mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

DRY out the runner lifecycle code (#12540)

Browse Source 
* DRY out the runner lifecycle code

Now that discovery uses the runners as well, this unifies the runner spawning code
into a single place.  This also unifies GPU discovery types with the newer ml.DeviceInfo

* win: make incremental builds better

Place build artifacts in discrete directories so incremental builds don't have to start fresh

* Adjust sort order to consider iGPUs

* handle cpu inference oom scenarios

* review comments
This commit is contained in:
Daniel Hiltgen
2025-10-23 11:20:02 -07:00
committed by GitHub
parent 1c093e97af
commit 3258a89b6e
16 changed files with 720 additions and 924 deletions
Download Patch File Download Diff File Expand all files Collapse all files

71
llm/memory.go
View File

@@ -4,27 +4,28 @@ import (
"fmt"
"log/slog"
"os"
"slices"
"sort"
"strings"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/discover"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/fs/ggml"
"github.com/ollama/ollama/ml"
)
// pickBestFullFitByLibrary will try to find the optimal placement of the model in the available GPUs where the model fully fits
// The list of GPUs returned will always be the same brand (library)
// If the model can not be fit fully within the available GPU(s) nil is returned
func pickBestFullFitByLibrary(f *ggml.GGML, modelPath string, projectors []string, adapters []string, opts api.Options, gpus discover.GpuInfoList, numParallel int) discover.GpuInfoList {
for _, gl := range gpus.ByLibrary() {
sgl := append(make(discover.GpuInfoList, 0, len(gl)), gl...)
func pickBestFullFitByLibrary(f *ggml.GGML, modelPath string, projectors []string, adapters []string, opts api.Options, gpus []ml.DeviceInfo, numParallel int) []ml.DeviceInfo {
for _, gl := range ml.ByLibrary(gpus) {
sgl := append(make([]ml.DeviceInfo, 0, len(gl)), gl...)
// TODO - potentially sort by performance capability, existing models loaded, etc.
// TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them
// Note: at present, this will favor most current available VRAM descending and ignoring faster GPU speed in mixed setups
sort.Sort(sort.Reverse(discover.ByFreeMemory(sgl)))
sort.Sort(sort.Reverse(ml.ByFreeMemory(sgl)))
if !envconfig.SchedSpread() {
// Try to pack into as few GPUs as possible, starting from 1 GPU
@@ -63,8 +64,8 @@ func pickBestFullFitByLibrary(f *ggml.GGML, modelPath string, projectors []strin
}
// If multiple Libraries are detected, pick the Library which loads the most layers for the model
func pickBestPartialFitByLibrary(f *ggml.GGML, projectors []string, adapters []string, opts api.Options, gpus discover.GpuInfoList, numParallel int) discover.GpuInfoList {
byLibrary := gpus.ByLibrary()
func pickBestPartialFitByLibrary(f *ggml.GGML, projectors []string, adapters []string, opts api.Options, gpus []ml.DeviceInfo, numParallel int) []ml.DeviceInfo {
byLibrary := ml.ByLibrary(gpus)
if len(byLibrary) <= 1 {
return gpus
}
@@ -81,10 +82,10 @@ func pickBestPartialFitByLibrary(f *ggml.GGML, projectors []string, adapters []s
}
// This algorithm looks for a complete fit to determine if we need to unload other models
func predictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
func predictServerFit(allGpus []ml.DeviceInfo, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
// Split up the GPUs by type and try them
var estimatedVRAM uint64
for _, gpus := range allGpus.ByLibrary() {
for _, gpus := range ml.ByLibrary(allGpus) {
var layerCount int
estimate := estimateGPULayers(gpus, f, projectors, opts, numParallel)
layerCount, estimatedVRAM = estimate.Layers, estimate.VRAMSize
@@ -97,14 +98,23 @@ func predictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, proj
return true, estimatedVRAM
}
}
if len(gpus) == 1 && gpus[0].Library == "cpu" && estimate.TotalSize <= gpus[0].FreeMemory {
return true, estimatedVRAM
}
}
return false, estimatedVRAM
}
func verifyCPUFit(f *ggml.GGML, modelPath string, projectors []string, adapters []string, opts api.Options, systemInfo ml.SystemInfo, numParallel int) bool {
estimate := estimateGPULayers(nil, f, projectors, opts, numParallel)
