Enable concurrency by default

This adjusts our default settings to enable multiple models and parallel requests to a single model. Users can still override these by the same env var settings as before. Parallel has a direct impact on num_ctx, which in turn can have a significant impact on small VRAM GPUs so this change also refines the algorithm so that when parallel is not explicitly set by the user, we try to find a reasonable default that fits the model on their GPU(s). As before, multiple models will only load concurrently if they fully fit in VRAM.
2025-12-11 08:17:03 +00:00 · 2024-05-06 17:47:52 -07:00
parent c7c2f3bc22
commit 17b7186cd7
4 changed files with 135 additions and 72 deletions
--- a/server/sched.go
+++ b/server/sched.go
@@ -23,6 +23,7 @@ type LlmRequest struct {
 	ctx             context.Context //nolint:containedctx
 	model           *Model
 	opts            api.Options
+	origNumCTX      int // Track the initial ctx request
 	sessionDuration time.Duration
 	successCh       chan *runnerRef
 	errCh           chan error
@@ -38,8 +39,8 @@ type Scheduler struct {
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex

-	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
-	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
+	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int)
+	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
 	getGpuFn     func() gpu.GpuInfoList
 	getCpuFn     func() gpu.GpuInfoList
 	reschedDelay time.Duration
@@ -65,13 +66,10 @@ func InitScheduler(ctx context.Context) *Scheduler {

 // context must be canceled to decrement ref count and release the runner
 func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
-	// allocate a large enough kv cache for all parallel requests
 	if opts.NumCtx < 4 {
 		opts.NumCtx = 4
 	}

-	opts.NumCtx *= envconfig.NumParallel
-
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
@@ -102,6 +100,7 @@ func (s *Scheduler) Run(ctx context.Context) {
 }

 func (s *Scheduler) processPending(ctx context.Context) {
+	maxRunnerFactor := 1 // number of GPUs or 1
 	for {
 		select {
 		case <-ctx.Done():
@@ -110,11 +109,25 @@ func (s *Scheduler) processPending(ctx context.Context) {
 		case pending := <-s.pendingReqCh:
 			// Block other requests until we get this pending request running
 			pending.schedAttempts++
+			if pending.origNumCTX == 0 {
+				pending.origNumCTX = pending.opts.NumCtx
+			}

 			if pending.ctx.Err() != nil {
 				slog.Debug("pending request cancelled or timed out, skipping scheduling")
 				continue
 			}
+			numParallel := envconfig.NumParallel
+			// TODO (jmorganca): multimodal models don't support parallel yet
+			// see https://github.com/ollama/ollama/issues/4165
+			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
+				numParallel = 1
+				slog.Warn("multimodal models don't support parallel requests yet")
+			}
+			// Keep NumCtx and numParallel in sync
+			if numParallel > 1 {
+				pending.opts.NumCtx = pending.origNumCTX * numParallel
+			}

 			for {
 				var runnerToExpire *runnerRef
@@ -130,7 +143,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						pending.useLoadedRunner(runner, s.finishedReqCh)
 						break
 					}
-				} else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners {
+				} else if envconfig.MaxRunners > 0 && loadedCount >= (maxRunnerFactor*envconfig.MaxRunners) {
 					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
 					runnerToExpire = s.findRunnerToUnload()
 				} else {
@@ -142,6 +155,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					} else {
 						gpus = s.getGpuFn()
 					}
+					maxRunnerFactor = max(len(gpus), 1)

 					// Load model for fitting
 					ggml, err := llm.LoadModel(pending.model.ModelPath)
@@ -152,26 +166,32 @@ func (s *Scheduler) processPending(ctx context.Context) {

 					// 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
+						if numParallel <= 0 {
+							numParallel = defaultParallel
+							pending.opts.NumCtx = pending.origNumCTX * numParallel
+						}
+
 						if loadedCount == 0 {
 							slog.Debug("cpu mode with first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus)
 						if runnerToExpire == nil {
 							slog.Debug("cpu mode with available system memory or first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						// else we need to expire a runner
 					} else if loadedCount == 0 {
 						// No models loaded. Load the model but prefer the best fit.
 						slog.Debug("loading first model", "model", pending.model.ModelPath)
-						g := pickBestFitGPUs(pending, ggml, gpus)
+						g := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
 						}
-						s.loadFn(pending, ggml, gpus)
+						s.loadFn(pending, ggml, gpus, numParallel)
 						break
 					}

@@ -186,10 +206,10 @@ func (s *Scheduler) processPending(ctx context.Context) {

 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus)
+						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
-							s.loadFn(pending, ggml, fitGpus)
+							s.loadFn(pending, ggml, fitGpus, numParallel)
 							break
 						}

@@ -350,8 +370,11 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	}()
 }

-func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
-	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+	if numParallel < 1 {
+		numParallel = 1
+	}
+	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, 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
@@ -375,6 +398,7 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 		loading:         true,
 		refCount:        1,
 	}
+	runner.numParallel = numParallel
 	runner.refMu.Lock()

 	s.loadedMu.Lock()
@@ -483,8 +507,9 @@ type runnerRef struct {
 	expireTimer     *time.Timer
 	expiresAt       time.Time

-	model     *Model
-	modelPath string
+	model       *Model
+	modelPath   string
+	numParallel int
 	*api.Options
 }

@@ -525,6 +550,9 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 		optsNew.NumGPU = -1
 	}

+	// Normalize the NumCtx for parallelism
+	optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel
+
 	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
@@ -611,22 +639,38 @@ func (a ByDuration) Less(i, j int) bool {

 // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
 // If the model can not be fit fully within the available GPU(s) nil is returned
-func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
+// If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
+// opts.NumCtx accordingly
+func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
+
+	var numParallelToTry []int
+	if *numParallel <= 0 {
+		// If no specific parallel setting was provided, try larger then smaller, always end with 1
+		numParallelToTry = append(numParallelToTry, 4, 1)
+	} else {
+		numParallelToTry = []int{*numParallel}
+	}
+
 	for _, gl := range gpus.ByLibrary() {
 		var ok bool
 		sgl := append(make(gpu.GpuInfoList, 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 more VRAM over faster GPU speed in mixed setups
 		sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl)))

 		// First attempt to fit the model into a single GPU
-		if !envconfig.SchedSpread {
-			for _, g := range sgl {
-				if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-					slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
-					return []gpu.GpuInfo{g}
+		for _, p := range numParallelToTry {
+			req.opts.NumCtx = req.origNumCTX * p
+			if !envconfig.SchedSpread {
+				for _, g := range sgl {
+					if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, 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 []gpu.GpuInfo{g}
+					}
 				}
 			}
 		}
@@ -636,9 +680,13 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family

 		// Now try all the GPUs
-		if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
-			return sgl
+		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 {
+				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
+			}
 		}
 	}
 	return nil