From 345420998e90090d2d6fba38ad5c2f3f5512adf4 Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <daniel@ollama.com>
Date: Mon, 22 Jul 2024 11:57:26 -0700
Subject: [PATCH] Prevent partial loading on mixed GPU brands

In mult-brand GPU setups, if we couldn't fully load the model we
would fall through the scheduler and mistakenly try to load across
a mix of brands.  This makes sure we find the set of GPU(s) that
best fit for the partial load.
---
 server/sched.go      | 31 +++++++++++++++++++++++++++----
 server/sched_test.go | 39 +++++++++++++++++++++++++++++++++++++++
 2 files changed, 66 insertions(+), 4 deletions(-)

diff --git a/server/sched.go b/server/sched.go
index 2daed3ab..92b8d508 100644
--- a/server/sched.go
+++ b/server/sched.go
@@ -212,9 +212,12 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					} 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, &numParallel)
+						g := pickBestFullFitByLibrary(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
+						} else {
+							// Only allow partial loads when this is the first model
+							gpus = pickBestPartialFitByLibrary(pending, ggml, gpus, &numParallel)
 						}
 						s.loadFn(pending, ggml, gpus, numParallel)
 						break
@@ -231,7 +234,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 
 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
+						fitGpus := pickBestFullFitByLibrary(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
 							s.loadFn(pending, ggml, fitGpus, numParallel)
@@ -668,11 +671,12 @@ func (a ByDuration) Less(i, j int) bool {
 // func (a BySize) Swap(i, j int)      { a[i], a[j] = a[j], a[i] }
 // func (a BySize) Less(i, j int) bool { return a[i].estimatedVRAM < a[j].estimatedVRAM }
 
-// pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
+// 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
 // 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 {
+func pickBestFullFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
 
 	var numParallelToTry []int
@@ -723,6 +727,25 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP
 	return nil
 }
 
+// If multiple Libraries are detected, pick the Library which loads the most layers for the model
+func pickBestPartialFitByLibrary(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
+	*numParallel = 1
+	byLibrary := gpus.ByLibrary()
+	if len(byLibrary) <= 1 {
+		return gpus
+	}
+	var bestEstimate uint64
+	var bestFit int
+	for i, gl := range byLibrary {
+		_, estimatedVRAM := llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+		if estimatedVRAM > bestEstimate {
+			bestEstimate = estimatedVRAM
+			bestFit = i
+		}
+	}
+	return byLibrary[bestFit]
+}
+
 // findRunnerToUnload finds a runner to unload to make room for a new model
 func (s *Scheduler) findRunnerToUnload() *runnerRef {
 	s.loadedMu.Lock()
diff --git a/server/sched_test.go b/server/sched_test.go
index 9ddd1fab..a186ce0e 100644
--- a/server/sched_test.go
+++ b/server/sched_test.go
@@ -666,6 +666,45 @@ func TestAlreadyCanceled(t *testing.T) {
 	require.Empty(t, scenario1a.req.successCh)
 }
 
+func TestHomogeneousGPUs(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+
+	s.getGpuFn = func() gpu.GpuInfoList {
+		// Set memory values to require the model to be spread
+		gpus := []gpu.GpuInfo{
+			{Library: "cuda"},
+			{Library: "rocm"},
+		}
+		gpus[0].TotalMemory = 1 * format.GibiByte
+		gpus[0].FreeMemory = 256 * format.MebiByte
+		gpus[1].TotalMemory = 1 * format.GibiByte
+		gpus[1].FreeMemory = 256 * format.MebiByte
+		return gpus
+	}
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
+		require.Len(t, gpus, 1)
+		return a.newServer(gpus, model, ggml, adapters, projectors, opts, numParallel)
+	}
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
+		require.Empty(t, s.pendingReqCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		t.Fatal("timeout")
+	}
+}
+
 type mockLlm struct {
 	pingResp           error
 	waitResp           error