if estimate.TotalSize > systemInfo.FreeMemory {
return false
}
slog.Info("new model will fit in available system memory for CPU inference, loading",
"model", modelPath,
"parallel", numParallel,
"required", format.HumanBytes2(estimate.TotalSize),
)
return true
}
type MemoryEstimate struct {
// How many layers we predict we can load
Layers int
@@ -141,7 +151,7 @@ type MemoryEstimate struct {
// Given a model and one or more GPU targets, predict how many layers and bytes we can load, and the total size
// The GPUs provided must all be the same Library
func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
func estimateGPULayers(gpus []ml.DeviceInfo, f *ggml.GGML, projectors []string, opts api.Options, numParallel int) MemoryEstimate {
// Graph size for a partial offload, applies to all GPUs
var graphPartialOffload uint64
@@ -175,10 +185,17 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
overhead := envconfig.GpuOverhead()
availableList := make([]string, len(gpus))
libraries := []string{}
for i, gpu := range gpus {
availableList[i] = format.HumanBytes2(gpu.FreeMemory)
if !slices.Contains(libraries, gpu.Library) {
libraries = append(libraries, gpu.Library)
}
}
slog.Debug("evaluating", "library", gpus[0].Library, "gpu_count", len(gpus), "available", availableList)
if len(libraries) == 0 {
libraries = []string{"cpu"}
}
slog.Debug("evaluating", "library", strings.Join(libraries, ","), "gpu_count", len(gpus), "available", availableList)
for _, projector := range projectors {
llamaEngineProjectorWeights += projectorMemoryRequirements(projector)
@@ -196,7 +213,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
}
useFlashAttention := envconfig.FlashAttention(f.FlashAttention()) &&
(discover.GpuInfoList)(gpus).FlashAttentionSupported() &&
ml.FlashAttentionSupported(gpus) &&
f.SupportsFlashAttention()
var kvct string
@@ -231,7 +248,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
}
// on metal there's no partial offload overhead
if gpus[0].Library == "Metal" {
if len(gpus) > 0 && gpus[0].Library == "Metal" {
graphPartialOffload = graphFullOffload
} else if len(gpus) > 1 {
// multigpu should always use the partial graph size
@@ -256,7 +273,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
gpuAllocations := make([]uint64, len(gpus))
type gs struct {
i int
g *discover.GpuInfo
g *ml.DeviceInfo
}
gpusWithSpace := []gs{}
for i := range gpus {
@@ -265,19 +282,11 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
gzo = gpuZeroOverhead
}
// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
if gpus[i].FreeMemory < overhead+gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
var compute string
if gpus[i].Library == "ROCm" {
compute = fmt.Sprintf("gfx%x%02x", gpus[i].ComputeMajor, gpus[i].ComputeMinor)
} else {
compute = fmt.Sprintf("%d.%d", gpus[i].ComputeMajor, gpus[i].ComputeMinor)
}
if gpus[i].FreeMemory < overhead+gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory()+2*layerSize {
slog.Debug("gpu has too little memory to allocate any layers",
"id", gpus[i].ID,
"library", gpus[i].Library,
"variant", gpus[i].Variant,
"compute", compute,
"compute", gpus[i].Compute(),
"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
"name", gpus[i].Name,
"total", format.HumanBytes2(gpus[i].TotalMemory),
@@ -291,7 +300,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
continue
}
gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
gpuAllocations[i] += gpus[i].MinimumMemory + layerSize // We hold off on graph until we know partial vs. full
gpuAllocations[i] += gpus[i].MinimumMemory() + layerSize // We hold off on graph until we know partial vs. full
}
var gpuZeroID int
@@ -397,7 +406,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
VRAMSize: 0,
GPUSizes: []uint64{},
inferenceLibrary: gpus[0].Library,
inferenceLibrary: strings.Join(libraries, ","),
layersRequested: opts.NumGPU,
layersModel: int(f.KV().BlockCount()) + 1,
availableList: availableList,
@@ -411,7 +420,7 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
projectorGraph: ollamaEngineProjectorGraph,
}
if gpus[0].Library == "cpu" {
if len(gpus) == 0 {
return estimate
}
if layerCount == 0 {