mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-11 18:36:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

next ollama runner (#7913)

Browse source 
feat: add new Ollama engine using ggml through cgo

This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.

- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations

This is the first implementation of the new engine. Follow up PRs will implement more features:

- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon

Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>
This commit is contained in:
Michael Yang 2025-02-14 00:31:21 +00:00 • committed by GitHub
parent 8cf16063a5
commit 58245413f4
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
57 changed files with 475427 additions and 494 deletions
Download patch file Download diff file Expand all files Collapse all files

25
server/sched.go
View file

@ -18,6 +18,7 @@ import (
"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/llm"
)
@ -41,8 +42,8 @@ type Scheduler struct {
loaded map[string]*runnerRef
loadedMu sync.Mutex
loadFn func(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoList, numParallel int)
newServerFn func(gpus discover.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
loadFn func(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel int)
newServerFn func(gpus discover.GpuInfoList, model string, f *ggml.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
getGpuFn func() discover.GpuInfoList
getCpuFn func() discover.GpuInfoList
reschedDelay time.Duration
@ -409,7 +410,7 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
}()
}
func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoList, numParallel int) {
func (s *Scheduler) load(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel int) {
if numParallel < 1 {
numParallel = 1
}
@ -417,12 +418,12 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoL
if req.sessionDuration != nil {
sessionDuration = req.sessionDuration.Duration
}
llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
llama, err := s.newServerFn(gpus, req.model.ModelPath, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
if err != nil {
// some older models are not compatible with newer versions of llama.cpp
// show a generalized compatibility error until there is a better way to
// check for model compatibility
if errors.Is(err, llm.ErrUnsupportedFormat) || strings.Contains(err.Error(), "failed to load model") {
if errors.Is(err, ggml.ErrUnsupportedFormat) || strings.Contains(err.Error(), "failed to load model") {
err = fmt.Errorf("%v: this model may be incompatible with your version of Ollama. If you previously pulled this model, try updating it by running `ollama pull %s`", err, req.model.ShortName)
}
slog.Info("NewLlamaServer failed", "model", req.model.ModelPath, "error", err)
@ -685,7 +686,7 @@ func (a ByDuration) Less(i, j int) bool {
// If the model can not be fit fully within the available GPU(s) nil is returned
// If numParallel is <= 0, this will attempt try to optimize parallelism based on available VRAM, and adjust
// opts.NumCtx accordingly
func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
func pickBestFullFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
var estimatedVRAM uint64
var numParallelToTry []int
@ -710,7 +711,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.Gpu
req.opts.NumCtx = req.origNumCtx * p
if !envconfig.SchedSpread() {
for _, g := range sgl {
if ok, estimatedVRAM = llm.PredictServerFit([]discover.GpuInfo{g}, ggml, 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); 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}
@ -726,7 +727,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.Gpu
// Now try all the GPUs
for _, p := range numParallelToTry {
req.opts.NumCtx = req.origNumCtx * p
if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
if ok, estimatedVRAM = llm.PredictServerFit(sgl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); 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
@ -737,7 +738,7 @@ func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.Gpu
}
// If multiple Libraries are detected, pick the Library which loads the most layers for the model
func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
func pickBestPartialFitByLibrary(req *LlmRequest, f *ggml.GGML, gpus discover.GpuInfoList, numParallel *int) discover.GpuInfoList {
if *numParallel <= 0 {
*numParallel = 1
req.opts.NumCtx = req.origNumCtx
@ -749,7 +750,7 @@ func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus discover.
var bestEstimate uint64
var bestFit int
for i, gl := range byLibrary {
_, estimatedVRAM := llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
_, estimatedVRAM := llm.PredictServerFit(gl, f, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
if estimatedVRAM > bestEstimate {
bestEstimate = estimatedVRAM
bestFit = i
@ -822,9 +823,9 @@ func (s *Scheduler) expireRunner(model *Model) {
// If other runners are loaded, make sure the pending request will fit in system memory
// If not, pick a runner to unload, else return nil and the request can be loaded
func (s *Scheduler) maybeFindCPURunnerToUnload(req *LlmRequest, ggml *llm.GGML, gpus discover.GpuInfoList) *runnerRef {
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, ggml, req.model.ProjectorPaths, req.opts)
estimate := llm.EstimateGPULayers(gpus, f, req.model.ProjectorPaths, req.opts)
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