matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-10 07:46:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml: Support heterogeneous KV cache layer sizes in memory estimation

Browse Source 
Gemma3 uses sliding windows for its context on 5/6 layers, significantly
reducing memory usage but leading to uneven usage across layers,
which makes allocation to the correct GPU difficult. We currently
estimate very conservatively by assuming all layers are consistent
at the max size.

Llama3.2-vision is also inconsistent between self attention and cross
attention layers - at moment, we calculate the correct total size
and then average this across layers. In some cases, this may lead
to crashes if a large layer is placed on a GPU sized by the average.

This allows memory estimation to calculate per-layer KV cache size
and take this account when placing layers onto GPUs. We already do
this for weights that vary per-tensor, so this is a logical extension.

Fixes #9730
Fixes #9890
This commit is contained in:
Jesse Gross
2025-03-24 13:39:07 -07:00
committed by Jesse Gross
parent f4f0992b6e
commit f66216e399
5 changed files with 49 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
server/sched.go
View File

@@ -711,7 +711,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
req.opts.NumCtx = req.origNumCtx * p
if !envconfig.SchedSpread() {
for _, g := range sgl {
if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
*numParallel = p
return []discover.GpuInfo{g}
@@ -727,7 +727,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuIn
// Now try all the GPUs
for _, p := range numParallelToTry {
req.opts.NumCtx = req.origNumCtx * p
if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, p); ok {
slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
*numParallel = p
return sgl
@@ -750,7 +750,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.Gp
var bestEstimate uint64
var bestFit int
for i, gl := range byLibrary {
_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, *numParallel)
if estimatedVRAM > bestEstimate {
bestEstimate = estimatedVRAM
bestFit = i
@@ -825,7 +825,7 @@ func (s *Scheduler) expireRunner(model *Model) {
// If not, pick a runner to unload, else return nil and the request can be loaded
func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList) *runnerRef {
slog.Debug("evaluating if CPU model load will fit in available system memory")
estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts, req.opts.NumCtx/req.origNumCtx)
if estimate.TotalSize <= gpus[0].FreeMemory {
slog.Debug("cpu inference mode, model fits in available system memory", "model", format.HumanBytes2(estimate.TotalSize), "available", format.HumanBytes2(gpus[0].FreeMemory))
return nil