduler) processPending(ctx context.Context) {
}
for {
cpus := s.getCpuFn()
var systemMem gpu.GpuInfo
if len(cpus) > 0 {
systemMem = cpus[0]
}
var runnerToExpire *runnerRef
s.loadedMu.Lock()
runner := s.loaded[pending.model.ModelPath]
@ -193,38 +188,6 @@ func (s *Scheduler) processPending(ctx context.Context) {
break
}
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
}
if loadedCount == 0 {
// If no other models are loaded, set the limit based on what's available
maxSize += gpu.FreeMemory
} else {
// Other models could be unloaded, favor total memory for limit
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" {
// simplifying assumption of defaultParallel when in CPU mode